Dataset statistics
| Number of variables | 4 |
|---|---|
| Number of observations | 19 |
| Missing cells | 0 |
| Missing cells (%) | 0.0% |
| Duplicate rows | 0 |
| Duplicate rows (%) | 0.0% |
| Total size in memory | 1.3 MiB |
| Average record size in memory | 70.9 KiB |
Variable types
| Categorical | 4 |
|---|
Alerts
source_doc_filename is highly overall correlated with source_doc_id and 2 other fields | High correlation |
source_doc_id is highly overall correlated with source_doc_filename and 2 other fields | High correlation |
source_doc_domain is highly overall correlated with source_doc_filename and 2 other fields | High correlation |
document_text is highly overall correlated with source_doc_filename and 2 other fields | High correlation |
source_doc_filename is uniformly distributed | Uniform |
source_doc_id is uniformly distributed | Uniform |
document_text is uniformly distributed | Uniform |
source_doc_filename has unique values | Unique |
source_doc_id has unique values | Unique |
document_text has unique values | Unique |
Reproduction
| Analysis started | 2023-05-22 09:55:54.385716 |
|---|---|
| Analysis finished | 2023-05-22 09:55:55.508283 |
| Duration | 1.12 second |
| Software version | ydata-profiling vv4.0.0 |
| Download configuration | config.json |
source_doc_filename
Categorical
HIGH CORRELATION UNIFORM UNIQUE
| Distinct | 19 |
|---|---|
| Distinct (%) | 100.0% |
| Missing | 0 |
| Missing (%) | 0.0% |
| Memory size | 2.1 KiB |
| ASRnlp_law_lecture_week_1_v_2_c_transcription_1.txt | 1 |
|---|---|
| OCR_ML4HLecture05-NLP.pptx_.txt | 1 |
| script_sunsetblvd..txt | 1 |
| script_strangersonatrain.txt | 1 |
| script_frozendisney.txt | 1 |
| Other values (14) |
Length
| Max length | 124 |
|---|---|
| Median length | 44 |
| Mean length | 47.105263 |
| Min length | 22 |
Characters and Unicode
| Total characters | 895 |
|---|---|
| Distinct characters | 57 |
| Distinct categories | 7 ? |
| Distinct scripts | 2 ? |
| Distinct blocks | 1 ? |
The Unicode Standard assigns character properties to each code point, which can be used to analyse textual variables.
Unique
| Unique | 19 ? |
|---|---|
| Unique (%) | 100.0% |
Sample
| 1st row | ASRnlp_law_lecture_week_1_v_2_c_transcription_1.txt |
|---|---|
| 2nd row | ASRnlp_law_lecture_week_2_v_2_c_transcription_2.txt |
| 3rd row | ASRnlp_law_lecture_week_3_part_1_v_2_c_transcription_3.txt |
| 4th row | ASR-whisper-rpunctuated_Noam Chomsky, Fundam_1669853561_0_part1.txt |
| 5th row | ASR-whisper-rpunctuated_Noam Chomsky, Fundam_1669853631_0_part2.txt |
Common Values
| Value | Count | Frequency (%) |
| ASRnlp_law_lecture_week_1_v_2_c_transcription_1.txt | 1 | 5.3% |
| OCR_ML4HLecture05-NLP.pptx_.txt | 1 | 5.3% |
| script_sunsetblvd..txt | 1 | 5.3% |
| script_strangersonatrain.txt | 1 | 5.3% |
| script_frozendisney.txt | 1 | 5.3% |
| script_findingnemo.txt | 1 | 5.3% |
| OCR_PAPER_Kandpal, Nieto, Jin - 2022 - Music Enhancement via Image Translation and Vocoding-annotated_.txt | 1 | 5.3% |
| OCR_PAPER_Hong et al. - 2022 - CogVideo Large-scale Pretraining for Text-to-Video Generation via Transformers-annotated_.txt | 1 | 5.3% |
| OCR_PAPER_dall-e-2-annotated_.txt | 1 | 5.3% |
| OCR_ML4HLecture04RepresentationLearning.pptx_.txt | 1 | 5.3% |
| Other values (9) | 9 |
Length
Histogram of lengths of the category
| Value | Count | Frequency (%) |
| 4 | 7.3% | |
| asr-whisper-rpunctuated_noam | 2 | 3.6% |
| 2022 | 2 | 3.6% |
| chomsky | 2 | 3.6% |
| via | 2 | 3.6% |
| gpt_peter_testing_group_exemplars.txt | 1 | 1.8% |
| ocr_paper_dall-e-2-annotated_.txt | 1 | 1.8% |
| ocr_ml4hlecture04representationlearning.pptx_.txt | 1 | 1.8% |
| asrnlp_law_lecture_week_2_v_2_c_transcription_2.txt | 1 | 1.8% |
| ocr_ml4hlecture02image_.txt | 1 | 1.8% |
| Other values (38) | 38 |
Most occurring characters
| Value | Count | Frequency (%) |
| t | 90 | 10.1% |
| e | 64 | 7.2% |
| _ | 62 | 6.9% |
| n | 54 | 6.0% |
| a | 53 | 5.9% |
| r | 44 | 4.9% |
| i | 36 | 4.0% |
| 36 | 4.0% | |
| s | 32 | 3.6% |
| o | 32 | 3.6% |
| Other values (47) | 392 |
Most occurring categories
| Value | Count | Frequency (%) |
| Lowercase Letter | 604 | |
| Uppercase Letter | 95 | 10.6% |
| Connector Punctuation | 62 | 6.9% |
| Decimal Number | 52 | 5.8% |
| Space Separator | 36 | 4.0% |
| Other Punctuation | 27 | 3.0% |
| Dash Punctuation | 19 | 2.1% |
Most frequent character per category
Lowercase Letter
| Value | Count | Frequency (%) |
| t | 90 | |
| e | 64 | |
| n | 54 | 8.9% |
| a | 53 | 8.8% |
| r | 44 | 7.3% |
| i | 36 | 6.0% |
| s | 32 | 5.3% |
| o | 32 | 5.3% |
| p | 29 | 4.8% |
| c | 25 | 4.1% |
| Other values (14) | 145 |
Uppercase Letter
| Value | Count | Frequency (%) |
| R | 16 | |
| C | 10 | |
| L | 9 | |
| A | 8 | |
| P | 8 | |
| O | 6 | 6.3% |
| S | 5 | 5.3% |
| M | 5 | 5.3% |
| E | 5 | 5.3% |
| T | 4 | 4.2% |
| Other values (9) | 19 |
Decimal Number
| Value | Count | Frequency (%) |
| 2 | 14 | |
| 1 | 8 | |
| 0 | 7 | |
| 6 | 6 | |
| 3 | 5 | 9.6% |
| 5 | 4 | 7.7% |
| 4 | 4 | 7.7% |
| 9 | 2 | 3.8% |
| 8 | 2 | 3.8% |
Other Punctuation
| Value | Count | Frequency (%) |
| . | 23 | |
| , | 4 | 14.8% |
Connector Punctuation
| Value | Count | Frequency (%) |
| _ | 62 |
Space Separator
| Value | Count | Frequency (%) |
| 36 |
Dash Punctuation
| Value | Count | Frequency (%) |
| - | 19 |
Most occurring scripts
| Value | Count | Frequency (%) |
| Latin | 699 | |
| Common | 196 | 21.9% |
Most frequent character per script
Latin
| Value | Count | Frequency (%) |
| t | 90 | 12.9% |
| e | 64 | 9.2% |
| n | 54 | 7.7% |
| a | 53 | 7.6% |
| r | 44 | 6.3% |
| i | 36 | 5.2% |
| s | 32 | 4.6% |
| o | 32 | 4.6% |
| p | 29 | 4.1% |
| c | 25 | 3.6% |
| Other values (33) | 240 |
Common
| Value | Count | Frequency (%) |
| _ | 62 | |
| 36 | ||
| . | 23 | 11.7% |
| - | 19 | 9.7% |
| 2 | 14 | 7.1% |
| 1 | 8 | 4.1% |
| 0 | 7 | 3.6% |
| 6 | 6 | 3.1% |
| 3 | 5 | 2.6% |
| , | 4 | 2.0% |
| Other values (4) | 12 | 6.1% |
Most occurring blocks
| Value | Count | Frequency (%) |
| ASCII | 895 |
Most frequent character per block
ASCII
| Value | Count | Frequency (%) |
| t | 90 | 10.1% |
| e | 64 | 7.2% |
| _ | 62 | 6.9% |
| n | 54 | 6.0% |
| a | 53 | 5.9% |
| r | 44 | 4.9% |
| i | 36 | 4.0% |
| 36 | 4.0% | |
| s | 32 | 3.6% |
| o | 32 | 3.6% |
| Other values (47) | 392 |
source_doc_id
Categorical
HIGH CORRELATION UNIFORM UNIQUE
| Distinct | 19 |
|---|---|
| Distinct (%) | 100.0% |
| Missing | 0 |
| Missing (%) | 0.0% |
| Memory size | 1.4 KiB |
| 5e311e20-4bb | 1 |
|---|---|
| adc6e224-1ea | 1 |
| deed3ee1-dae | 1 |
| 9e6bfae4-7c2 | 1 |
| 0abeb1f8-b6c | 1 |
| Other values (14) |
Length
| Max length | 12 |
|---|---|
| Median length | 12 |
| Mean length | 12 |
| Min length | 12 |
Characters and Unicode
| Total characters | 228 |
|---|---|
| Distinct characters | 17 |
| Distinct categories | 3 ? |
| Distinct scripts | 2 ? |
| Distinct blocks | 1 ? |
The Unicode Standard assigns character properties to each code point, which can be used to analyse textual variables.
Unique
| Unique | 19 ? |
|---|---|
| Unique (%) | 100.0% |
Sample
| 1st row | 5e311e20-4bb |
|---|---|
| 2nd row | 016e8d29-288 |
| 3rd row | 07af2cf9-15a |
| 4th row | fed834b5-a04 |
| 5th row | aa279e3b-2d1 |
Common Values
| Value | Count | Frequency (%) |
| 5e311e20-4bb | 1 | 5.3% |
| adc6e224-1ea | 1 | 5.3% |
| deed3ee1-dae | 1 | 5.3% |
| 9e6bfae4-7c2 | 1 | 5.3% |
| 0abeb1f8-b6c | 1 | 5.3% |
| 04a90337-527 | 1 | 5.3% |
| 110b05be-f8d | 1 | 5.3% |
| 66f03e4f-bd9 | 1 | 5.3% |
| 3f42d484-d96 | 1 | 5.3% |
| 65105d7b-502 | 1 | 5.3% |
| Other values (9) | 9 |
Length
Histogram of lengths of the category
| Value | Count | Frequency (%) |
| 5e311e20-4bb | 1 | 5.3% |
| 016e8d29-288 | 1 | 5.3% |
| 07af2cf9-15a | 1 | 5.3% |
| fed834b5-a04 | 1 | 5.3% |
| aa279e3b-2d1 | 1 | 5.3% |
| 7a72cd85-984 | 1 | 5.3% |
| 3210a55b-6fd | 1 | 5.3% |
| 6adec8a8-d94 | 1 | 5.3% |
| 67f6cc9a-83c | 1 | 5.3% |
| 65105d7b-502 | 1 | 5.3% |
| Other values (9) | 9 |
Most occurring characters
| Value | Count | Frequency (%) |
| - | 19 | 8.3% |
| e | 18 | 7.9% |
| a | 18 | 7.9% |
| d | 16 | 7.0% |
| 2 | 15 | 6.6% |
| b | 14 | 6.1% |
| 0 | 14 | 6.1% |
| 1 | 13 | 5.7% |
| 6 | 13 | 5.7% |
| f | 12 | 5.3% |
| Other values (7) | 76 |
Most occurring categories
| Value | Count | Frequency (%) |
| Decimal Number | 121 | |
| Lowercase Letter | 88 | |
| Dash Punctuation | 19 | 8.3% |
Most frequent character per category
Decimal Number
| Value | Count | Frequency (%) |
| 2 | 15 | |
| 0 | 14 | |
| 1 | 13 | |
| 6 | 13 | |
| 5 | 12 | |
| 4 | 12 | |
| 8 | 12 | |
| 9 | 11 | |
| 3 | 10 | |
| 7 | 9 |
Lowercase Letter
| Value | Count | Frequency (%) |
| e | 18 | |
| a | 18 | |
| d | 16 | |
| b | 14 | |
| f | 12 | |
| c | 10 |
Dash Punctuation
| Value | Count | Frequency (%) |
| - | 19 |
Most occurring scripts
| Value | Count | Frequency (%) |
| Common | 140 | |
| Latin | 88 |
Most frequent character per script
Common
| Value | Count | Frequency (%) |
| - | 19 | |
| 2 | 15 | |
| 0 | 14 | |
| 1 | 13 | |
| 6 | 13 | |
| 5 | 12 | |
| 4 | 12 | |
| 8 | 12 | |
| 9 | 11 | |
| 3 | 10 |
Latin
| Value | Count | Frequency (%) |
| e | 18 | |
| a | 18 | |
| d | 16 | |
| b | 14 | |
| f | 12 | |
| c | 10 |
Most occurring blocks
| Value | Count | Frequency (%) |
| ASCII | 228 |
Most frequent character per block
ASCII
| Value | Count | Frequency (%) |
| - | 19 | 8.3% |
| e | 18 | 7.9% |
| a | 18 | 7.9% |
| d | 16 | 7.0% |
| 2 | 15 | 6.6% |
| b | 14 | 6.1% |
| 0 | 14 | 6.1% |
| 1 | 13 | 5.7% |
| 6 | 13 | 5.7% |
| f | 12 | 5.3% |
| Other values (7) | 76 |
source_doc_domain
Categorical
| Distinct | 9 |
|---|---|
| Distinct (%) | 47.4% |
| Missing | 0 |
| Missing (%) | 0.0% |
| Memory size | 1.3 KiB |
| Script | |
|---|---|
| ASR | |
| OCR | |
| OCR_academic_paper | |
| ASR_cleaned | |
| Other values (4) |
Length
| Max length | 18 |
|---|---|
| Median length | 14 |
| Mean length | 8.6842105 |
| Min length | 3 |
Characters and Unicode
| Total characters | 165 |
|---|---|
| Distinct characters | 21 |
| Distinct categories | 3 ? |
| Distinct scripts | 2 ? |
| Distinct blocks | 1 ? |
The Unicode Standard assigns character properties to each code point, which can be used to analyse textual variables.
Unique
| Unique | 4 ? |
|---|---|
| Unique (%) | 21.1% |
Sample
| 1st row | ASR |
|---|---|
| 2nd row | ASR |
| 3rd row | ASR |
| 4th row | ASR_cleaned |
| 5th row | ASR_cleaned |
Common Values
| Value | Count | Frequency (%) |
| Script | 4 | |
| ASR | 3 | |
| OCR | 3 | |
| OCR_academic_paper | 3 | |
| ASR_cleaned | 2 | |
| academic_paper | 1 | 5.3% |
| conversation | 1 | 5.3% |
| adversarial | 1 | 5.3% |
| literature | 1 | 5.3% |
Length
Histogram of lengths of the category
Common Values (Plot)
| Value | Count | Frequency (%) |
| script | 4 | |
| asr | 3 | |
| ocr | 3 | |
| ocr_academic_paper | 3 | |
| asr_cleaned | 2 | |
| academic_paper | 1 | 5.3% |
| conversation | 1 | 5.3% |
| adversarial | 1 | 5.3% |
| literature | 1 | 5.3% |
Most occurring characters
| Value | Count | Frequency (%) |
| a | 19 | |
| e | 16 | 9.7% |
| c | 15 | 9.1% |
| r | 13 | 7.9% |
| p | 12 | 7.3% |
| i | 11 | 6.7% |
| R | 11 | 6.7% |
| S | 9 | 5.5% |
| _ | 9 | 5.5% |
| d | 7 | 4.2% |
| Other values (11) | 43 |
Most occurring categories
| Value | Count | Frequency (%) |
| Lowercase Letter | 119 | |
| Uppercase Letter | 37 | 22.4% |
| Connector Punctuation | 9 | 5.5% |
Most frequent character per category
Lowercase Letter
| Value | Count | Frequency (%) |
| a | 19 | |
| e | 16 | |
| c | 15 | |
| r | 13 | |
| p | 12 | |
| i | 11 | |
| d | 7 | 5.9% |
| t | 7 | 5.9% |
| m | 4 | 3.4% |
| l | 4 | 3.4% |
| Other values (5) | 11 |
Uppercase Letter
| Value | Count | Frequency (%) |
| R | 11 | |
| S | 9 | |
| C | 6 | |
| O | 6 | |
| A | 5 |
Connector Punctuation
| Value | Count | Frequency (%) |
| _ | 9 |
Most occurring scripts
| Value | Count | Frequency (%) |
| Latin | 156 | |
| Common | 9 | 5.5% |
Most frequent character per script
Latin
| Value | Count | Frequency (%) |
| a | 19 | |
| e | 16 | |
| c | 15 | |
| r | 13 | 8.3% |
| p | 12 | 7.7% |
| i | 11 | 7.1% |
| R | 11 | 7.1% |
| S | 9 | 5.8% |
| d | 7 | 4.5% |
| t | 7 | 4.5% |
| Other values (10) | 36 |
Common
| Value | Count | Frequency (%) |
| _ | 9 |
Most occurring blocks
| Value | Count | Frequency (%) |
| ASCII | 165 |
Most frequent character per block
ASCII
| Value | Count | Frequency (%) |
| a | 19 | |
| e | 16 | 9.7% |
| c | 15 | 9.1% |
| r | 13 | 7.9% |
| p | 12 | 7.3% |
| i | 11 | 6.7% |
| R | 11 | 6.7% |
| S | 9 | 5.5% |
| _ | 9 | 5.5% |
| d | 7 | 4.2% |
| Other values (11) | 43 |
document_text
Categorical
HIGH CORRELATION UNIFORM UNIQUE
| Distinct | 19 |
|---|---|
| Distinct (%) | 100.0% |
| Missing | 0 |
| Missing (%) | 0.0% |
| Memory size | 1.3 MiB |
| Welcome everyone! This is natural Language processing for Law and Social Science. Thanks for joining remotely today. It still is a bit up in the air how we will do the hybrid verses in person versus Zum format. This term, you hear, I'm a little stuffy today. This is true. Avoid nineteen case I Caught it from my daughter who caught it in daycare. It's very mild so I hope if any of you catch it, it's not worse than this. It's definitely manageable. You can see I'm here to teach even though I have it so it's not that bad. I'm a little congested and we might end a bit early today. I Hope that's all right, but going forward. I Would like to do the course hybrid where we have some impression meetings at least, but before text money you'll hear from me about that. Thank you for all of you who filled in the students survey. There was a broad agreement that we should have an online aspect or at least should be recorded so well. we will work with that. So I have a few slides to introduce the course and then we'll have a chance to answer any sex questions about the format. So this course is a applied Natural Language Processing. It's not a course where we will just start with different texts, data tools, or different help models and learn how to come them up in there. We care just as much about the applications of those methods in the law and in social science and this is in the news all the time. Here is an example from a recent legal product called Clarity which uses in all tools to to analyse contracts and for example terms of use to highlight different clauses that are unusual. You also hear these really exciting ideas such as the World's First Robot Lawyer I'm excited about this, I think everybody should you. I think that this technology is is improving rapidly and dramatically and there is scope for many interesting things happening in the law and inoup world. But there also is a lot of hype and we will take a skeptical view of of these strong statements such as the World's First Robot Lawyer And in turn I Think that while there is a lot of interesting about tools coming about for law and social science and other applications, I Do not think we're close to having a judge to be replaced by a contact. some other reasons to be skeptical or to be concerned about the arrival of these legal inopetuls is that they can be biased So northwest. One of the classic examples is different languages have different degrees of being rendered having notions of gender, mail and female in the language and if you translate from English such as she is a doctor he is a nurse to Turkish which does not have notions of gender pronouwns and then you translate back the gender switches so basically they they have since fixed to this in google translate but it used to be where if you to see as a doctor translated to Turkish and then translated it back it will change to him as a doctor just because similarly he as a nurse would be transformed she as a nurse and this is just because. Theories this basis this statistical correlation in language where doctors tend to be male and nurses tend to be female and statistical language models and translation systems will capture that bias. These issues are based. the language models are as the technology comes more powerful, these issues become more intense to get more intense benefits but also more more intense risks and good. It is now a few years a few years old but this is language whose hole it came out in the Tousadand nineteen that could. It was basically among many other things a fake news production engine and it could produce a lot of context appropriate prose. So you can imagine to know Twitter and email. can the news being filled with all this machine produced speech that would drown out all other speech and I think that those those concerns are still relevant. but now that got two has been out for it for three years and there's an even better version called just There that's been out for a year and we have not seen the internet employee. That means that maybe it was not as bad as we thought and so in this course we want to know. Can we take legal Gptto illegal? Get there to help judges in their work? So this is the course. It's natural Language processing for law and Social science and our engineering goals are doing these kind of two pieces. This course we're going to develop skills in applied natural language processing which will include machine analysis, interpretation, generation of documents and those could be on news articles or contracts or judicial opinions or political speeches. And we want to also take a social science approach where we do not just care about sequencing language in different ways, we care about relating it to attend data, and to understand the social forces that underlie these documents. What are their determinations and what are their outcomes and so you knowsome. Of the examples that we will frequently come back to are: what are the motivations for judges? Can we analyze judicial opinions and see what's driving their decisions? Can we look at the writings and the speeches of politicians and to the end where they're coming from And I Think this is kind of the broader agenda or the goal in this research area is Knpowders language matter in society in human relationships s and what can help do to help understand that? So what we will do. We're going to read text documents as data so there's you know many ways to do this and will go over many of them. We will use supervise learning techniques for dimension reduction, topic modeling, groups interpretation, supervise learning for text regression and text classification can be predict from a speech. Is this form a left wing politician or a right wing politician will get at World embeddings, document embeddings, a lot of exciting technologies there for producing these learned representations of language of words and concepts in geometric spaces. and towards the end will get into disclosure analytics. So this is where the linguistic side of natural language processing, cynicism, and a summarization question answering I'm checking. These are all really relevant to legal applications for example. So some course logistics or beating times that will be two even fourteen to it in an sixteen so we'll have a ten minute break in the middle going back to what I started mentioning at the beginning. These are the survey results for the the course format and there were only a handful of you who would be register if there only online and everybody else wanted some online aspect or the indifferent. The based on these surveys we will certainly have a online component. like everything in the class will be durable online. but I'm hoping that we can have some impression component as well. so there's a question in the chat I have the the chat year so I'm sorry but it in general how keep track of the chat so you can always ask questions three were asked to. We need to have knowledge about law I said are to be a good in the class. The answers no no not at all so you do not to have any knowledge of it, you need to open to learning about it. So if you have some interest in social science applications or legal applications of help it will make the class much more enjoyable. but there will be no substantive aspects of health or social science that will be tested. and so given that we will have this important online component to the class I Want to make the most of the hybrid learning. The lectures will be recorded by but in some view that contacted me about different special cases which is fine but if you can it's going to make the class more fun and and more better for everyone if everybody comes in if you're going to be absent let me or the tea now and if you have questions or comments from online you can just type them in the chat as you as the doing or you can use the the raise and front in which a no monitor so help asks are people who either know Pytha nor beeligible for this course so aim going to get to that in a little bit. But the short answer is if you've never seen Python I do not recommend taking this course, it will be too difficult. I mean you can try to stay for the first two weeks and see if it's manageable for you. but in previous versions of the class, people who were new to Python and people who had never done any text analysis it was frustrating for them. and so I do not recommend the course for anyone who's sure asked and well tell you that as some emails if you'regoing to be absent for a lecture, a email email after the tea to let her know and if you have to miss a significant number of courses the email of cause you might have to do an additional assignment. So ya so relax. If you're anyone who is kind of the new to Python or has not done any ex data and turnout sure let me know you can talk about it so avoid asks and can homework only be permitted in Python or can we draw also try to this in or sure yeah you're welcome to try it in our for me I should. I wouldbe great if if anyone is interested in converting the course materials to war that would actually be a great course project so we can arrange to get extra credit for that report. asks what's the do registration deadline? There is not an official one I'm not sure exactly. I think it's varies a bit by department but I do not have an official de registration that line. If you're going to be just for for our grading purposes, it's better to do it before the first response essay which is like six weeks in five or six weeks in because others take a long time for grading and so I would rather you deregister before that. So I would say I think by five or six weeks you should know whether you will stay or not. So smart. Asks if we attend the lecturers and only watch the videos. there will be additional projects yes, so mandatory. The live attendance is mandatory and so if you're going to just watch the videos then you have to do another assignment. but I have not decided what that is yet. Okay so yes, so this is related to newly keep track of course participation through in class activities. So young asks, do you recommend someone who also general machine her knowledge but just to experience with help. If you're actually pretty comfortable machine learning with Python then this course actually wopolity Fine. So if I think that if you're doing the first two assignments, the first two home work assignments and they're not taking you a long time to do if you can finish them within of hours then then your on track. but it mainly do not recommend it. I mean it, if you're quite new to Python then I do not recommend it if you have some machine learning experience than that's good. But as I said some text analysis or snap experiences is recommended. So we have course syllabus I've already sent by email and I be in oppose to league to it again so also asks why this course worked for people who intend buillier of it judge course at so if you took my course in the spring and you l in off if you've done if you finish the assignments of the course in the spring then this course will find freedom so there's there's a little bit of overall. So I would say that he saw in the fall course it was say in the fall course ably report judge it would be fine as a prerequisite for this course. If you've done that then this should be fine for you. So those links are a bit is going to change to it screenshare to the syllabus so I just pose I did a link to this in the home Here's my contact details. Here's area's contact details: the lecture schedule that sessions which I'll cook you a bit in a second but those are at ten a man on Fridays they're not mandatory but these will. They also be recorded and Afro will go over the example coding notebook for the week and then also go over the previous week's homework. This is our daughter's the structure, All the slides including Iardaploi today slides here there in the slides thotfolder notebooks. These are the simple notebooks for learning the material and so before before doing the assignment. You should read the notebooks so you can see. You can kind of skim though you can read through these, ensure you understand them and everything that's in the homework which is here under homework. The homeowners will follow will have similar content to what's in the notebook so you can see we fill in part of the notebook and then you have to add something in a new column text which contains the lower case, title and lead. So here's lead. here's title and so No Nights is an example and here you can just like to nowtype lower so thiswill be how to get lower case. So this is just to show you that these the notebooks and the homework are designed off for self motivated learning of all the coding as aspects of the course. so find asked how do the homers omissions work? So there's there's a homework every week and so it's like this: homework Here you download this Jupiter notebook and fill it out and finish it and then you upload it to add you that it's not working yet but it's going to be on the course model. There's a submission system or using coal ufous up load it on the website and going to be due. The homework are done on thousands but the first homework is done next. Thursday So I can not actually show you if you scroll down so everything you need is going to be highlighted here. So for example, do this week. next week the homework one is done on Thursday fin is that what you were asking? I'm going to come back to that as let me know if you have some other questions about it. So here's me. I'm going to put in the correct system still working on this but camera acts are all homework mandatory if you want it mean you lose point if you do not do them but they are. The homework are a completion grade so you know we're not grading them. We're not grading all the answers but if will check that you did like you tried to do every piece. and if you say you get full credit and basically so in terms of the grading it's thirty percent for the programming homeowners and I do not go eleven homework so mistake. three points per homework or that's thirty percent and so for every lecture will have the slides. I'm going to post the links to the recordings here so like after it today, you'll be able to get the recording for for everyone here. there's going to be a tea session on free, there will be a recording link heiress about. So unique asks what can we think, what the response essays are. Can you be a bit more specific? like do you want to see an example? Okay, well get to that. We'll get to that next week. It may be the We attributes. You do not have to do one of those for a time until a month from now. time is talking about the response essays. Whether's some information here I'll provide some example response essays from previous previous years, but it was not going to get into that into detail today because it's a more complex topic but you can read about them here. But basically what it is is reading a paper one of these papers in writing a response as I about it. Like a review here, I have a link to the bibliography of references. So these are all of like the materials that the slides are based on so you do not. Someone of these are required readings but it's worth skimming through this just to see where things come from. and if you are interested and you want to to go back to add to fill in your knowledge from the slides then you can read these the other. the other required assignment is there is going to be there required readings for example in week for you have to read one of these papers and then we'll do an inner class activity about them but it's going to be. We will form groups and do short presentations to each other about these press but I'm going to provide more information to that in week there before we do that. So the the three pieces of the assessment are the homework on the coding homework which I showed you the response essays which I mentioned or reading a paper and writing a report about it and in third there's a end of there's an end of the course assignment and its in the you So we would call them an exam but I think here you would just say it's an end of course assignment where you have a few days to do an assign. For those of you who were in my class in the fall you know this is like it's a questionbasicly a question about every like sure in some questions about the required readings and so that the end of the course assignment is one the things that we will cover in the lecture are covered their of sotthat's how the course will be assessed I'm going to cover some of that information again now just in the slides so it mentioned awards the is the first that session will be on fairly area's here as well. After do will introduce yourself sure here one man after I'm a packed student at an centre and I hope to see you in the first session. So in these is sessions it's what would expect far will go over the notebooks they code note books from the bathtub and then the last week's homework after you've submitted it and then usually there will be some time left over an area can turn the recorder off and you can ask some office hours time questions. I'm going to pose course announcements on Model and if you were registered for the course this morning you should have gotten an email about that if you did not send me a note after class and so we can try to figure out your muddle registration. it's not ready yet but I'm going to work with airfare to post it but we will have this to in a forum on model and so you can post questions there before next week's class and I'll go over them at the beginning of class or I'll just answer on the model. So I wanted to make a note about the course work load because this is not like other science and Perspectives classes like it's not much work to the extent that I've gotten negative. I mean I just want to say expectations I have got a negative feedback on the class because people thought it was too much work and so the thing is, it's actually a lot of work for me too because I have degraded the response essays. So it's actually easier if there's fewer students. So if if you're worried about the course load, then there's other classes you can take that do not take as much time, but according to it, would increase. The number of credit points at I is not the credit system is the maximum for a Science and Perspectives course, but the number of hours for most people If you satisfied the course prerequisites such as having some Phantom background and a little bit of blip background. the amount of work in this course is less than ninety hours. And so it's twelve lectures, eleven programming assignments. There required readings to response essays, and then the final assignment. So that's about sixty hours of time just actually doing these things. And so that includes three more hours. So that includes the tea sessions and then study time if you are new to pythem especially, but if you're new to help then it will take longer. So I just wanted I Want to say expectations about that beforehand? Also, if you were interested in this topic of applied Up for Social science then I would highly recommend you also sign up for the two additional credits for the course project so we'll talk about that more after class next week. So if your interested in it, just stay after it you. This is simply recommended for people who might want to do graduate research after because the previous course projects have turned into conference and journal publications. two of the projects were part of into Swiss startups as well. So if your interested in legal tracker or other entrepreneurial projects based on applied help then the course project could be interesting for you so then asked one where doing for the submission of the project. there's there's a link from the syllabus on course projects that has the rough deadlines. Basically you need you haveyouhave to pick a topic within the next month and then have an outline within the next month and then the full draft of the paper is then day until remember September first so you can work on it over are so a related system of what we've talked about already. Thank you to everybody who filled out the course survey. if you registered since I said this out, send me a note, email me a no because it send you a link to the course survey. Oab'll just send out another link so you can fill it out as well if be curious who else has joined. It's about half master students and few old students and then the rest bachelor students and mostly computer science some data science. He's actually zero people from law which is somebody asked do we need substantive law background so if we did not, we would lose all these students. So we do not require that so that two you guys are So I Already went through this a bit in the syllabus, but the required readings are indicated in the syllabus schedule. In addition, there's the bibliography of references that has additional readings if you want to complement the slides in the link related to the response essays. there's the list of applications papers for response essays, which will talk about more next to be. So I wanted to just give a quick outline of some of the books that were mentioned in the references list. Again, none of these are required readings, but for those who want a deeper understanding, I do coming these books. So Natural Language Processing with Perception is the book that accompanies the Natural Language Tooloquate, which is just this classic blip trouble with kind of more standard classical like old school machine, old school natural language wing tools. If you want to learn machine learning, this is my favorite book for earmachine learning with Physicist learn and wood courses and Monster Flow. It's more generic and's not about Inop specifically, but there are a lot of top applications and for those of you in my course in the Fall you would have already seen this and this is available on oil through the Earth Library. you should be able to get it. This is a free book if you're interested in more of the theory and guess for natural language processing more product than mathematical formalization. If you want to do graduate work research in Blip, then I really recommend this book the Your Goldberg book. I think this is available for download as well on the Earth Libraries. If you can not find it, let me know I can get you a Pdf even though it came out in to this and seventeen. It's actually a little bit out of date unfortunately, so it basically has everything up until Transformers, which as we'll talk about have kind of remade inilp. but a lot of the issues and approaches here are still quite good. Another kind of classic textbook is necessary in Martin. Its kind of more than the does not really focus on neural nets inalp and is kind of more than the older generation of help research, but it's very good for some of the more linguistics oriented in semantics oriented part of the course, so this came up a bit already. Python is a course prerequisite see here for the example notebooks and you know I'm sure many of you as I said, Python is a country register, so you should already have it set up on fairy affairs can provide some help in setting things up, but we would trust everybody. Everybody should be able to get their own another environment running. As a prerequisite to this course. these are the main piping packages that we will be using. As I mentioned in all to is this broad collection of older Inalp tools. Finish is great for topic models and award embedding. Spicy is another kind of generic tool. It's great for named in any recognition, parsing in reference resolution, things like this as well as a library of pre trade world factors. and then as it mentioned this new inilp this new neural net architecture called Transformers in particular large pre train transformer models that are trained on large corporate these have really remade how help is done and hugging base transformers as the hugging base system is the standard for that. To provide an overview on the course, here are your objectives: seventy percent if you want to learn how to use help tools and fifty their parents if you want to learn how to apply opinion tools for law and social science so this is great. We're going to do both in this course which are the followings best Matches your goals for learning in top: sixty percent want to learn it for Engineering in Software development Thirty seven percent for social science research and fifty three percent for computer science research. This is good. We're going to be doing all three of these goals are going to be covered in this course so avoid asks if we need to into processor to no no and maybe you're asking if you know like at you for the assignments. The short answer is no and you do not need any special computer equipment the yeah so we you should be able to the examples on the books and assignments. We use kind of small corporate things so you do not need you do not need any specialized competition for that. If you have problems you can use a Google collaps right and so Afro will cover that in the tea. sure you can just those Google could for everything. So why this course right now we live in this really amazing time I Think for language processing where with our lifetimes there's been these new social structures that have remade the data landscape. the Internet Social Media digit join efforts by governments and Google Books for example just as amazing amazing initiatives for digitizing tons of text at the same time as having these huge crops are. We also have had this amazing increase in computational resources as from cheap disease to efficient databases, solutions for quarrying all of those corporate and then having cups give rise to go Pus and then also tips for training these gigantic volunteers and in particular for natural language analysis. We have these really interesting tools in machine learning, a blip and casual inference for the legal and the social science applications of these tools. And for me I Think it's fair to say that at least from a research perspective a lot of as these trends are especially amplified in the law and Legal language. Political Science and Political Language Here many doors that are being opened in these old fields by these new technologies and so we care about legal and political institutions such as what judges write in their opinions, what politicians say speeches, what's written in patents or in newspaper articles about the law or in legislation and regulations. Those are all millions of lines or millions of documents of unstructured texts. and there's no way that humans could read them even if they wanted to add. So this is why bring in these tools for computers to read them is so exciting. So manual asks, could you share the response to the questionable students background acknowledging presence is up. I do not have that in the slides but if all talk about that next week. I don' think there's anything not notable from that. or do have a specific question manual right? All talk about that a bit next to be but but you do not need to worry about that. So here's an outline of the course and actually I would say let's will all just go through this and they will take them break. So I know this is like an eyefool and I made this last year, but we're actually going to follow basically the same format and justice. but you can visualize everything that we're going to learn in this chorus from this gap. And you what? what we're starting with as raw data today. and next week we go from raw data to segmented documents that are pre processed in different ways. And once you have these documents, you can only use these in some social science analysis just to say oh well, how long are the documents you know? How many bonus do they use? What's the word link to the sentence Link This public a measure of reliability or sophistication in language. The second would be dictionary accounts. and I Think if you're a example researcher, a computer scientist, the fact that you should just go and count different words and count the number of the times the word good shows up as a measure of sentiment that just seems so primitive. it's like the stoneage it. But it I think that we should consider those models cape seriously and I'll give you a good reason at the end of today why dictionary methods are are not to be ignored. And so next week we'll get into tocanization. So the different ways that documents are split up in to sendances and words and looking at part of speech things like this. Once we have documents as these end up being teprimitives for all the models that we will analyse including in gram. So that's converting tokens to phrases. Topic models that's converting documents to distributions over topics so you can imagine in illegal groups there's a crime in contracts on tutors and patterns and things. Each stations are left wing politician I Think my internet might be unstable but I'll give it a second. Can you go hear me now? Can you put it in the cash? I back market thank you So think asks do do we take a look at how he's methods roughly work or do we may learn to use them or what were weregoing to do both rateboth so we will in the notebooks in homework. In the tax sessions we're going to be learning how to do these things in Python we will implement them, but in the lectures were going to be focusing on whether's the purpose, how whatever, going to help us understand the judges or the lawyers things and's so after machine learning will get into neural nets and a particular if we'll learn about word embendings which is a way to represent language in a low dimensional space we'll get into passing, getting at syntax, the relationship between subjects and objects, agents and patients. This is getting into linguistic sides things. We will then get into Transformers and that part of the course which will lead to are ample language modeling knowledge graph's entitlement. So this is getting into asking a computer does a empty public, does sentence A empty B or unknown We do information and extra going to extract relations from a corpus and learn about what the corpus is about and towards the end will be getting into these of more global semantic aspects of summarization. Question answer, automatic claim checking, casual inference from documents, identifying casual relations in documents and a lot of this gets way past that social scientists are using right now. But I think these technologies are improving rapidly in the case of the legal domain at least the there going to be clear applications that really have not been done yet but or will be running right to the frontier in this course to take a look at this if you like will be following us as long as we go in the course. Okay so I know we've already gotten abunchab logistical questions but I wanted to take break now and give everybody a chance to ask a question and then we'll take a quick coffee break. So are there any questions at the moment that have not been covered? You can put them in a chat or the race hand function so manual ask how relevant is reinforcement learning to snap. That's a very interesting question You next's not universal, but it is certainly relevant. There are many very interesting reinforcement learning applications help for example the recent paper that cannot used as reinforcement learning to to improve the quality of summaries and I have a paper with a old student which actually came out of this course using reinforcement learning for attractive summarization. So if you understand reinforcement learning, there's a lot of interesting applications and I can provide some more resources on that. So so fun. Also, memory can note of that area. can you make it note of that? They set up environment script that must have been something from last year so we'll fix that thank you find so report access. Is it possible to do the course this semester near the party next year? Sure it the mean things change right but I'm planning to teach the course again next year so should be fine. Thank you for asking that. So think asks on right, think is asking. theatre's not a simple yes or no. answered to that question Sometimes yes he mean it depends on it mean we're not. We're not going to be like trying to implement them and see Plus Pals or something so but you knew will be talking about you know will have some review of Nstarcastic Radiant Descent. You know how volunteers learn. You know it is not as it said it and this is not a substitute for machine learning in our Pop course so we will do some. but if you need that then you had take a different course or take both of right? So we're going to take a breaknow and will return in nine minutes at there fifteen. I'm also going to turn the recorder off now So if you want to ask a question the recorder please do of really's We really started the content of the course for the remainder of today so nfuzum in on this on the course outline We get to the Corpera. so text data is a sequence of characters called documents and the set of documents is the corpus which we will call them in this course. And what makes text data challenging but also interesting is that it's structured. It's just this stream of characters and the information that we want is induced in the symbolic information in the those characters and it's mixed up with a lot of information that we do not need for any and task and so a lot of what we're going to do in this course is its information extraction or if you prefer information disposal. we're trying to get rid of the information we do not need and will be the main focus of what will do next week and a week after. And to recap something that was mentioned earlier, this course is appealing in place and we do not care about that much of the documents by themselves. What we care about is related into better data and that could even just be like the time Theatre document is published. So we might say well, how syntimate towards a impose social group, How a sintimate towards immigrants changed over time And we can. we can. make a systematic measure toward immigrants and end show that How that's evolved over the last ten years less one hundred years. And the important point there is that we have met a data on the time and so just to say that it be more specifically new might start off with some positive negative syntimate capital by and judicial opinions. And that by itself is not that interesting for a lawyer or for such a scientist. But what if we had the dynamite in opinion is by Judge J at time It and so we will often use this type of notation with these subscripts for the indexing corresponding to the meta data that the document corresponds to. And we can say you know how to sintimate very over time. Or let's say we have the information on the political party of the judge are they do. They have more negative sentiment towards defendants from groups Go. So let's say that go is a dummy variable enjoying one for cases where the defendant is an immigrant and so then we can is information in our data set to say. Well, the right wing judges. They have more negative sentiment towards immigrants for example and you can see that one you relate the text features to meet data. There's many interesting questions that one can answer and a precursor question to you. This type of analysis is what's the document. So now often have this big data set. If we might have a bunch of books on what do we do with them, we need to zoom input to some degree to find out which document, which observation is useful for our meditative variation. And so you know if if we do not want to just arbitrarily make the documents really small because they do not, they do not help us answer a research question such as you know our republican judge is more negative towards of immigrants. If we made a data seat at the sentence level for that, the sentence would both abstinence. data would be super high dimensional, but we would not have any additional variation comparing the right wing judges. did the left win judges. And so this is going to be one of our first Islands activities. what should we use as a document in these contexts? So I Want to take about five minutes six minute to answer these questions? We're going to set up them, going to set up breakout groups to put you guys into pairs and this is just a way too please pull up the slides on your own computer because you're going to lose it when we stop sharing and I'm going to put everybody into pairs and this is just a way for us to start to know each other. So you going to be. You're in a breakout room of two or three people. introduce yourself and said that your major and what you are interested in the chorus and answer these three questions together. What counters the document and were is going to open the breakout rooms for six minutes and will be back at Fifteen Twenty seven so only so handle in corporate. So we have some examples in the notebook about some breed processing to especially if our working on a project not just in this course but you knlater on in your career in life. It's good to think about from given questions or given to ask the data sets out there that have been compiled and so far for court cases like in the U is for example court listeners good but in social media there is really excellent data from Twitter and Facebook for example. for Red it is also for Wikipedia All these data that's are really useful for such a science analysis. This is not in part of the course necessarily, but you know it will be useful for you later on to learn about queueing ships running web scalpers doing processing to remove home markup and there is also the hugging face hub and hugging face system. They have a lot of amazing data sets so it's good to just kind of produce through that ski that a bit can fu access should learn or should have learnt so it would say it mean that should learn because it do not be tested in this course but it will help you to know it for you for other things. So I recommend learning it but you can do it after that. We do it in the summer so all everything that we talk about this course is kind of language agnostic I'm a native English speaker so everything is going to be focused on English. but for your course projects and things you're welcome to do things in other languages. After one of area's special teas is multilingual implants so she can help you get started with it. And there's also just really great machine translation So Elevthos asks why should we use Cyllinium is not cycle evercovtalate that it depends on what what you trying to do then I think that doing a webscraper with that you are a Lib is like a nice start. but if you keep going on up for social science for data science then we will come to a point where you will need a crime driver or something on those lines. So how can we use the quantity of text as data So you know most of this course is going to be about the semantic or conceptual or symbolic content of language. but there is also interesting things to be learned just from the service features just the statistics in the characters in documents and one thing that me and old students in my group had looked at is how the style of writing changes over the life cycle for judges. and one of odd or curious thing we found is that older judges that use shorter words but longer sentences and so whether this is like better or worse writing thing is kind of subjective but it shows that there is. It seems like there is this of biological component of writing style for judges. Another relevant debate in the case of legal quantity of text, the law is on legal complexity where this depends on what country you're from but like in the U's and Finance for example they're always complaining about the law being too complex on but using help we can actually try turn this into an empirical question and ask how complex is the law and certain bedroom which is one of the applications. Papers linked to the syllabus is about measuring the complexity across different areas of the legal code and they produce the number of words in a code title which is a section of the code and they also produce a word invite measure which is basically it's the tiny of the probability distribution across words and what they show is that Public health and the Tax code for example is very complex. It has a a lot of words in it but if you look at the codes if you look at the code titles that have high word intern so a more diverse vocabulary it's quite different and so you. The Commerce in Trade or Public Health and Welfare scores highly on both measures of complexity and so from a tax lawyer perspective. This is kind of interesting because it shows that even though the Internal Revenue Code is complex in terms the number of words, it's very repetitive so it might not be as complex as if sounds. So the next set of methods that will talk about today our dictionary based method which is counting words or patterns and so in the notebook we show some examples of using regular expressions which is this really powerful string matching system and this is going to be depending on what your research, question or engineering task is how you would want to to do this. So one theme or question in the law or judicial behaviour is do judges? Are they making decisions based on justice or morality or are they doing it based on cost and benefit in analysis or efficiency? And so you can imagine using a dictionary method to go through to what different judges say and say, do they say justice demands or do they say efficiency demands And so a dictionary based method could help you get it that. There are also general dictionaries such as Wardnet and Luck which will talk to you about in a second. One example from Economics and this is also one of the applications readings is Baker Bloom and Divas where they wanted to know what is the amount of uncertainty in the macro economy and this is like if you're a Mcroeconomister, you're going to finance a big beam of when there's more uncertainty in the markets, they're more volatile. That could be costly. And so they built this data set of newspapers and they use this kind of simple dictionary method. Does the article that the word uncertain does it have a word related to the economy and then does it have some words related to the government and then they ploddedin that and this is actually just the main result in the paper is just showing that this policy of uncertainty index based on the newspapers, it spikes during these kind of stock market shocks. So like Black Mundane or the Wars elections, nine even. This is the financial crisis to two thousands and nine the Euro Death Sealing crisis. And so you can see that these kind of intuitive moments of market uncertainty are captured by this newspaper based measure. There are are some problems with that to be sure if you're curious about using this in like financial applications is recommend to keep that all paper. another set more fundamental dictionary methods are are available in World That which is this nice python package with a database of it's a big dictionary with all the words in English and how it related to each other and so you can see for example, all all the different senses of the word bass or bass are located here. so it's like music or voices or the fish. There's many meanings of the word base or bass and the the word that it captures these synonym sets of groups of beer anonymous and has information on the anthiponyms and also parts and holes and then also all towns are organized in this categorical hierarchy and so you could have employers which are the higher word in a category and then symptoms are the members of a category and so you. There's many ways this could be used, so if you want to do definition reduction, you could replace words by their hypernem for example. And if you keep on going up Twilight of Categories word that has twenty six lexacer, gray graphic suppresses and so like action, animal partifact person relations shape things like this. I think it's pretty cool that these linguists have gone together and really tried to organize all of language in this and it's all automated now and a Python and so they did the same thing for verbs as well. So this is for bonus and this is for verbs and so you could take a corpus in, say, well, which books are mostly about perception versus possession for example. and you can imagine there's a kind of digital humanities or cultural analytical types applications for these methods. Some other general dictionaries which are relevant for us include lists of function words so thinking of words like four rather than also these get at the style features of text so most of them have you going to drop them because they're not going be very useful for topics, but they can be used to get it nonpolitical dimensions. So if you want to identify authors of texts without that being correlated with the topics, then using software is a good way to do like. or luck is it's kind of the standard licenses for linguistics and Social Psychology and the like team. They have seventy lists of word's of category relevant words including the commotion, cognition, family, positive, negative. We will see some applications using Luck later on in the course and in more recently there is these big lists of words that are initiated by people on crowd scouring platforms. So Mohammed and turns have joy and sadness, anger, fear, trust, disgust, anticipation, surprise and can warmer at all. They could fourteen thousand words along violence, arousal to dominance dimension. So these last two on kind of emotional content. Those are part of this broader set of tools in blip on sentiment analysis and this. You'll see this in many papers, but also in an industry like in advertising digital marketing, they really care about determine right for obvious reasons and we want to know for a given like a review of a movie. Is it positive, negative or neutral And the standard approach could be licensed. base research for the word good, but it's easy to break that right. Like what if somebody said though the movie was not good or the movie was not very good and so just like amends other words totally transforms the meaning. and so it means that just counting words often is not going to work very well. And the more moderate approach is to use pre trained transformer based syntimate models in area I Think is going to add an example of this into this week's notebook to show how you can download a pre trained sentiment analysis from the Hugging Faced Model hub and apply that to documents and you should be careful about these off the shelf scores through these are trained models because they might be trained on a corpus that's quite different from yours. So if you take a corpus that's like initiated tweets and then you apply it key contracts, that problem will work right and so you have to be careful about that. Check that it works in your new domain and there is also some methods which will get too in the world embeddings for making some domain specific licences so you can start off with some positive negative words and then you might find that in read it. the positive words are different than the ones in Instagram for example. So I wanted to just point out a specific issue with syntimate awareness that are based on initiated data. So if you take a syntimate analysis like from hugging bass or something and this is from from some slides by crime car I saw on his Twitter feed. so I do not know exactly which model this is but he made this where if you just take let's go get initial food versus let's go get medical food this has a very positive sentiment and this has a low sentiment. This is bit so just changing the word from relation to Mexican but just that soon as by itself the sentiments exactly the same right? but they changed it from relation to medical and the sentiment went from positive to almost negative and then this is an even kind of more striking example. If you say my name is email, you get this positive sentiment. but if you say my name is unique while you get negative sentiment and so this is like really kind of striking and important and concerning issue in all Pi systems and you want to ask the mean, Is this sentimental model racist? What's what's going on here? How did that happen? And this shows how you have to be careful about using symptomatic scores that are trained on initiated data sets because they're also going to be learning this correlated information in the data set. so that's not unsintimate, but then it will learn it and apply it when you use it in a new domain. And so this is just part of this broader problem or challenge or issue in an Apple for social science. But also this is going to be relevant to many things for products as well that we care about. some true sentiment in language. It but what we get is a systematic scorer. and the model that we trained to learn a sentimate score from language has all these confounding factors. So you nonwhite examples for medical food versus Italian food. You can kind of see how that might have happened right where initial restaurants. maybe they tend to be more of scale or like thecritics like them more. and so because the language model or the competent classifies trained on these biased data sets. that food this, let's say, food critics like Italian food, then that gets baked into the intimate model even though it has nothing to do with even though it's using these words that are syntimate neutral. So this is not easy to fix me. You know, because there's not going to be, you can not just make data set that's neutral, like every data set going to have some biases in it. And so I Think that trying to fix it is this really important and interesting area of upon research that's happening and and this is not unique to determine either. This is a universal problem that I want you guys to be thinking about Throughout this whole chorus is that we are trying to measure this true quantity in language, trying to extract it to learn about science or to learn about to solve a legal task to make a prediction and whizzl we care about. But we get this. We can only make a measurement of its indirect measurement and it's going to be confounded by other language features and sometimes non language features like where it comes from or the large age or style and supervised models are just by construction how they're built. they learn features that are correlated with the label being initiated and unsupervised models you in a topic model or world embodying there also going to learn those correlations and so you like. A classic example is like the similarity between the word banana and the word yellow is actually pretty low. but the similarity between the word banana and the word green is actually really high. And it's not because bananas are green, but it's because if a banana is yellow you we just never mention it right and so you have to be very. This is just some examples of these these problems or limitations with these language models you have to be careful about when you're interpreting their their outputs. But and this is what I mentioned at the beginning dictionary methods. They do have these other limitations But They very significantly mitigate this problem. And because the researcher is familiar with the context, they know what would the words mean and so they can always regularize out any serious surroundings. And so if I'm like trying to make a sentiment analysis and my model tells me that captain means high positive, I can easily fix that with a dictionary method, right? And so this is like a defense for dictionary methods. potentially. And I think it's why economists in particular and just other empirical social scientists. they might still use dictionary methods because of this reason. And I mean they have these other limitations which you you can not. Those are difficult to deal with. but we have to be careful with these blackbox models. Okay, so let's wrap up so the first assignment is already put on the gatehub as well as the coding example that Afro will go over on. Friday. So this came up. We explained it a little bit earlier actually, so I'm just going I'm going to stop for a second and answer elithereosthis question sorry is missed that earlier Those elyptherias asked are those general dictionaries useful any more since we can not easily measure the similarly between words. and also such dictionaries require a lot of human labor to be kept up to date. So I think that's an important point. So I mean the general dictionaries. They are built up over time by these research groups. but I think they have limitations and I mean I think they should be updated over time but meant I think of now all the methods that we'll talk about in this class in terms of dictionaries and classified things. They're all going to have kind of prose and cons and we's want to identify those problems and think of ways to address them. So the way that time the timeline will be in this course about the coding practice and the homework is that for last week to so like week one, the notebook, the coding notebook for week it is got to be reviewed in the tea session on fairy week it and the home work for a week it is going to be done on Thursday week to plus one. So for example, week one notebook is going to be reviewed this fairly homework one is going to be done next Thursday arch third and then reviewed in the week to to session. the notebook two will be reviewed in the week to that session and so on. All of this is in the syllabus in terms of the days, so you now something is confusing. We'll cut you guy some slack in the first few weeks of course, but I think it'll be self explanatory as we move forward. That's the first week we have five minutes left for any questions. | 1 |
|---|---|
| medical_ data_ science Ezurich Lecture Machine Learning for Health Care" (261-5120-00L) Basics of Natural Language Processing Gunnar Ratsch, Rita Kuznetsova Biomedical Informatics group, Institute for Machine Learning; Department of Computer Science DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Outline for today Introductionlmotivation Basic preprocessing steps Basic text features LDA algorithm Embeddings: from BoW to word embedding models POS tagging Language Modelling DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Towards Comprehensive Patient Models for Precision Medicine Distill Data Embed Data Connect Data Clinical Trial Design Precision Medicine Health Records Pathology Images Genomic Data X=w( p(phenotype x, t) Drugs Predict Treatment Mobile Health = argmax p(survival X, DINFK Gunnar Ratsch 15. 3. 2022 3 medical_ data_ science Ezurich Usefulness of the clinical texts Classification of the clinical notes Binary: mortality Multiclass: phenotyping Sentiment analysis of the clinical reports is diagnosis confirmed, suspected or excluded Topic modelling diseases, treatment etc Medical Named Entity Recognition diseases, drugs etc Text generation medical report generation (for example based on images) Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 Latent Representation DINFK medical_ data_ science Ezurich Problems with clinical texts Ungrammatical, has misspellings and concatenations. Contains short telegraphic phrases, acronyms, abbreviations, which are often overloaded It can contain many things that can be typed or pasted, such as long sets of lab values or vital signs Institution-specific template-use is common Pervasive fear; misunderstanding, and confusion around security, privacy, de-identification, and anonymization => significant barriers to get access Some sections might be long and detailed, other sections might be empty or only contain some captions DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 5 medical_ data_ science Ezurich Basic Text Processing Most NLP tasks needs to do text preprocessing and normalization: a Segmenting tokenizing words b_ Normalization C. Stop-words removal d. Punctuation removal e Lemmatization Istemming Disclaimer: Natural Language Processing is a huge topic on its own and we can only cover the absolute basics. Check out lectures on Natural Language processing and understanding offered by colleagues in the department: DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Basic preprocessing steps Tokenization "This is a cat:' 99 "This" "is" "a" "cat" 66 9) Issues: "whatre, Im, isn't" "What are, |am; is not" Ways to do: spacy, nltk, custom tokenizer Normalization The indexed text and the query terms must have the same form: e. g., the US, USA & U. S. A. are the same; could be done with RegExp lowercasing Stop-words removal from nltk. corpus import stopwords ['out' on 'off' 'over' under' 'again' 1 further' then' 'once here there'when'where '.. ] Punctuation removal from string import punctuation DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Stemming vs Lemmatization Stemming reduce terms to their stems in information retrieval: Is crude chopping of affixes automate(s), automatic, automation automat Consult; consultant; consulting, consultative, consultants consult Lemmatization have to find the correct dictionary headword form; use of a vocabulary and morphological analysis of words: Lemma: same stem, part of speech a lemma is the dictionary form of a set of words (headword): cat and cats = same lemma (cat) run; runs, ran, running = same lemma (run) Reduce inflections or variant forms to base form Car; cars, car's, cars car Am, are, is 3 be DINFK NLP Stanford Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Bag of Words (BoW) How to map the term representation into vector space? One way to do this is to represent documents as the bag (multiset) of its words, disregarding grammar and even word order; but keeping multiplicity "No, my brain did not hurt 9} "Perhaps it was more exasperating this way than if it had."66 would have preferred it to hurt me. We assign every word w from vocabulary Wa one-hot vector: Vw [0, 0, 1, 0,., 0] e RII W 0 The document is represented as d = {w_1, w_2, "J w_n}, then we could assign the vector for the document d: Vd = Vw wed vocab size W O1oo, 000) 0 1 "dog 0 DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Bag of Words (BoW) & Term Frequency How to map the term representation into vector space? One way to do this is to represent documents as the bag (multiset) of its words, disregarding grammar and even word order; but keeping multiplicity "No, my brain did not hurt 9} "Perhaps it was more exasperating this way than if it had."would have preferred it to hurt me. Term Frequency (TF): raw count of term in document: number of times that term t occurs in document d. {"no" 1, "my": 1, "brain": 1, "did": 1, "not": 1, "hurt": 1} {"Perhaps": 1, "it": 2, 66 was": 1, "more": 1, "exasperating": 1, "this": 1, "way": 1, "if: 1, 66 "had': 1} {": 1, 66 would": 1, "have' 1, "preferred": 1, 66 "it": 1, "to": 1, "hurt": 1, "me": 1} W) Easy to use bag of word representation for vanilla ML models (like SVMs) as inputs (example) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 10 medical_ data_ science Ezurich TF-IDF (term frequency-inverse document frequency) TF suffers from a critical problem: all terms are considered equally important: In fact, certain terms have little or no discriminating power in determining relevance. How to scale down the term weights? (Simple idea: scale down the term weights of terms with high frequency, defined to be the total Inumber of occurrences of a term in the document collection: Document Frequency (IDF): the number of documents in the collection that contain a term t TF-IDF: assigns to term t a weight in document d that is highest when t occurs many times within a small number of documents (thus lending high discriminating power to those documents) lower when the term occurs fewer times in a document; or occurs in many documents (thus offering a less pronounced relevance signal); lowest when the term occurs in virtually all documents DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 11 medical_ data_ science Ezurich TF-IDF (term frequency-inverse document frequency) nt TF (t, d) k nk nt the number of times that term t occurs in document d. Denominator the total number of terms in document d. IDI IDF(t, D) = log Kdi EUR D lt e di } Usage: similarity computation feature vector for the classifier (as a baseline) IDl total number of documents in the collection Kd; e D l t e d } number of documents from the collection, where the term t appears, "t = 0. TF-IDF = TF(t, d)* IDF (t, D) TF-IDF values could be obtained with sklearn feature extraction. text. TfidfVectorizer DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 12 medical_ data_ science Ezurich Latent Dirichlet Allocation (LDA) Topics What is the Topic Modelling? geae 8. 82 genetic 0. 01 Word: element in vocabulary set; W Document: collection of words, d (BoW assumption) life 0. 02 evolve 0. 01 Corpus: collection of documents, D organism 0. 01 Documents Topic proportions and assignments Seeking Life's Bare (Genetic) Necessities COl SPix HARBOR; Vnl FORK nalI "Fectak Hou {mcnWthe 72 Lnme"the dunl Saks t lun "Wuak Urt #itctnt ArFr 'ch'Fnentol Cl Wi U cuulelre: Wetd. Ju" IItIun ( #urh tT= t cnr canluku Wa mtd 4 ~llte Ieuh ~ua I Thetulie he c#l Uk nce MTT Ute T" Arcu: Muhun: nen htum7e et > Icuu Fuky' JeN R Mult RuuueeI k a IntruE tkut Iueuntit R theMnln/ Cmfirin: Go J Fkns m Juthe wat Alhw_h ohe nmker: dent 1 J UEuch Mtl. I~ Frehcts Topic: collection of words (elements from vocabulary set) Document is represented by latent mixture of topics: p(wld) p(wlz)p(zld) where z topic. Rearon 8. 82 nerve 0. 01 Gonomo appinp and SoqvancCcld Sprng Hor Ncx Yok Mey 8 lo 12 Stripping down @omeuter ana y vieis @nes = Iatr ofina Minimum Nocain and &n0 EURnt Denonio; data 0. 02 nunber 0. 02 computer 0. 01 C'IENLE Ml : Mi Each topic is a distribution over words Each document is a mixture of corpus-wide topics Each word is drawn from one of those topics Material from Tutorial by David Blei, 2012 For given corpus, we learn: 1_ Topic: from full vocabulary important subsets 2_ Topic proportion: for each document what is about? Vocabulary ~ basis vector extraction (topics) represent d in topic space (topic proportion) (topic proportion could be used as a feature vector for downstream tasks) Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 13 DINFK medical_ data_ science Ezurich Latent Dirichlet Allocation LDA is a generative probabilistic model for collections of discrete data such as text corpora. Proportions Per-word n EUR RV, a EUR RK are the model parameters parameter topic assignment Per-document Observed Assume there are K topics across all the documents: topic proportions word Material from Tutorial by David Blei, 2012 Topic parameter Topics For k in (1, K): choose per-corpus topic distribution Bk e RV ~ Dir(n) For d in (1, D): choose per-document topic proportion 0d EUR RK ~ Dir(a) 04 Bk K n Za;n Wa, n N D For each word w_n: choose topic Zd, n EUR Lx Multinomial(0a) choose word Wd, n eZv Multinomial(Wd, nlzd, m Bk) p(B, 0, 2, wla, n) = K N P(B;In) [ [r(oala) Mp(zd, n| Oa)p(Wd, nl B1. K, _ Zd, n) d=1 a is the parameter of the Dirichlet prior on the per-document topic distributions B is the parameter of the Dirichlet prior on the per-topic word distribution is the topic distribution for document m mn is the topic for the n-th word in document m W is the specific word mn DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 14 medical_ data_ science _ Ezurich From basic text features to word embeddings Goal: to map each word to the vector: BoW and One-hot: Easy to build; Big dimensionality; Do not reflect the relationship of words in the text: Distributional hypothesis: words that are occur in the same contexts have similar meanings: DINFK Harris (1954). Distributional structure Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 15 medical_ data_ science _ Ezurich NeurIPS 2013 Distributed Representations of Words and Phrases and their Compositionality word2vec Tomas Mikolov Google Inc Mountain View mikolov@google com Ilya Sutskever Google Inc. Mountain View ilyasu@google com Kai Chen Google Inc _ Mountain View kai@google com arXiv 2013 Greg Corrado Jeffrey Dean Google Inc. Mountain View jeff@google com Efficient Estimation of Word Representations in Vector Space Tomas Mikolov Google Inc.. Mountain View, CA tmikolov@google com Kai Chen Google Inc. Mountain View, CA kaichen@google com Greg Corrado Jeffrey Dean Google Inc, Mountain View, CA gcorrado@google com Google Inc, Mountain View, CA jeff@google com DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 16 https: / /code. google. com/p/ordZvecl medical_ data_ science _ 1 Ezurich word2vec The basic idea is to predict a missing word from a context: What is context? Example: "the quick red fox jumps over the lazy brown dog' 1_ Continuous bag of words (CBOW) the 'context' is the sum (or mean) of vectors appearing in the sentence. Based on this context we predict the central' word. dog fox brown quick lazy the jumps red the over 2 Skip-gram the 'context' is each word from the surrounding central word. Based on the central word we predict each word from this surrounding: jumps the jumps quick jumps DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 17 dog medical_ data_ science _ Ezurich CBOW VS Skip-gram Continuous bag of words (CBOW) Skip-gram INPUT PROJECTION OUTPUT INPUT PROJECTION OUTPUT w(t-2) w(t-2) w(t-1) w(t-1) SUM w(t) w(t) w(t+1) w(t+1) N log P(w;b i-1 wi-k' wi+k) i=1 N k log P(Wi+ilw;) i=1 j=-k j=O w(t+2) w(t+2) Objective: maximise log-probability DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 18 medical_ data_ science _ Elzurich Skip-gram model Back to the optimisationl Maximise this: N k log P(Wi+jlwi_ i=1 j=-k j=O This prime is importantl T W0 UwI exp p(wolwi) W T w=1 exp vW Uw[ For each word, J learn two representations: 1. as the context jumps {the quick red fox over the lazy brown dog} Distributed Representations of Words and Phrases and their Compositionality Mikolov; Sutskever; Chen, Corrado, Dean, NIPS 2013 Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 19 2. as the word itself DINFK medical_ data_ science _ Ezurich Paradigmatic relations This distinction is importantl "the quick red fox jumps over the lazy brown dog" "the quick red fox leaps over the lazy brown dog 'leaps' and 'jumps' are similar because they fit the context: We don't expect to see them occurring together in a sentence, because they have a paradigmatic relationship. (This means they're somehow interchangeable:) The Distributional Hypothesis Magnus Sahlgren, PhD dissertation (2006) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 20 medical_ data_ science _ Ezurich Country and Capital Vectors Projected by PCA China Beijing Russias Japans Moscow Ankara 3 Tokyo Turkey 2 1. 5 0. 5 Polandk Germany France Warsaw Berlin Paris0. 5 Italy Greece Athens Rome Spains Madrid Lisbon1. 5 Portugal 22 2150. 5 0. 5 1. 5 2 figure 2 from Mikolov et al. II Distributed Representations of Words and Phrases and their Compositionality NIPS 2013 Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 21 DINFK medical_ data_ science _ Ezurich So does it work? Tested vectors using an analogical reasoning task A : B as X : Y'? (e. g: 'puppy is to dog as kitten is to cat) This means asking: vec(A) 5 vec(B) 2 vec(X) vec(Y) Or 2 vec(A) 5 vec(B) + vec(Y) vec(X) Where, as mentioned before, the 'similarity' here is cosine similarity. DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 22 medical_ data_ science _ Ezurich So does it work? Created a test set with ~9k semantic questions and ~11k syntactic Examples: calm calmly safe safely Athens Greece Japan big bigger small smaller old oldest best Poland zloty Hungary forint move moving fly flying Austin Texas Honolulu Hawaii Ireland Irish Egypt Egyptian dancing danced saying said girl uncle aunt man men cat cats (these are actually most of the categories) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 23 Tokyo good boy medical_ data_ science _ Ezurich So does it work? Table 4: Comparison of publicly available word vectors on the Semantic-Syntactic Word Relationship test set; and word vectors from our models: Full vocabularies are used Model Vector Training Accuracy [%] Dimensionality words Semantic Syntactic Total Collobert-Weston NNLM 50 660M 9. 3 12. 3 11. 0 Turian NNLM 50 37M 1. 4 2. 6 2. 1 Turian NNLM 200 37M 1. 4 2. 2 1. 8 Mnih NNLM 50 37M 1. 8 9. 1 5. 8 Mnih NNLM 100 37M 3. 3 13. 2 8. 8 Mikolov RNNLM 80 320M 4. 9 18. 4 12. 7 Mikolov RNNLM 640 320M 8. 6 36. 5 24. 6 Huang NNLM 50 990M 13. 3 11. 6 12. 3 neural net Our NNLM 20 6B 12. 9 26. 4 20. 3 Our NNLM 50 6B 27. 9 55. 8 43. 2 language model Our NNLM 100 6B 34. 2 64. 5 50. 8 CBOW 300 783M 15. 5 53. 1 36. 1 Skip-gram 300 783M 50. 0 55. 9 533 yes! this takes 2 weeks DINFK on 180 cores! 2. 5 days on 125 cores Efficient Estimation of Word Representations in Vector Space' Mikolov, Chen; Corrado, Dean, arXiv 2013 medical_ data_ science _ Ezurich Analogical reasoning The oft-cited example of successful analogical reasoning is: vec(king) vec(man) vec(woman) = vec(queen) Intuitively; this means vec(king) vec(man) vec("sense of royalty") Which is a pretty fascinating idea: What if. _ vec(gleevec) vec(leukemia)~ vec("treatment")? Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 25 DINFK medical_ data_ science _ Ezurich Analogical reasoning Then we could do.. vec(melanoma) vec("treatment") = vec("?? 2") This would certainly be useful for a medical Jeopardy. _ It's also the kind of information we want our embeddings to encode 9 for enabling medical language processing: So we ran wordZvec on some MSKCC text. DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 26 medical_ data_ science _ Elzurich An example vec(melanoma) [vec(gleevec) vec (leukemia)] =? 22 Top 10 closest results to melanoma: Top 10 closest results to gleevec: Top 10 closest results to Zeukemia: 3 me Lanoma 340 ~2. 22044604925e-16 gLeevec crel 2148 1. 11022302463e-16 Leukemia 1129 ~2. 22044604925e-16 dfsp 12007 0. 156969851964 dasatinib 4299 0. 0436408836789 itp A3744 0. 216802690111 neurotropic 17723 0. 18892856986 imatinib 2596 0. 0444753136031 myelodysplasia 8348 0. 220480414542 neurotrophic 22261 0. 193823458626 nilotinib 5211 0. 0565038545667 cll 1270 0. 229098946201 SCC 4457 0. 199310695303 anagrelide 6961 0. 0806562061213 aml 2236 0. 232815740704 amelanotic 10967 0. 205920966083 hydroxyurea 3921 0. 0824481117079 cmmol 8659 0. 236774032333 choroidal 9357 0. 208689368781 gleevac 16087 0. 0843472720016 mds 2858 0. 23788044386 fibrosarcoma 8679 0. 223705333644 ruxolitinib 11279 0. 0845686271568 coexisting 16242 0. 241202871701 eccrine Jl13344 0. 22791107347 "ieeeie nexavar 7350 0. 0862700069714 Leukemia/sll 35544 0. 245014293364 fibrohistiocytoma 11045 0. 239171736259| hydrea 6445 0. 100871239337 Igl Bxa10616 0. 246571161984 cancer/ 27891 0. 243011714361 afinitor 10465 0. 10846339025 hypogammaglobulinemia 6544 0. 249632640917 Top 10 closest results to UNM+ G-L and we get:.. CMLIALL ponatinib 14236 0. 42982079986 diascopy 23369 0. 435802215172 #I#th 20379 0. 44739503288 eruption 3763 0. 447999880188 gleevac 16087 0. 448643975007 nexavar 7350 0. 452329959825 hive 18576 0. 455971559975 pustule 11842 0. 455989743044 gleevec Ae2148'0. 458117608185 dabrafenib 10448 0. 459866794241 desatinib 32409 0. 46007721604 typo :) sorafenib (kidneylliver cancer drug) (BRAF-mutated, metastatic) MELANOMAI DINFK CMLIALL medical_ data_ science _ Ezurich Other embedding techniques 1. Glove [1] aggregated global word-word co-occurrence statistics from a corpus. 2. FastText [2] ~ training process with character n-grams, helps with OOV (Out of vocabulary) problem: 1 GloVe: Global Vectors for Word Representation; Pennington, J. Socher; R,, & Manning; C. D."EMNLP 2014. 2. Enriching_Word_Vectors_with_Subword Information; Bojanowski, P, Grave, E., Joulin, A"& Mikolov, T. TACL, 2017. DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 28 medical_ data_ science Ezurich Part of Speech (PoS) Tagging Given a sentence W. Wn and a tagset of lexical categories, find the most likely tag t-tn for each word in the sentence Penn_Treebank PQSTags] SecretariatINNP isIVBZ expected/VBN toITO race/VB tomorrow/NN Problem: Many of the words may have unambiguous tags But enough words are either ambiguous or unknown non-trivial task Brown corpus is a general corpus in corpus linguistics (500 samples of English texts, ~IM words). Most words in English have only one Brown Corpus tag: unambiguous (1 tag) 35, 340 words Many of the most common words are ambiguous (over 40% tokens are ambiguous) Obvious strategies may be suggested based on intuition: to/TO race/VB thelDT racelNN Methods: simple baseline with unigrams hardcoded rules vs supervised unsupervised ML approaches. NNP Proper noun, singular; VBZ Verb, 3rd person singular present; VBN Verb, past article; VB verb, base form, NN Noun; singular or mass: DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 29 medical_ data_ science Ezurich Simplest strategy Choose the most likely tag for each ambiguous word, independent of previous words assign each token the POS category it occurred as most often in the training set This strategy gives 90% accuracy in controlled tests Which POS is more likely in a corpus (1, 273, 000 tokens)? race: NN: 400 VB: 600 P(NNlrace) = P(race, NN) / P(race) by the definition of conditional probability P(race) = 1000/1, 273, 000 =. 0008 P(race, NN) = 400/1, 273, 000 =. 0003 P(race, VB) = 600/1, 273, 000 =. 0004 SO we obtain: P(NNIrace) = P(race, NN)P(race) =. 0003/. 0008 =. 375 P(VBlrace) = P(race, VBYP(race) =. 0004/. 0008 =. 5 DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 30 medical_ data_ science Ezurich HMMs: A Probabilistic Approach We want to find the "best sequence' 9) of PoS tags T=T_ T for & sentence 1' W=W_ 1"W n' where T; is the PoS tag for word W In other words we want to find a PoS tags T that maximizes P(TIW) Using Bayes' Rule, we can say P(WIT) * P(T) P(TIW) P(W) We want to find the value of T which maximizes the right hand side. note: denominator can be discarded (same for every T) Find T which maximizes P(WIT) * P(T) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 31 medical_ data_ science Elzurich Independence Assumptions Assume that current event is based only on previous n-1 events n 1 P(T1, _, Tn) IIP(TIT;_1) i=1 assumes that the event of a PoS tag occurring is independent of the event of any other PoS tag occurring, except for the immediately previous PoS tag from a linguistic standpoint; this seems an unreasonable assumption, due to long-distance dependencies 2_ P(W1.. WnITi. Tn) II P(WAT;) i1 assumes that the event of a word appearing in a category is independent of the event of any surrounding word or tag, except for the tag at this position: Linguists know both these assumptions are incorrect nevertheless, statistical approaches based on these assumptions work pretty well for part-of-speech tagging: Hidden Markov Models (HMMs) is widely used in both PoS-tagging and speech recognition, among other problems DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 32 medical_ data_ science Ezurich PoS Tagging Based on HMM Problem: Find T which maximizes P(W | T) * P(T) Here W=W, 1W n and T=T Tn Using the HMM; we get: Transition probabilities (prob. of transitioning from one stateltag to another): n P(T1, Tn) IIP(T;IT;_1) i=1 Emission probabilities (prob. of emitting a word at a given state): P(W1. WnITi.. Tn) II P(wt;) i] We want to find the value of T_ 1"~T which maximizes: n n II P(T;) * P(TilTi-1) {=1 DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 33 medical_ data_ science Ezurich POS probabilities: transitions P(T1, _ '' 1 Tn) ~ IIP(TIT;_1) 0. 8 i=1 1. 0 PRP MD 04 NN 0. 6 0. 3 "He will race J} Possible tag series for T = T1'T2'T3 T =PRP MD NN T = PRP NN NN T = PRP MD VB T =PRP NN VB POS bigram probabilities from training corpus can be used for P(T) 0. 2 NN 0_ B PRP: personal pronoun MD: modal (auxiliary verb) NN: noun VB Verb, base form Which is the most likely path? DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 34 medical_ data_ science Ezurich Lexical generation probabilities n From the training corpus, we need to find the T; which maximizes: [[ P(WAT;) * P(TilT;_1) C E i] 0. 4 0. 8 MD NN A B 1Q PRP 0. 6 0. 3 0. 2 NN VB 0_ C Emission Probabilities 0. 4 willMD 0. 8 racelNN 0. 4 MD NN VB PRP 0. 8 A B 0. 6 he S | $ 10 helPRP 1. 0 will 0. 8 0. 2 0. 3 0. 22 0. 7 race 0 0. 4 0. 6 willINN 0. 2 racelVB 0. 6 Note 1: sum over column should sum up to 1, the MLE of the emission probability is how many times this word W appears as this tag t, divided by how many times we saw the tag t in training data: DINFK Note 2: the whole table extends further down than is depicted, as there are other words in the dictionary whose emission probability we're not interested in for this sentence. Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 35 medical_ data_ science Ezurich Using Dynamic Programming To find the most likely sequence of categories for a sequence of words, we don't need to enumerate all possible sequences of categories. Due to the Markov assumption, if you keep track of the most likely sequence found so far for each possible ending category; you can ignore all the other less likely sequences: multiple edges coming into a state, but only keep the value of the most likely path, i. e. use of DP The algorithm to do this is called the Viterbi algorithm: DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 36 medical_ data_ science Ezurich Viterbi Algorithm recap (informal) 1_ Assume we are at state h in the HMM: a States H, _H m all come into h 2. Obtain a the best probability of each previous state H_Hm b the transition probabilities: P(hIH;); P(hHm) C the emission probability for word w at h: P(wlh) 3_ Multiply the probabilities for each new path: Best(l) = Maxh <h [Best(H)* P(hlH)]* P(wlh) 4_ One of these states (H_ 1"H m) will give the highest probability Only keep the highest probability when using h for the next state E 0 willlM D racelN N 0. 4 0_ 8 A B 1Q helPR P S | 0 03 0 0_ 2 Find the most likely sequencel DINFK willIN N 82 0. 7 racelV B 0. 6 F medical_ data_ science Ezurich Introduction to the Transformer model Transformer is the encoder-decoder architecture with self-attention: The key component of the Encoder is Self-Attention mechanism. Layer normalization Input data: X e Rnxd Encoder overview: Q-queries, K keys, V values: linear projections of the input data X, d dimension: # QKT Latent Representation softmax )V Residual 1 Va connections Add & Normalize Feed Forward Feed Forward Projection layer / Add & Normalize Self-Attention Attention weights Positional embeddings POscoDMG Self attention Positional Embeddings is used to learn the order in the sequence. Residual connections mitigates vanishing gradients. Layer normalization helps with numerical stability: POSITIONAL ENCODING EMBEDDINGS X1 Pictures from The Ilustrated transformer_Jay Alammar DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 38 Vaswani, Ashish, et al. "Attention is allyou need"Advances in neural information processing systems 30 (2017). medical_ data_ science Ezurich Language Modeling Deep Learning Language Model (LM) computes the probability of p(w_1, w_n) for any W_1, _ w_n = V (vocabulary) How we do LM in modern world? state-of-the art Transformer-based models BERT Input [CLS] my is cute [SEP] he likes play ##ing [SEP] Pretrained WordPiece embeddings Token Embeddings E [CLS] E my E "dog E cute E 'play [SEP] ~likes [SEP] A marker indicating Sentence A or Sentence B Segment Embeddings EA EB To learn about the position in sentence Position Embeddings Eo E Ez E4 E5 E6 E7 Eg E1o 1 [CLS] token is added in the beginning a special classification token, the final hidden state corresponding to this token used as the aggregate sequence representation for classification tasks. 2_ [SEP] token is inserted at the end of each sentence. One sentence also can feed as input. 3_ Both sentences A and B are encoded with the help of Byte-pair encoding (BPE): BPE is a simple form of data compression in which the most common pair of consecutive bytes of data is replaced with a byte that does not occur within that data. ZabdZabac DINFK Devlin et al. 'Bert: Pre-training of_deep bidirectional transformers for language understanding arXiv preprint arXiv:1810. 04805 (2018). Z-aa dog 8 &ne E#ring 4 4 4 4 4 8 8 8 8 S 8 medical_ data_ science Elzurich Training details 1. Masked-LM: replace n% words in the input by special [MASK] token predict these tokens variants altering Structural DNA [MASK] can modify gene function by [MASK] transcript sequences 2. Next sentence prediction: binarized task; from the [CLS] token need to predict whether one sentence follows another in the text. Class 1: Cancer is a group of diseases involving abnormal cell growth It has the potential to invade or spread to other parts of the body: Class 0: Cancer is a group of diseases involving abnormal cell growth A bear is sunbathing: The basis of the BERT architecture = encoder of the Transformer model Way of using BERT: have pretrained model finetune for the required task on the specific corpora DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 40 medical_ data_ science Ezurich BERT in biomedical domain Data: MMC I clinical notes, PubMed abstracts, papers from Semantic Scholar etc Tasks: Named Entity Recognition, Q&A, Sentence Similarity etc Hugging Face Search models, datasets, users__ Models Datasets Spaces Docs Solutions Pricing Tasks Models 296 bio Fill-Mask Question Answering dmis-lab/biobert-base-cased-V1. 1 Updated Oct 14, 2020 1. 01M Summarization Table Question Answering Text Classification Text Generation Text2Text Generation 88 Token Classification emilyalsentzer/Bio_ClinicalBERT Fill-Mask Updated 16 days ag0 533k 14 Translation Zero-Shot Classification 218 Sentence Similarity +14 microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext Fill-Mask Updated Sep 22, 2021 121k 19 Libraries PyTorch TensorFlow JAX 4 24 microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract Fill-Mask Updated Sep" 22, 2021 85. 2k Datasets wikipedia common_voice squad dmis-lab/biobert-V1. 1 Feature Extraction Updated May19, 2021 bookcorpus c4 glue conll2003 73k 0 4 dcep europarl jrc-acquis 828 bionlp/bluebert_pubmed_mimic_uncased_L-12_H-768_A-12 Updated Sep 24, 2021 62,. 6k Gunnar Ratsch & Rita Kuznetsova Languages DINFK 22. 3. 2022 41 medical_ data_ science _ Ezurich Summary & Take Home messages Clinical reports contain relevant information about patient's health state, but are written in challenging, specialized language Standard basic text processing is often applied: tokenization, stemming, lemmatization, Latent Dirichlet Allocation (LDA) is a probabilistic model for text content using topics Word embedding are a recently developed powerful tool in NLP Can be learned unsupervisedly and represent semantic in a vector space Part of Speech Tagging is an import task in NLP that is often solved with HMMs Recent NLP techniques use deep learning and have shown great promise DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 42 | 1 |
| <b><!-- </b>if (window!= top) top.location.href=location.href <b>// --> </b> Sunset Boulevard SUNSET BOULEVARD Charles Brackett Billy Wilder D.M. Marshman, Jr. March 21,1949 SEQUENCE "A" A-l-4 START the picture with the actual street sign: SUNSET BOULEVARD, stencilled on a curbstope. In the gutter lie dead leaves, scraps of paper, burnt matches and cigarette butts. It is early morning. Now the CAMERA leaves the sign and MOVES EAST, the grey asphalt of the street filling the screen. As speed accelerates to around 40 m.p.h., traffic de- marcations, white arrows, speed-limit warnings, man- hole covers, etc., flash by. SUPERIMPOSED on all this are the CREDIT TITLES, in the stencilled style of the street sign. Over the scene we now hear MAN'S VOICE sirens. Police squad cars Yes, this is Sunset hurtle toward the camera, Boulevard, Los Angeles, turn off the road into a California. It's about driveway with squealing five o'clock in the brakes. Dismounted motor- morning. That's the cycle cops stand directing Homicide Squad, com- the cars in. plete with detectives and newspaper men. A-5 PATIO AND POOL OF A murder has been re- MANSION ported from one of those great big houses in the The policemen and news- ten thousand block. paper reporters and You'll read all about photographers have it in the late editions, jumped out of the cars I'm sure. You'll get and are running up to it over your radio, the pool, in which a and see it on tele- body is seen floating. vision -- because an Photographers' bulbs old-time star is in- flash in rapid suc- volved. one of the big- cession. gest. But before you hear it all distorted and blown out of proportion, before those Hollywood columnists get their hands on it, maybe you'd like to hear the facts, the whole truth... A-6 FLASH OF THE BODY MAN'S VOICE Angle up through the If so, you've come to the water from the bottom right party... You see, of the pool, as the the body of a young man body floats face down- was found floating in the ward. It is a well- pool of her mansion, with dressed young man. two shots in his back and one in his stomach. No- body important, really. Just a movie writer with a couple of "B" pictures to his credit. The poor dope. He always wanted a pool Well, in the end he got himself a pool -- SLOW DISSOLVE TO: only the price turned out to be a little high... Let's go back about six A-7 HOLLYWOOD, SEEN FROM months and find the day THE HILLTOP AT IVAR when it all started. & FRANKLIN STREETS It is a crisp sunny I was living in an day. The voice con- apartment house above tinues speaking as Franklin and Ivar. CAMERA PANS toward Things were tough the ALTO NIDO APART- at the moment. I hadn't MENT HOUSE, an ugly worked in a studio for Moorish structure ofsat a long time. So I stucco, about four there grinding stories high. CAMERA out original stories, MOVES TOWARD AN OPEN two a week. Only I WINDOW on the third seemed to have lost floor, where we look my touch. Maybe they in on JOE GILLIS' APART- weren't original MENT. Joe Gillis, bare- enough. Maybe they footed and wearing no- were too original. thing but an old bath- All I know is they robe. is sitting on didn't sell. the bed. In front of him. on a straight chair, is a portable typewriter. Beside him, on the bed, is a dirty ashtray and a scattering of type written and pencil- marked pages. Gillis is typing. with a pencil clenched bet- ween his teeth. A-8 JOE GILLIS' APARTMENT It is a one-room affair with an unmade Murphy bed pulled out of the wall at which Gillis sits typing. There are a couple of worn-out plush chairs and a Spanish-style, wrought-iron standing lamp. Also a small desk littered with books and letters, and a chest of drawers with a portable phonograph and some records on top. On the walls are a couple of repro- ductions of characterless paintings, with laundry bills and snapshots stuck in the frames. Through an archway can he seen a tiny kitchenette, complete with unwashed coffee pot and cup, empty tin cans, orange peels, etc. The effect is dingy and cheerless -- just another furnished apartment. The buzzer SOUNDS. GILLIS Yeah. The buzzer SOUNDS again. Gillis gets up and opens the door. Two men wearing hats stand outside one of them carrying a briefcase. NO. 1 Joseph C. Gillis? GILLIS That's right. The men ease into the room. No. 1 hands Gillis a business card. NO. 1 We've come for the car. GILLIS What car? NO. 2 (Consulting a paper) 1946 Plymouth convertible. Calif- ornia license 97 N 567. NO. 1 Where are the keys? GILLIS Why should I give you the keys? NO. 1 Because the company's played ball with you long enough. Because you're three payments behind. And because we've got a Court order. Come on -- the keys. NO. 2 Or do you want us to jack it up and haul it away? GILLIS Relax, fans. The car isn't here. NO. 1 Is that So? GILLIS I lent it to a friend of mine. He took it up to Palm Springs. NO. 1 Had to get away for his health, I suppose. GILLIS You don't believe me? Look in the garage. NO. 1 Sure we believe you, only now we want you to believe us. That car better be back here by noon tomorrow, or there's going to be fireworks. GILLIS You say the cutest things. The men leave. Gillis GILLIS' VOICE stands pondering beside Well, I needed about two the door for a moment. hundred and ninety dollars Then he walks to the and I needed it real center of the room and, quick, or I'd lose my car. with his back to the It wasn't in Palm Springs CAMERA, slips into a and it wasn't in the pair of gray slacks. garage. I was way ahead There is a metallic of the finance company. noise as some loose change and keys drop from the trouser pockets. As Gillis bends over to pick them up, we see that he has dropped the car keys, identifiable be- cause of a rabbit's foot and a miniature license plate attached to the key-ring. Gillis pockets the keys and as he starts to put on a shirt DISSOLVE TO: A-9 EXTERIOR OF RUDY'S GILLIS' VOICE SHOESHINE PARLOR (DAY) I knew they'd be coming A small shack-like build- around and I wasn't tak- ing, it stands in the ing any chances, so I corner of a public park- kept it a couple of ing lot. Rudy, a blocks away in a parking colored boy, is giving lot behind Rudy's Shoe- a customer a shine. shine Parlor. Rudy never asked any quest- ions. He'd just look at your heels and know the score. PAN BEHIND the shack to GILLIS' CAR, a yellow 1946 Plymouth convertible with the top down. Gillis enters the SHOT. He is wearing a tweed sport jacket, a tan polo shirt, and moooasins. He steps into the car and drives it off. Rudy winks after him. A-10 THE ALLEY NEXT TO SIDNEY'S MEN'S SHOP ON BRONSON AVE. GILLIS' VOICE I had an original story Gillis drives into the kicking around Paranount. alley and parks his car My agent told me it was right behind a delivery dead as a doornail. but truck. PAN AND FOLLOW I knew a big shot over HIM as he gets out, walks there who'd always liked around the corner into me, and the time had Bronson and then toward come to take a little the towering main gate of advantage of it. His Paramount. A few loafers, name was Sheldrake. He studio cops and extras are was a smart producer, lounging there. with a set of ulcers to prove it. DISSOLVE TO: A-11 SHELDRAKE'S OFFICE It is in the style of a Paramount executive's office -- mahogany, leather, and a little chintz. On the walls are some large framed photographs of Paramount stars, with dedications to Mr. Sheldrake. Also a couple of framed critics' awards certificates, and an Oscar on a bookshelf. A shooting schedule chart is thumb-tacked into a large bulletin board. There are piles or scripts, a few pipes and, somewhere in the background, some set models. Start on Sheldrake. He is about 45. Behind his wor- ried face there hides a coated tongue. He is en- gaged in changing the stained rilter cigarette in his Zeus holder. SHELDRAKE All right, Gillis. You've got five minutes. What's your story about? GILLIS It's about a ball player, a rookie shortstop that's batting 347. The poor kid was once mixed up in a hold- up. But he's trying to go straight -- except there's a bunch of gamblers who won't let him. SHELDRAKE So they tell the kid to throw the World Series, or else, huh? GILLIS More or less. Only for the end I've got a gimmick that's real good. A secretary enters, carrying a glass or milk. She opens a drawer and takes out a bottle of pills for Sheldrake. SHELDRAKE Got a title? GILLIS Bases Loaded. There's a 4O-page outline. SHELDRAKE (To the secretary) Get the Readers' Department and see what they have on Bases Loaded. The secretary exits. Sheldrake takes a pill and washes it down with some milk. GILLIS They're pretty hot about it over at Twentieth, but I think Zanuck's all wet. Can you see Ty Power as a GILLIS (cont'd) shortstop? You've got the best man for it right here on this lot. Alan Ladd. Good change of pace for Alan Ladd. There's another thing: it's pretty simple to shoot. Lot of outdoor stuff. Bet you could make the whole thing for under a million. And there's a great little part for Bill Demarest. One of the trainers, an oldtime player who got beaned and goes out of his head sometimes. The door opens and Betty Schaefer enters -- a clean- cut, nice looking girl of 21, with a bright, alert manner. Dressed in tweed skirt, Brooks sweater and pearls, and carrying a folder of papers. She puts them on Sheldrake's desk, not noticing Gillis, who stands near the door. BETTY Hello, Mr. Sheldrake. On that Bases Loaded. I covered it with a 2-page synopsis. (She holds it out) But I wouldn't bother. SHELDRAKE What's wrong with it? BETTY It's from hunger. SHELDRAKE Nothing for Ladd? BETTY Just a rehash of something that wasn't very good to begin with. SHELDRAKE I'm sure you'll be glad to meet Mr. Gillis. He wrote it. Betty turns towards Gillis, embarrassed. SHELDRAKE This is Miss Kramer. BETTY Schaefer. Betty Schaefer. And right now I wish I could crawl into a hole and pull it in after me. GILLIS If I could be of any help... BETTY I'm sorry, Mr. Gillis, but I just don't think it's any good. I found it flat and banal. GILLIS Exactly what kind of material do you recommend? James Joyce? Dostoosvsky? SHELDRAKE Name dropper. BETTY I just think pictures should say a little something. GILLIS Oh, you're one of the message kids. Just a story won't do. You'd have turned down Gone With the Wind. SHELDRAKE No, that was me. I said, Who wants to see a Civil War picture? BETTY Perhaps the reason I hated Bases Loaded is that I knew your name. I'd always heard you had some talent. GILLIS That was last year. This year I'm trying to earn a living. BETTY So you take Plot 27-A, make it glossy, make it slick -- SHELDRAKE Carefull Those are dirty words! You sound like a bunch of New York critics. Thank you, Miss Schaefer. BETTY Goodbye, Mr. Gillis. GILLIS Goodbye. Next time I'll write The Naked and the Dead. Betty leaves. SHELDRAKE Well, seems like Zanuck's got himself a baseball picture. GILLIS Mr. Sheldrake, I don't want you to think I thought this was going to win any Academy Award. SHELDRAKE (His mind free-wheeling) Of course, we're always looking for a Betty Hutton. Do you see it as a Betty Hutton? GILLIS Frankly, no. SHELDRAKE (Amusing himself) Now wait a minute. If we made it a girls' softball team, put in a few numbers. Might make a cute musical: It Happened in the Bull Pen -- the story of a Woman. GILLIS You trying to be funny? -- because I'm all out of laughs. I'm over a barrel and I need a job. SHELDRAKE Sure, Gillis. If something should come along - GILLIS Along is no good. I need it now. SHELDRAKE Haven't got a thing. GILLIS Any kind of assignment. Additional Dialogue. SHELDRAKE There's nothing, Gillis. Not even if you were a relative. GILLIS (Hating it) Look, Mr. Sheldrake, could you let me have three hundred bucks yourself, as a personal loan? SHELDRAKE Could I? Gillis, last year some- body talked me into buying a ranch in the valley. So I borrowed money from the bank so I could pay for the ranch. This year I had to mortgage the ranch so I could keep up my life insurance so I could borrow on the insurance so I could pay my income tax. Now if Dewey had been elected - GILLIS Goodbye, Mr. Sheldrake. DISSOLVE TO: A-12 EXT. SCHWAB'S DRUG STORE (EARLY AFTERNOON ACTIVITY) GILLIS' VOICE After that I drove down MOVE IN toward drug store to headquarters. That's and the way a lot of us think about Schwab's Drug Store. DISSOLVE TO: Actors and stock girls and waiters. Kind of a combination office,Kaffee- A-13 INT. SCHWAB'S DRUG STORE Klatsch and waiting room. Waiting, waiting for the The usual Schwabadero gravy train. crowd sits at the fount- ain, gossips at the cigar-stand, loiters by the magazine display. MOVE IN towards the TWO TELEPHONE BOOTHS. In I got myself ten nickels one of them sits Gillis, and started sending out a stack of nickels in a general S.O.S. Couldn't front of him. He's get hold of my agent, doing a lot of talking naturally. So then I into the telephone, called a pal of mine,name hanging up, dropping of Artie Green -- an awful another nickel, dialing, nice guy, an assistant talking again. director. He cquld let me have twenty, but twenty wouldn't do. GILLIS' VOICE (Cont.) Then I talked to a couple of yes men at Twentieth. To me they said no. Finally I located that agent of mine, the big faker. Was he out digging up a job for poor Joe Gillis? Hmph! He was hard at work in Bel Air, making with the golf clubs. Gillis hangs up with a curse, opens the door of the booth, emerges, wiping the sweat from his forehead. He walks toward the exit. He is stopped by the voice of SKOLSKY Hello, Gillis. Gillis looks around. At the fountain sits Skolsky, drinking a cup of coffee. GILLIS Hello, Mr. Skolsky. SKOLSKY Got anything for the column? GILLIS Sure. Just sold an original for a hundred grand. The Life of the Warner Brothers. Starring the Ritz Brothers. Playing opposite the Andrew Sisters. SKOLSKY (With a sour smile) But don't get me wrong -- I love Hollywood. Gillis walks out. DISSOLVE TO: A-14 THE BEL AIR GOLF LINKS On a sun-dappled green edged with tall sycamores, stands Morino, the agent, a caddy and a nondescript opponent in the background. Gillis has evidently stated his problem already. MORINO So you need three hundred dollars? Of course, I could give you three hundred dollars. Only I'm not going to. GILLIS No? MORINO Gillis, get this through your head. I'm not just your agent. It's not the ten per cent. I'm your friend. He sinks his putt and walks toward the next tee, Gillis following him. GILLIS How's that about your being my friend? MORINO Don't you know the finest things in the world have been written on an empty stomach? Once a talent like yours gets into that Mocambo- Romanoff rut, you're through. GILLIS Forget Romanoff's. It's the car I'm talking about. If I lose my car it's like having my legs out off. MORINO Greatest thing that could happen to you. Now you'll have to sit behind that typewriter. Now you'll have to write. GILLIS What do you think I've been doing? I need three hundred dollars. MORINO (Icily) Maybe what you need is another agent. He bends down to tee up his ball. Gillis turns away. DISSOLVE TO: A-15 GILLIS IN HIS OPEN CAR GILLIS' VOICE driving down Sunset As I drove back towards town towards Hollywood. He I took inventory of my pros- drives slowly. His pects. They now added up to mind is working. exactly zero. Apparently I just didn't have what it takes, and the time had come to wrap up the whole Hollywood deal and go home. Maybe if I hocked all my junk there'd be enough for a bus ticket back to Ohio, back to that thirty-five- dollar-a-week job behind the copy desk of the Dayton Evening Post, if it was still open. Back to the smirking delight of the whole office. All Gillis stops his car at right you wise guys. why don't a red light by the main you go out and take a crack at entrance to Bel Air. Hollywood? Maybe you think Suddenly his eyes fall you could -- Oh-oh! on: A-16 ANOTHER CAR It is a dark-green Dodge business coupe, also waiting for the light to change. but headed in the opposite direction. In it are the two finance company men. They spot Gillis in his car and exchange looks. From across the intersection Gillis recognizes them and pulls down the leather sunshade to screen his face. As the light changes. Gillis gives his car the gun and shoots away. The men narrowly avoid hitting another car as they make a U-turn into oncoming traffic and start after him. A-17 THE CHASE to A-21 Very short, very sharp, told in FLASHES. (Use locations on Sunset between Bel Air and Holmby Hills). The men lose Gillis around a bend, catch sight of him and then -- while they are trapped behind a slow- moving truck. he disappears again. A-22 GILLIS He is driving as fast as he dares, keeping an eye out for pursuit in his rear-view mirror. Suddenly his right front tire blows out. Gillis clutches desperately at the steering wheel and manages to turn the careening car into A-23 A DRIVEWAY It is overgrown with weeds and screened from the street by bushes and trees. Gillis stops his car about thirty feet from the street and looks back. GILLIS' VOICE Was I far enough ahead? A-24 THE OTHER CAR shoots past the driveway, still looking for Gillis. A-25 GILLIS He watches his pursuers GILLIS' VOICE shoot past and out of Yeah... sight. He opens the door and looks down at I had landed myself in the the flat tire. Then he driveway of some big mansion looks around to see that looked run-down and where he is. deserted. At the end of the drive was a lovely sight A-26 DRIVEWAY WITH GARAGE indeed -- a great big empty garage, just standing there An enormous, five-car going to waste. If ever there affair. neglected and was a place to stash away a empty-looking. limping car with a hot license number... A-27 GILLIS He gets back into his There was another occupant in car and carefully pilots that garage: an enormous the limping vehicle into foreign-built automobile. It one of the stalls. In must have burned up ten gallons the adjoining one is a to a mile. It had a 1932 large, dust-covered license. I figured that's Isotta-Fraschini propped when the owners moved out... up on blocks. He closes I also figured I couldn't go the garage door and walks back to my apartment now that up the driveway. In idle those bloodhounds were on to curiosity he mounts a me. The idea was to get Artie stone staircase which Green's and stay there till I leads to the garden. could make that bus for Ohio. CAMERA IN BACK OF HIM. Once back in Dayton I'd drop At the top of the steps the credit boys a picturepost- he sees the somber pile card telling them where to of pick up the jallopy. NORMA DESMOND'S HOUSE GILLIS' VOICE It is a grandiose -- It was a great big white Italianate structure, elephant of a place. The kind mottled by the years, crazy movie people built in the gloomy, forsaken, crazy Twenties. A neglected little formal garden house gets an unhappy look. completely gone to This one had it in spades. It seed. was like that old woman in Great Expectations -- that Miss From somewhere above Haversham in her rotting wed- comes ding dress and her torn veil, taking it out on the world be- cause she'd been given the go- by. A WOMAN'S VOICE You there! Gillls turns and looks. A-28 UPSTAIRS LOGGIA Behind a bamboo blind there is a movement of a dark figure. WOMAN'S VOICE Wlly are you so late? Why have you kept me waitlng so long? A-29 GILLIS He stands flabbergasted. A new noise attracts his attention -- the creak of a heavy metal-and-glass door being opened. He turns and sees A-3O THE ENTRANCE DOOR OF THE HOUSE Max von Mayerling stands there. He is sixty, and all in black, except for immaculate white cotton gloves, shirt, high, stiff collar and a white bow tie. His coat is shiny black alpaca, his trousers ledger-atriped. He is semi-paralyzed. The left side of his mouth is pulled down, and he leans on a rubber-ferruled stick. MAX In here! Gillis enters the shot. GILLIS I just put my car in the garage. I had a blow-out. I thought -- MAX Go on in. There is authority in the gesture of his white- gloved hand as he motions Gillis inside. GILLIS Look, maybe I'd better take my car -- MAX Wipe your feet! Automatically, Gillis wipes his feet on an enormous shabby cocoanut mat. MAX You are not dressed properly. GILLIS Dressed for what? THE WOMAN'S VOICE Max! Have him come up, Max! MAX (Gesturing) Up the stairs! GILLIS Suppose you listen just for a minute - MAX Madame is waiting. GILLIS For me? Okay. Gillis enters. A-31 INT. NORMA DESMOND'S ENTRANCE HALL It is grandiose and grim. The whole place is one of those abortions of silent-picture days, with bowling alleys in the cellar and a built-in pipe organ, and beams imported from Italy, with California termites at work on them. Portieres are drawn before all the windows, and only thin slits or sunlight find their way in to fight the few electric bulbs which are always burning. Gillis starts up the curve of the black marble staircase. It has a wrought-iron rail and a worn velvet rope along the wall. MAX (From below) If you need help with the coffin call me. The oddity of the situation has caught Gillis' imagination. He climbs the stairs with a kind of morbid fascination. At the top he stops, undecided, then turns to the right and is stopped by WOMAN'S VOICE This way! Gillis swings around. Norma Desmond stands down the corridor next to a doorway from which emerges a flickering light. She is a little woman. There is a curious style, a great sense of high voltage about her. She is dress- ed in black house pyjamas and black high-heeled pumps. Around her throat there is a leopard-pat- terned scarf, and wound around her head a turban of the same material. Her skin is very pale, and she is wearing dark glasses. NORMA In here. I put him on my massage table in front of the fire. He always liked fires and poking at them with a stick. Gillis enters the SHOT and she leads him into A-32 NORMA DESMOND'S BEDROOM It is a huge, gloomy room hung in white brocade which has beconle dirty over the years and even slightly torn in a few places. There's a great, unmade gilded bed in the shape of a swan, from which the gold had begun to peel. There is a disorder of clothes and negligees and faded photographs of old-time stars about. In an imitation baroque fireplace some logs are burn- ing. On the massage table before it lies a small form shrouded under a Spanish shawl. At each end on a baroque pedestal stands a three-branched cande- labrum, the candles lighted. NORMA I've made up my mind we'll bury him in the garden. Any city laws against that? GILLIS I wouldn't know. NORMA I don't care anyway. I want the coffin to be white. And I want it specially lined with satin. White, or deep pink. She picks up the shawl to make up her mind about the color. From under the shawl flops down a dead arm. Gillis stares and recoils a little. It is like a child's arm, only black and hairy. NORMA Maybe red. bright flaming red. Gay. Let's make it gay. Gillis edges closer and glances down. Under the shawl he sees the sad, bearded face of a dead chimpanzee. Norma drops back the shawl. NORMA How much will it be? I warn you - don't give me a fancy price just because I'm rich. GILLIS Lady. you've got the wrong man. For the first time. Norma really looks at him through her dark glasses. GILLIS I had some trouble with my car. Flat tire. I pulled into your garage till I could get a spare. I thought this was an empty house. NORMA It is not. Get out. GILLIS I'm sorry, and I'm sorry you lost your friend, and I don't think red is the right color. NORMA Get out. GILLIS Sure. Wait a minute -- haven't I seen you -- ? NORMA Or shall I call my servant? GILLIS I know your face. You're Norma Desmond. You used to be in pictures. You used to be big. NORMA I am big. It's the pictures that got small. GILLIS I knew there was something wrong with them. NORMA They're dead. They're finished. There was a time when this busi- ness had the eyes of the whole wide world. But that wasn't good enough. Oh, nol They wanted the ears of the world, too. So they opened their big mouths, and out came talk, talk, talk... GILLIS That's where the popcorn business comes in. You buy yourself a bag and plug up your ears. NORMA Look at them in the front offices -- the master minds! They took the idols and smashed them. The Fairbankses and the Chaplins and the Gilberts and the Valentinos. And who have they got now? Some nobodies -- a lot of pale little frogs croaking pish-poshl GILLIS Don't get sore at me. I'm not an executive. I'm just a writer. NORMA You are! Writing words, words! You've made a rope of words and strangled this businessl But there is a microphone right there to catch the last gurgles, and Technicolor to photograph the red, swollen tongue! GILLIS Ssh! You'll wake up that monkey. NORMA Get out! Gillis starts down the stairs. GILLIS Next time I'll bring my autograph album along, or maybe a hunk of cement and ask for your footprints. He is halfway down the staircase when he is stopped by NORMA Just a minute, you! GILLIS Yeah? NORMA You're a writer, you said. GILLIS Why? Norma starts down the stairs. NORMA Are you or aren't you? GILLIS I think that's what it says on my driver's license. NORMA And you have written pictures, haven't you? GILLIS Sure have. The last one I wrote was about cattle rustlers. Before they were through with it, the whole thing played on a torpedo boat. Norma has reached him at the bottom of the staircase. NORMA I want to ask you something. Come in here. She leads him into A-33 THE HUGE LIVING ROOM It is dark and damp and filled with black oak and red velvet furniture which looks like crappy props from the Mark of Zorro set. Along the main wall, a gigantic fireplace has been freezing for years. On the gold piano is a galaxy of photographs of Norma Desmond in her various roles. On one wall is a painting -- a California Gold Rush scene, Carthay Circle school. (We will learn later that it hides a motion picture screen.) One corner is filled with a large pipe organ, and as Norma and Gillis enter, there is a grizzly moaning sound. Gillis looks around. NORMA The wind gets in that blasted pipe organ. I ought to have it taken out. GILLIS Or teach it a better tune. Norma has led him to the card tables which stand side by side near a window. They are piled high with papers scrawled in a large, uncertain hand. NORMA How long is a movie script these days? I mean, how many pages? GILLIS Depends on what it is -- a Donald Duck or Joan or Arc. NORMA This is to be a very important picture. I have written it myself. Took me years. GILLIS (Looking at the piles of script) Looks like enough for six impor- tant pictures. NORMA It's the story or Salome. I think I'll have DeMille direct it. GILLIS Uh-huh. NORMA We've made a lot of pictures together. GILLIS And you'll play Salome? NORMA Who else ? GILLIS Only asking. I did't know you were planning a comeback. NORMA I hate that word. It is a return. A return to the millions of people who have never forgiven me for deserting the screen. GILLIS Fair enough. NORMA Salome -- what a woman! What a part! The Princess in love with a Holy man. She dances the Dance of the Seven Veils. He rejects her, so she demands his head on a golden tray, kissing his cold, dead lips. GILLIS They'll love it in Pomona. NORMA (Taking it straight) They will love it every place. (She reaches for a batch of pages from the heap) Read it. Read the scene just before she has him killed! GILLIS Right now? Never let another writer read your stuff. He may steal it. NORMA I am not afraid. Read it! NORMA (Cont'd) (Calling) Max! Max! (To Gillis) Sit down. Is there enough light? GILLIS I've got twenty-twenty vision. Max has entered. NORMA Bring something to drink. MAX Yes. Madame. He leaves. Norma turns to Gillis again. NORMA I said sit down. There is compulsion in her voice. Gillis looks at her GILLIS' VOICE and starts slowly Well. I had no pressing reading. engagement, and she'd men- tioned something to drink.. Max comes in, wheeling Sometimes it's interesting a wicker tea wagon on to see just how bad bad which are two bottles o writing can be. This prom- f champagne and two ised to go the limit. I red Venetian glasses, wondered what a handwriting a box of zwieback and expert would make of that a jar of caviar. Norma childish scrawl of hers. sits on her feet. deep Max wheeled in some champagne in a chair, a gold ring and some caviar. Later, I on her forefinger with found out that Max was the a clip which holds a only other person in that cigarette. She gets up grim Sunset castle, and I and forces on Gillis found out a few other things another batch of script, about him... As for her, she goes back to her chair. sat coiled up like a watch spring, her cigarette clamped in a curious holder... I could sense her eyes on me from behind those dark glasses, defying me not to like what I read, or maybe begging me in her own proud way to like it. It meant so much to her... A-34 SHOT OF THE GILLIS' VOICE CEILING It sure was a cozy set-up. That bundle of raw nerves,and PAN DOWN to the moan- Max, and a dead monkey upstair ing organ. PAN OVER and the wind wheezing through TO THE ENTRANCE DOOR. that organ once in a while. Max opens it, and a Later on, just for comedy solemn-faced man in relief, the real guy arrived undertaker's clothes with a baby coffin. It was brings in a small all done with great dignity. white coffin. (Thru He must have been a very these shots the room important chimp. The great has been growing grandson of King Kong, maybe. duskier.) DISSOLVE TO: A-35 GILLIS It got to be eleven. I was feeling a little sick at my reading. The lamp stomach, what with that sweet beside him is now champagne and that tripe I'd really paying its been reading -- that silly way in the dark room. hodgepodge of melodramatic A lot of the manu- plots. However, by then I'd script pages are started concocting a little piled on the floor plot of my own... around his feet. A half-empty champagne glass stands on the arm of his chair. THE CAMERA SLOWLY DRAWS BACK to include Norma Desmond sitting in the dusk, just as she was before. Gillis puts down a batch of script. There is a little pause. NORMA (Impatiently) Well? GILLIS This is fascinating. NORMA Of course it is. GILLIS Maybe it's a little long and maybe there are some repetitions... but you're not a professional writer. NORMA I wrote that with my heart. GILLIS Sure you did. That's what makes it great. What it needs is a little more dialogue. NORMA What for? I can say anything I want with my eyes. GILLIS It certainly could use a pair of shears and a blue pencil. NORMA I will not have it butchered. GILLIS Of course not. But it ought to be organized. Just an editing job. You can find somebody. NORMA Who? I'd have to have somebody I can trust. When were you born -- I mean, what sign of the zodiac? GILLIS I don't know. NORMA What month? GILLIS December twenty-first. NORMA Sagittarius. I like Sagittarians. You can trust them. GILLIS Thank you. NORMA I want you to do this work. GILLIS Me? I'm busy. Just finished one script. I'm due on another assignment. NORMA I don't care. GILLIS You know, I'm pretty expensive. I get five hundred a week. NORMA I wouldn't worry about money. I'll make it worth your while. GILLIS Maybe I'd better take the rest of the script home and read it - NORMA Oh no. I couldn't let it out of my house. You'll have to finish it here. GILLIS It's getting kind of late -- NORMA Are you married, Mr. -- ? GILLIS The name is Gillis. I'm single. NORMA Where do you live? GILLIS Hollywood. The Alto Nido Apart- ments. NORMA There's something wrong with your car, you said. GILLIS There sure is. NORMA You can stay here. GILLIS I'll come early tomorrow. Norma takes off her glasses. NORMA Nonsense. There's room over the garage. Max will take you there...Max! THE CAMERA MOVES GILLIS' VOICE TOWARD NORMA'S FACE, She sure could say a lot of right up to her things with those pale eyes of eyes. hers. They'd been her trade mark. They'd made her the Num- ber One Vamp of another era. I remember a rather florid des- cription in an old fan magazine which said: "Her eyes are like two moonlit waterholes, where strange animals come to drink." DISSOLVE TO: A-36 SMALL STAIRCASE, LEAD- GILLIS'VOICE ING TO ROOM OVER GARAGE I felt kind of pleased with the way I'd handled the sit- Max, an electric light uation. I'd dropped the hook, bulb in his hand, is and she'd snapped at it. Now leading Gillis up. my car would be safe down Gillis carries a batch below, while I did a patch- of the manuscript. up job on the script. And there should be plenty of money in it... Max pushes open a door at the top of the stairs. MAX (Opening the door) I made your bed this afternoon. GILLIS Thanks. (On second thought) How did you know I was going to stay, this afternoon? Max doesn't answer. He walks across to the bed, screws a bulb in the open socket above it. The light goes on, revealing: A-37 A GABLED BEDROOM There are dirty windows on two sides, and dingy wall- paper on the cracked plaster walls. For furniture there is a neatly made bed, a table and a few chairs which might have been discarded from the main house. MAX This room has not been used for a long time. GILLIS It will never make house Beautiful. I guess it's O.K. for one night. Max gives him an enigmatic look. MAX (Pointing) There is the bathroom. I put in soap and a toothbrush. GILLIS Thanks. (He starts taking off his coat) Say, she's quite a character, that Norma Desmond. MAX She was the greatest. You wouldn't know. You are too young. In one week she got seventeen thousand fan letters. Men would bribe her mani- curist to get clippings from her fingernails. There was a Maharajah who came all the way from Hyderabad to get one of her stockings. Later, he strangled himself with it. GILLIS I sure turned into an interesting driveway. MAX You did, sir. GILLIS' VOICE He goes out. Gillis I pegged him as slightly looks after him, hangs cuckoo, too. A stroke maybe. his coat over a chair, Come to think of it, the walks over to the win- whole place seemed to have dow, pulls down the been stricken with a kind of rickety Venetian blind. creeping paralysis, out of As he does so, he looks beat with the rest of the down at: world, crumbling apart in slow motion ... A-38 THE TENNIS COURT OF GILLIS' VOICE THE DESMOND HOUSE There was a tennis court, or (MOONLIGHT) rather the ghost of a tennis court, with faded markings The cement surface is and sagging net ... cracked in many places, and weeds are growing high. A-39 GILLIS - IN THE WINDOW He looks away from the court to: A-40 THE DESMOND SWIMMING POOL GILLIS' VOICE There is no water in And of course she had a pool. it, and hunks of Who didn't then? Mabel Norm- mosaic which lines its and and John Gilbert must enormous basin are have swum in it ten thousand broken away. midnights ago, and Vilma Banky and Rod La Roque. It was empty now....or was it? A-41 GILLIS - IN THE WINDOW He stares down, his stomach slowly turning. A-42 THE SWIMMING POOL At the bottom of the basin a great rat is eating a decaying or,ange. From the inlet pipe crawl two other rats, who join battle with the first rat over the orange. A-43 GILLIS -IN THE WINDOW He starts away, but some- GILLIS' VOICE thing attracts his atten- There was something tion. He turns back and else going on below: looks down again. the last rites for that hairy old chimp, performed with the A-44 THE LAWN BELOW utmost seriousness -- as if she were laying Norma Desmond and Max are to rest an only child. carrying the white coffin Was her life really towards a small grave as as empty as that? which has been dug in the dead turf. Norma carries one of the candelabra, all of its candles flickering in the wind. They reach the grave and lower the coffin into it. Then, Norma lighting his task with the candelabrum, Max takes a spade from the loose earth and starts filling in the grave. A-45 GILLIS - IN THE WINDOW He watches the scene be- GILLIS' VOICE low, then turns into the It was all very queer, room, goes to the door but queerer things to lock it. There is no were yet to come. key, and only a hole where the lock has been gouged out. Gillis moves a heavy overstuffed chair in front of the door, then walks towards the bed, throws himself on it, picking up some of the manuscript pages to read. DISSOLVE END OF SEQUENCE "A" SEQUENCE "B" DISSOLVE IN ON: B-1 LONG SHOT THE DESMOND HOUSE - (MORNING) The day is overcast. The SOUND: (Distant organ house is shrouded in low music - improvisations fog. on an odd, mournful theme - not too loud, continuing throughout B-2 THE TENNIS COURT, blurred the scene.) over with fog. B-3 THE EMPTY SWIMMING POOL Its dark outline even more melancholy under the misty blanket. B-4 THE ROOM OVER THE GARAGE Muted daylight seeps GILLIS' VOICE through the blinds. Gillis That night I'd had a lies on the bed, under a mixed-up dream. In it shabby quilt. The manu- was an organ grinder. script is beside him, some I couldn't see his of the pages scattered on face, but the organ the floor. He is just was all draped in opening his eyes. It takes black, and a chimp was him a moment to adjust him- dancing for pennies. self to the strange sur- When I opened my eyes, roundings. His eyes, wander- the music was still ing about the room. suddenly there... Where was stop, startled. He lifts I? himself on one elbow and stares at - B-5 THE DOOR The heavy chair he had set Oh yes, in that empty against it the night before room over her garage. has been pushed back. The Only it wasn't empty door is wide ajar. any more. Somebody had brought in all my belongings - my B-6 GILLIS books, my typewriter, my clothes... He jumps out of bed. He wears, shirt, trousers and socks. Suddenly he realizes that all his possessions have GILLIS' VOICE been brought in. In What was going on? the closet hang his shirts. His books and typewriter are neatly arranged on the table. His phonograph-radio combination is all installed. Gillis looks around startled, then sits down and starts putting on his moccasins hastily. DISSOLVE TO: B-7 A PAIR OF HANDS IN WHITE GLOVES, PLAYING THE ORGAN PULL BACK: They belong to Max von Mayerling. He is sitting erect, his bull neck taut as a wrestler's as he rights out somber chord after somber chord. He sits in a shaft of gray light coming from an open French window. Through the far archway, Gillis storms into the big room. GILLIS Hey, you -- Max -- whatever -your- name-is -- what are my things doing here? No answer. GILLIS I'm talking to you. My clothes and things are up in the room. MAX Naturally. I brought them myself. GILLIS (Furiously) Is that so! MAX Why are you so upset? Is there anything missing? GILLIS Who said you could? Who asked you to? Norma Desmond's shadow moves into the shaft of light. NORMA'S VOICE I did. Gillis looks around. On the couch by the fireplace reclines Norma Desmond, dressed in a negligee. She rises. NORMA I don't know why you should be so upset. Stop that playing, Max. (To Gillis again) It seemed like a good idea -- if we are to work together. GILLIS Look, I'm supposed to fix up your script. There's nothing in the deal about my staying here. NORMA You'll like it here. GILLIS Thanks for the invitation, but I have my own apartment. NORMA You can't work in an apartment where you owe three months' rent. GILLIS I'll take care of that. NORMA It's all taken care of. It's all paid for. GILLIS I'm used to paying my own bills. NORMA You proud boy, why didn't you tell me you were having difficulties. GILLIS Okay. We'll deduct it from my salary. NORMA Now, now, don't let's be small about such matters. We won't keep books. (To Max) Go on, unpack Mr. Gillis' things. GILLIS Unpack nothing. I didn't say I was staying. NORMA (Her glasses off again) Suppose you make up your mind. Do you want this job or don't you? DISSOLVE TO: B-8 BIG ROOM, NORMA DESMOND'S HOUSE - (DAY) GILLIS' VOICE Gillis sits at an impro- So I let him unpack my vised table, his typewriter things. I wanted the in front of him, working dough, and I wanted to hard at the manuscript. get out of there as Pencils, shears and a quickly as possible. paste-pot at hand. I thought if I really got going I could toss Facing him at some dis- it off in a couple or tance sits Norma,dressed weeks. But it wasn't in another version of her so simple, getting some favorite lounging pajamas, coherence into that wild, the cigaette contraption scrambled melodrama on her finger. She is she'd concocted. What autographing large photo- made it tougher was that graphs of herself and put- she was around all the ting them in envelopes. time -- hovering over me, afraid I'd do injury to that precious brain- child of hers. Gillis takes two or three pages from Norma's hand- written script, crosses them out and puts them to one side. Norma rises, crosses towards Gillis, looks over his shoulder. NORMA What's that? GILLIS Just a scene I cut out. NORMA What scene? GILLIS The one where you go to the slave market. You can cut right to the scene where John the Baptist - NORMA Cut away from me? GILLIS Honestly, it's a little old hat. They don't want that any more. NORMA They don't? Then why do they still write me fan letters every day. Why do they beg me for my photo- graphs? Because they want to see me, me, me! Norma Desmond. GILLIS (Resigned) Okay. He pulls the page from his typewriter. As he does so he glances over towards Norma. GILLIS' VOICE On the table in front I didn't argue with her. of her are the photo- You don't yell at a graphs which she is sign- sleepwalker-- he may fall ing. On the long table and break his neck.That's in the living room is a it -- she was still gallery of photographs sleepwalking along the in various frames -- all giddy heights of a lost Norma Desmond. On the career --plain crazy piano more photographs. when it came to that one Above the piano an oil subject: her celluloid portrait of her. On the self, the great Norma highboy beside him still Desmond. How could She more photographs. breathe in that house, so crowded with Norma DISSOLVE TO: Desmonds? More Norma Desmond and still more Norma Desmond. B-9 THE BIG ROOM - (NIGHT) GILLIS' VOICE Shooting towards the big It wasn't all work - of Gold Rush painting. Max, course. Two or three white gloves and all, times a week Max would steps into the shot, shoves haul up that enormous oil the painting up towards painting that had been the ceiling,revealing a presented to her by some motion picture screen. Nevada Chamber of Com- Max exits. merce, and we'd see a movie,right in her living room. B-1O NORMA AND GILLIS GILLIS' VOICE They sit on a couch,facing "So much nicer than going the screen. On a table in out," she'd say. The front of them are champagne, plain fact was that she cigarettes and coffee. was afraid of that world Above their heads are the outside. Afraid it typical openings for a pro- would remind her that jector. The lights go off. time had passed. From the opening above their heads shoots the wide beam of light. B-11 MAX, IN THE PROJECTION They were silent movies, BOOTH BEHIND THE ROOM and Max would run the projection machine, which The light of the machine was just as well -- it flickering over his face, kept him from giving us which is frozen, a somber an accompaniment on enigma. that wheezing organ. B-12 NORMA AND GILLIS She'd sit very close to watching the screen. me, and she'd smell of Gillis looks down and sees tuberoses, which is not that Norma's hand is clasp- my favorite perfume, not ing his ann tight. He by a long shot. Sometines doesn't like it much but as we watched, she'd c he can't do anything about lutch my arm or my hand it. However. when she for forgetting she was my a second lets go his arm employer becoming just a to pick up a glass of fan, excited about that champagne, he gently with- actress up there on the draws his arm, leans away screen....I guess I don't from her and crosses his have to tell you who the arms to discourage any star was. They were resumption of her approach. always her pictures -- Norma puts the glass down that's all she wanted doesn't find his arn, but to see. is not aware of any signifi- cance in his maneuver. They both watch the screen. B-13 THE OTHER END OF THE BIG ROOM. WITH THE SCREEN On it flickers a famous scene from one of Norma's old silent pictures. It is not to be a funny scene. It is old-fashioned, but shows her incredible beauty and the screen presence which made her the great star of her day. B-14 NORMA AND GILLIS ON THE COUCH NORMA Still wonderful, isn't it? And no dialogue. We didn't need dialogue. We had faces. There just aren't any faces like that any more. Well, maybe one -- Garbo. In a sudden flareup she jumps to her feet and stands in the flickering beam of light. NORMA Those idiot producers! Those imbeciles! Haven't they got any eyes? Have they forgotten what a star looks like? I'll show them. I'll be up there again. So help me! DISSOLVE TO: B-15 THE BIG ROOM - (NIGHT) It is apparently empty. GILLIS' VOICE The elaborate lamps Sometimes there'd be a make pools of light. little bridge game in the house, at a twentieth-of- THE CAMERA PULLS BACK a cent a point. I'd get AND PANS to reveal a half her winnings. Once card table around they ran up to seventy which sit Norma and cents, which was about three friends - three the only cash money I actors of her period. ever got. The others They sit erect and play around the table would with grim seriousness. be actor friends - dim figures you may still Beside Norma sits remember from the silent Gillis, kibitzing on a days. I used to think of game which bores him them as her Wax Works. extremely. An ashtray on the card table is full and Norma holds it out for Gillis to take away. He crosses the room to the fire- place. but his eyes fall on the entrance door and he stops. B-16 THE ENTRANCE HALL - (FROM GILLIS' POINT OF VIEW) Max stands in the open door. Outside are the two men who came to the apartment for Gillis' car. B-17 GILLIS He steps back so that he cannot be seen from the door. A second later Max appears, looking for him. MAX (Quietly) Some men are here. They asked for you. GILLIS I'm not here. MAX That's what I told them. GILLIS Good. MAX They found your car in the garage. They are going to tow it away. Gillis doesn't know what to do. From offstage comes: NORMA'S VOICE The ashtray, Joe dear! Can we have the ashtray? Gillis dumps the cigarette butts into the cold fire- place, crosses to the bridge table, puts the ashtray down, leans over and speaks into Norma's ear. GILLIS I want to talk to you for a minute. NORMA Not now, my dear. I'm playing three no trump. GILLIS They've come for my car. NORMA Please. Now I've forgotten how many spades are out. GILLIS I need some money right now. NORMA Can't you wait till I'm dummy? 3.22.49 GILLIS No. NORMA (Angry by now) Please! Gillis stands frustrated, hideously embarrassed by the stares of the waxworks. He turns away and hurries to the door. B-18 ENTRANCE DOOR TO THE HOUSE It is half open. Gillis comes into the shot and, taking cover, looks out. B-19 COURTYARD (FROM GILLIS' ANGLE) The men from the finance company are cranking up the car. Max stands watching silently. When they finish the cranking job, the men climb into the front seat of the truck. B-2O GILLIS - AT THE DOOR Over the shot the SOUND of the truck being started and the cars moving away. Gillis moves out into the courtyard and stands staring after the car. From the house comes Norma. NORMA Now what is it? Where's the fire? GILLIS I've lost my car. NORMA Oh...and I thought it was a matter of life and death. GILLIS It is to me. That's why I came to this house. That's why I took this job -- ghost writing! NORMA Now you're being silly. We don't need two cars. We have a car. And not one of thuse cheap new things made of chromium and spit. An Isotta-Fraschini. Have you ever heard of Isotta-Fraschinis? All hand-made. Cost me twenty-eight thousand dollars. THE CAMERA HAS PANNED over to the garage and FOCUSES on the dirty Isotta-Fraschini on its blocks. DISSOLVE TO: B-21 NORMA'S ISOTTA-FRASCHINI DRIVING IN THE HILLS ABOVE SUNSET (DAY) Max is at the wheel, GILLIS' VOICE dressed as usual except So Max got that old bus for a chauffeurfs cap. down off its blocks and polished it up. She'd take me for rides in the B-22 INSIDE THE CAR hills above Sunset. Gillis sits beside Norma, The whole thing was up- who is wearing a smart holstered in leopard tailleur and her eternal skin, and had one of sun glasses. Gillis those car phones, all wears his sport jacket- gold-plated. flannel trousers-moccasin combinatIon. He sits uncomfortably. Norma is studying him. NORMA That's a dreadful shirt you're wearing. GILLIS What's wrong with It? NORMA Nothing, if you work in a fill- ing station. And I'm getting rather bored with that sport jacket, and those same baggy pants. (She picks up the car phone) Max, what's a good men's shop in town? The very best... Well, go there ! GILLIS I don't need any clothes, and I certainly don't want you buy- ing them for -- NORMA Why begrudge me a little fun? I just want you to look nice, my stray little boy. By this time Max has made a U-turn. QUICK DISSOLVE TO: B-23 INT. MEN'S DEPARTMENT, AN ELEGANT WILSHIRE STORE Gillis stands in front of a full-length triple mirror, surrounded by a couple of salesmen and the tailor, who is busily working out alterations. Gillis wears a double-breasted gray flannel coat with chalk stripes. His trousers belong to another suit of glen plaid. Norma is running the show. NORMA There's nothing like gray flannel with a chalk stripe. (she points at the trousers) This one single-breasted, of course. (to another salesman) Now we need a topcoat. Let's see what you have in camel's hair. The salesman leaves. NORMA How about some evening clothes? GILLIS I don't need a tuxedo. NORMA Of course you do. A tuxedo and tails. GILLIS Tails. That's ridiculous. NORMA You'll need them for parties. You'll need them for New Year's Eve. (to a salesman) Where are your evening clothes? SALESMAN This way, Madame. He leads her off. The other salesman arrives with a selection of topcoats. SALESMAN Here are some camel hairs, but I'd like you just to feel this one. It's Vicuna. Of course, it's a little more expensive. GILLIS A camel's hair will do. SALESMAN (With an insulting inflection) As long as the lady is paying for it, why not take the Vicuna? DISSOLVE: END OF SEQUENCE "B" SEQUENCE "C" DISSOLVE IN: C-1 LONG SHOT DESMOND HOUSE A day in December. Rain. QUICK DISSOLVE TO: C-2 INT. ROOM OVER GARAGE Water is drizzling from GILLIS' VOICE two or three spots in the The last week in December ceiling into pans and the rains came -- a great bowls set to catch it, big package of rain. one bowl right on the Over-sized, like every- bed. The room is almost thing else in California. emptied of Gillis' be- longings by now. Max It came right through is carrying out a hand- the old roof of my room full of new suits on above the garage. She hangers. He has a had Max move me to the dressing gown over his main house. I didn't shoulder. Gillis holds much like the idea -- the a stack of shirts, his only time I could have typewriter, and some to myself was in that manuscript. He surveys room -- but it was better the room for the last than sleeping in a rain- time, to see whether coat and galoshes. he's forgotten any- thing. He has. He puts down the typewriter and picks up from under the bed a pair of very smart red leather bedroom slippers. He tucks them under his arm, picks up the typewriter and leaves. QUICK DISSOLVE TO: C-3 A BEDROOM IN TIiE MAIN HOUSE It is obviously a man's room -- heavy Spanish furniture -- one wall nothing but a closet with shelves and drawers for shirts and shoes. Max is hanging up the suits. Gillis throws the shirts on a big chair, tosses the slippers at the foot of the bed, places the typewriter and manuscript on a desk at the window. GILLIS Whose room was this? MAX It was the room of the husband. Or of the husbands, I should say. Madame has been married three times. Slightly embarrassed, Gillis picks up his toilet kit with razor, toothbrushes, soap, etc., and starts towards the bathroom, pausing en route at a rain- splattered window. GILLIS I guess this is the one you can see Catalina from. Only this isn't the day. He proceeds towards the half-opened door leading to the bathroom. Something strikes his attention and he stops. As in the door to the room above the garage, this lock, too, has been gouged out. GILLIS Hey, what's this with the door? There isn't any lock. MAX There are no locks anywhere in this house. He points to the entrance door of the room, and to another door. GILLIS How come? MAX The doctor suggested it. GILLIS What doctor? MAX Madame's doctor. She has moments of melancholy. There have been some suicide attempts. GILLIS Uh-huh? MAX We have to be very careful. No sleeping pills, no razor blades. We shut off the gas in her bed- room. GILLIS Why? Her career? She got enough out of it. She's not forgotten. She still gets those fan letters. MAX I wouldn't look too closely at the postmarks. GILLIS You send them. Is that it, Max? MAX I'd better press your evening clothes, sir. You have not for- gotten Madame's New Year's party. GILLIS No, I haven't. I suppose all the waxworks are coming? MAX I don't know, sir. Madame made the arrangements. Max leaves. Gillis comes out of the bathroom, picks up his shirts, goes over to a closet, opens it. As he does so one of the doors without a lock swings slightly open. Gillis looks through the half-open door and sees. C-4 NORMA DESMOND'S ROOM It is empty. The rainy GILLIS' VOICE day does nothing to There it was again - that help its gloom. room of hers, all satin and ruffles, and that bed like a gilded rowboat. The per- fect setting for a silent movie queen. Poor devil, still waving proudly to a parade which had long since passed her by. He pushes the door shut and walks back into the room. DISSOLVE TO: C-5 STAIRCASE OF DESMOND HOUSE (NIGHT) Gillis is coming down the GILLIS' VOICE stairs in his tailcoat It was at her New Year's adjusting the handkerchief party that I found out in his pocket. He obviously how she felt about me. feels a little uneasy in Maybe I'd been an idiot this outfit. From below not to have sensed it comes a tango of the Twen- was coming - that sad, ties. played by a small embarrassing revelation. orchestra. Gillis stops in the archway leading to the big room and looks around. C-6 THE BIG ROOM has been deco- rated for the occasion with laurel garlands. Dozens of candles in all the sconces and candelabra are ablaze. Their flickering flames are reflected in the waxed sur= face of the tile floor. There is a buffet, with buckets of champagne and caviar on ice. In one corner on a little platform banked with palms. a four-piece orchestra is playing. At the buffet are Max and Norma. She is drinking a glass of champagne. She is wearing a diamonte evening dress. very high style. with long black gloves and a headdress of paradise feathers. Her eyes fall on Gillis. She puts down the glass of champagne. picks up a gardenia boutonniere and moves toward him. NORMA Joe, you look absolutely divine. Turn around! GILLIS (Embarrassed} Please. NORMA Come on! Gillis makes a slow 36O-degree turn. NORMA Perfect. Wonderful shoulders. And I love that line. She indicates the V from his shoulders to his hips. GILLIS All padding. Don't let it fool you. NORMA Come here! She puts the gardenia on his lapel. GILLIS You know, to me dressing up was always just putting on my dark blue suit. NORMA I don't like those studs they've sent. I want you to have pearls. Nice big pearls. GILLIS Now, I'm not going to wear ear- rings, I can tell you that. NORMA Cute. Let's have some drinks. She leads him over to the buffet. GILLIS Shouldn't we wait for the others? NORMA (Pointing at the floor) Careful, it's slippery. I had it waxed. They reach the buffet. Max is ready with two glasses of champagne. Norma hands Gillis a glass. NORMA Here's to us. They drink. NORMA You know, this floor used to be wood but I had it changed. Valentino said there is nothing like tiles for a tango. She opens her arms. GILLIS Not on the same floor with Valentino! NORMA Just follow me. They start to tango. After a moment -- NORMA Don't bend back like that. GILLIS It's those feathers. They tickle. Norma pulls the paradise feathers from her hair and tosses them away. C-7 THE ORCHESTRA As they play the tango, the musicians eye the danc- ing couple, take in the situation, exchange glances and turn away with professional discretion. C-8 NORMA AND GILLIS, TANGOING Gillis glances at his wrist watch. GILLIS It's a quarter past ten. What time are they supposed to get here? NORMA Who? GILLIS The other guests? NORMA There are no other guests. We don't want to share this night with other people. This is for you and me. GILLIS I understand some rich guy bought up all the tickets for a perfor- mance at the Metropolitan and sat there listening to La Traviata, all by himself. He was afraid of catching cold. NORMA Hold me tighter. GILLIS Come midnight, how about blind- folding the orchestra and smash- ing champagne glasses on Max's head? NORMA You think this is all very funny. GILLIS A little. NORMA Is it funny that I'm in love with you? GILLIS What's that? NORMA I'm in love with you. Don't you know that? I've been in love with you all along. They dance on. Gillis is acutely embarrassed. THE CAMERA SLOWLY PULLS BACK, PANS past the faces of the musicians, who play on with a rather overe- mphasized lack of interest. Finally it winds up on Max, behind the buffet. He stands watching Gillis, a faint trace of pity in his eyes. DISSOLVE TO: C-9 NORMA'S FINGER, WITH THE CIGARETTE GADGET, as she GILLIS' VOICE inserts a cigarette. I'm sure a lot of you will laugh about this. Ridicu- lous situation, wasn't it? -- a woman almost twice my age ... It got to be about a quarter of eleven. I felt caught, like a cig- arette in the prongs of that contraption on her finger. PULL BACK TO: NORMA AND GILLIS sitting on a couch in front of the cavernous fireplace. Norma holds out her cigarette to Gillis, who lights it. NORMA. What a wonderful next year it's going to be. What fun we're going to have. I'II fill the pool for you. Or I'll open my house in Malibu, and you can have the whole ocean. Or I'll buy you a boat and we'll sail to Hawaii. GILLIS Stop it. You aren't going to buy me anything more. NORMA Don't be silly. (She reaches under a pillow of the couch and brings out a leather box) Here. I was going to give it to you at midniglht. Gillis opens the box. It contains a matched gold cigarette case and lighter. NORMA Read what's inside. Gillis snaps open the case. Engraved inside the cover is: TO JOE FROM NORMA, and two bars of music. GILLIS What are the notes? NORMA "Mad about the boy." GILLIS Norma, I can't take it. You've bought me enough. NORMA Shut up. I'm rich. I'm richer than all this new Hollywood trash. I've got a million dollars. GILLIS Keep it. NORMA I own three blocks downtown. I have oil in Bakersfield -- pumping, pumping, pumping. What's it for but to buy us anything we want. GILLIS Cut out that us business. He rises. NORMA What's the matter with you? GILLIS What right do you have to take me for granted? NORMA What right? Do you want me to tell you? GILLIS Has it ever occurred that I may have a life of my own? That there may be some girl I'm crazy about? NORMA Who? Some car hop, or a dress extra? GILLIS Why not? What I'm trying to say is that I'm all wrong for you. You want a Valentino -- somebody with polo ponies -- a big shot -- NORMA (Getting up slowly) What you're trying to say is that you don't want me to love you. Is that it? Gillis doesn't answer. Norma slaps his face and rushes from the room and upstairs. Gillis stands paralyzed, the slap burning his cheek. C-1O THE TOP OF THE STAIRCASE AND CORRIDOR Norma rushes up the last few steps, down the corridor and into her bedroom, banging the door. MOVE THE CAMERA toward the closed door, centering on the gouged-out lock. C-11 GILLIS, IN THE BIG ROOM He still stands motionless. He glances around fur- tively, to see if his humiliation has been observed. C-12 THE ORCHESTRA The musicians are playing away. They have turned their eyes away from Gillis rather too ostentatious- ly for comfort. C-13 GILLIS His eyes move over toward C-14 MAX He is subtler than the musicians. He appears very busy at the buffet, putting empty bottles and used glasses on a tray. He walks across the room with them. C-15 GILLIS He starts slowly out. As he does so his long gold key chain catches on a carved ornament of the sofa and holds him for a second of additional embarrass- ment. He yanks it loose and walks with as much nonchalance as he can muster to C-16 THE HALL Crossing towards the coat closet, Gillis throws a look upstairs. Then he pulls the Vicuna coat from its hangar and slips into it as he crosses to the entrance door. He opens the door on the darkness of the courtyard. C-17 EXT. DESMOND HOUSE (NIGHT - RAIN) Gillis shuts the door. GILLIS'VOICE He takes a few steps I didn't know where I was forward, then stands going. I just had to get for a while breathing out of there. I had to be deep. The rain is with people my own age. I balm to that cheek had to hear somebody laugh where the slap still a again. I thought of Artie burns. He walks for- Green. There was bound to ward with a great be a New Year's shindig sense of relief. going on in his apartment down on Las Palmas -- the hock shop set -- not a job C-18 DRIVEWAY LEADING TO in the room. but lots of fun on the cuff. Gillis walks to the street, which is dark and empty. He starts down Sunset in an Easterly direction. A car passes. He tries to thumb a ride, without success. However, the second car, a florist's delivery wagon, stops. Gillis jumps in and the car drives off. DISSOLVE TO: C-19 ARTIE GREEN'S APARTMENT It is the most modest one-room affair, jam packed with young people flowing over into the miniature bathroom and the microscopic kitchenette. The only drink being served is punch from a pressed-glass bowl -- but everybody is having a hell of a time. Most of the men are in slacks and sweaters, and only a few of the girls in something that vaguely suggests party dress. Abe Burroughs sits at a small, guest-festooned piano and sings Tokio Rose. By the door, a group of young men and girls respond to the song by sing1ng Rinso White or Dentyne Chewing Gum or something similar, in the manner of a Bach choral. Artie Green, a dark haired, pleasant-looking guy in his late twenties, is conducting with the ladle from the punch bowl. The door behind some of the singers is pushed open, jostling them out of their places. In comes Gillis, his hair and face wet, the collar of his Vicuna coat turned up. Artie stops conducting, but the commer- cial goes right on. ARTIE Well, what do you know ! Joe Gillis ! GILLIS Hi, Artie. ARTIE Where have you been keeping that gorgeous face of yours? GILLIS In a deep freeze. ARTIE I almost reported you to the Bureau of Missing Persons. (To the company) Fans, you all know Joe Gillis, the well-known screen writer, opium smuggler and Black Dahlia suspect. Gillis greets some of the kids by name as he and Artie push their way into the room. ARTIE Give me your coat. GILLIS Let it ride for a while. ARTIE You're going to stay, aren't you? GILLIS That was the general idea. ARTIE Come on. Artie starts peeling the coat off Gillis. Its texture takes his breath away. ARTIE What is this - mink? He has taken the coat. He looks at Gillis standing there in tails. ARTIE Judas E. Priest, who did you borrow that from? Adolphe Menjou? GILLIS Close, but no cigar. Gillis stands embarrassed While Artie rolls up the Vicuna coat and tucks it above the books on a book- shelf. ARTIE Say, you're not really smuggling opium these days, are you? GILLIS Where's the bar? The two make their way toward the punch bowl. It's a little like running the gauntlet for Gillis. There are whistles and 'stares of astonishlnent at his tails. When they reach the punch bowl, Artie picks up a half-filled glass and fills it. GILLIS Good party. ARTIE The greatest. They call me the Elsa Maxwell of the assistant directors. (To some guests who are dipping their empty cups into the punch bowl) Hey, easy on the punch bowl. Budget only calls for three drinks per extra. Fake the rest. GILLIS Listen, Artie, can I stick around here for a while? ARTIE Sure, this'll go on all night. GILLIS I mean, could you put me up for a couple of weeks? ARTIE It just so happens we have a vacancy on the couch. GILLIS I'll take it. ARTIE I'll have the bell-hop take care of your luggage. He runs his finger across the decollete back of a girl standing in a group next them. ARTIE Just register here. The girl turns around. She is Betty Schaefer. BETTY Hello, Mr. Gillis. ARTIE You know each other? Gillis looks at her a little puzzled. BETTY Let me help you. Betty Schaeter, Sheldrake's office. GILLIS Sure. Bases Loaded. ARTIE Wait a minute. This is the woman I love. What's going on? Who was loaded? GILLIS Don't worry. She's just a fan for my literary output. BETTY (to Artie) Hurt feelings department. GILLIS About that luggage. Where's the phone? ARTIE Over by the Rainbow Room. Gillis squeezes his way through groups of people to the telephone, which is next to an open door leading to the bathroom. The phone is busy. A girl sits listening to it, giggling wildly. Another girl beside her is laughing too. They are apparently sharing a conversation with some man on the other end of the wire. The telephone passes from hand to hand. Gillis watches impatiently, then GILLIS When youlre through with that thing, can I have it? The girl just nods, going on with her chattering. Gillis stands waiting, and Betty Schaefer comes up with his glass. BETTY You forgot this. GILLIS Thanks. BETTY I've been hoping to run into you. GILLIS What for? To recover that knife you stuck in my back? BETTY I felt a little guilty, so I got out some of your old stories. GILLIS Why, you sweet kid. BETTY There's one called....Window... something with a window. GILLIS Dark Windows. How did you like it? BETTY I didn't. GILLIS Thank you. BETTY Except for about six pages. You've got a flashback there ... There is too much racket for her. BETTY Is there someplace we can talk? GILLIS How about the Rainbow Room? They squeeze their way towards the bathroom, past Artie. ARTIE I said you could have my couch. I didn't say you could have my girl. BETTY This is shop talk. She and Gillis go through the open door into C-20 ARTIE'S BATHROOM It's a little less noisy, although there are some guests there, chatting and having fun. Betty and Gillis sit down on the edge of the tub. GILLIS Now if I got you correctly, there was a short stretch of my fiction you found worthy of notice. BETTY The flashback in the courtroom, when she tells about being a school teacher. GILLIS I had a teacher like that once. BETTY Maybe that's why it's good. It's true, it's moving. Now why don't you use that character... GILLIS Who wants true? Who wants moving? BETTY Drop that attitude. Here's some- thing really worth while. GILLIS Want me to start right now? Maybe there's some paper around. BETTY I'm serious. I've got a few ideas. GILLIS I've got some ideas myself. One of them being this is New Year's Eve. How about living it up a little? BETTY As for instance? GILLIS Well.... BETTY We could make some paper boats and have a regatta. Or should we just turn on the shower? GILLIS How about capturing the kitchen and barricading the door? BETTY Are you hungry? GILLIS Hungry? After twelve years in the Burmese jungle. I am starving, Lady Agatha -- starving for a white shoulder -- BETTY Phillip, you're mad! One of the girls who was on the phone comes to the door. GIRL You can have the phone now. GILLIS (Paying no attention) Thirsting for the coolness of your lips - BETTY No, Phillip, no. We must be strong. You're still wearing the uniform of the Coldstream Guards! Furthermore, you can have the phone now. GILLIS O.K. (He gets up, starts out, turns) I find I'm terribly afraid of losing you. BETTY You won't. (She takes the glass out of his hand) I'll get us a refill of this awful stuff. GILLIS You'll be waiting for me? BETTY With a wildly beating heart. GILLIS Life can be beautiful! He leaves. C-21 THE MAIN ROOM Gillis squeezes himself through some guests to the phone. He has to stand in a cramped position, holding the instrument close to him as he dials a number. GILLIS Max? This is Mr. Gillis. I want you to do me a favor. C-22 NORMA DESMOND HOUSE Max is at the phone, in the lower hall. MAX I am sorry, Mr. Gillis. I cannot talk now. C-23 GILLIS ON THE PHONE GILLIS Yes you can. I want you to get my old suitcase and I want you to throw in my old clothes -- the ones I came with, and my typewriter. I'll have somebody pick them up. C-24 MAX AT THE PHONE MAX I have no time to talk. The doctor is here. C-25 GILLIS ON THE PHONE GILLIS What doctor? What's going on? C-26 MAX AT THE PHONE MAX She got the razor from your room. She cut her wrists. Max hangs up, moves toward the staircase. C-27 GILLIS AT THE PHONE GILLIS Max ! Max ! He hangs up the dead receiver, stands numb with shock. Betty elbows her way up to him, carrying the two punch glasses filled again. BETTY I just got the recipe: take two packages of cough drops, dissolve in one gallon of lukewarm grape juice -- Gillis looks up at her. Without a word he pushes her aside so that she spills the drink. He makes his way through the guests to the Vicuna coat, pulls it from the shelf, some books tumbling with it, and rushes towards the door and out. Betty stands look- ing after him, completely bewildered. DISSOLVE TO: C-28 EXT. DESMOND HOUSE - (NIGHT, RAIN) The doctor's car is parked in the driveway. A taxi pulls up. Gillis, in his Vicuna coat now, jumps out, throws a couple of dollars to the rdriver and runs toward the house. C-28a DOORWAY, NORMA DESMOND HOUSE> Max is opening the door to let out the doctor, a professional looking man carrying a black bag. Gillis runs into the SHOT. GILLIS How is she? MAX She is upstairs. Gillis starts to push past Max. Max grabs his arm. MAX Be careful. Do not race up the stairs. The musicians must not know what has happened. Gillis goes into the house. C-29 ENRANCE HALL AND STAIRCASE Gillis crosses the hall and starts up the stairs. C-3O INT. NORMA DESMOND'S ROOM Only one alabaster lamp lights the big, cold room. On the bed lies Norma in her evening dress. She is white as a sheet. Her wrists are bandaged. Her eyes are wide open, staring at the ceiling. One of her shoes has halt slipped off her foot. The other is on. Gillis opens the door and stands there tor a second. Then he slowly moves to the toot of the bed. He takes the shoes from her feet and puts them on the floor. NORMA Go away. GILLIS What kind of a silly thing was that to do? NORMA To fall in love with you -- that was the idiotic thing. GILLIS It sure would have made attractive headlines: Great Star Kills Her- self for Unknown Writer. NORMA Great stars have great pride. She puts one bandaged forearm over her eyes, sobbing. Gillis walks slowly over to the mantelpiece, stands there for awhile. NORMA Go away. Go to that girl of yours. GILLIS Look, I was making that up because I thought the whole thing was a mistake. I didn't want to hurt you. You've been good to me. You're the only person in this stinking town that has been good to me. NORMA Why don't you just say thank you and go, go, go -- GILLIS Not until you promise to act like a sensible human being. NORMA I'll do it again, I'll do it again, I'll do it again! Gillis stands looking at her helplessly. C-31 LIVING ROOM, THE DESMOND HOUSE The candles burned down, the orchestra playing to the emptiness. The orchestra leader looks at his watch, rises, silences the orchestra, then starts them in on Auld Lang Syne. C-32 INT. NORMA'S ROOM Gillis still stands. Norma lies on the bed, arms over her eyes, sobbing. GILLIS Happy New Year. Norma continues to sob. Gillis goes to the bed, puts his arms on her shoulders and turns her around. GILLIS Happy New Year. Norma looks at him, tears in her eyes. Slowly she enfolds him in her bandaged arms. NORMA Happy New Year. darling. She kisses him. DISSOLVE END OF SEQUENCE "C" SEQUENCE "D" DISSOLVE IN ON: D-1 INT. HALLWAY, NORMA GILLIS' VOICE DESMOND'S HOUSE (DAY) Around the middle of May some incidents happened The telephone is heard which I think I should tell ringing. Max comes from you about. living room to the phone, picks it up. MAX Hello ... Yes? D-1a BETTY SCHAEFER, AT THE PHONE ON HER DESK IN THE READERS' DEPARTMENT BETTY Is this Crestview 5-1733? ... I'm sorry to bother you again, but I've confirmed the number. I must speak to Mr. Gillis. D-1b MAX, AT THE PHONE MAX He is not here. D-1c BETTY ON THE PHONE BETTY Where can I reach him? Maybe somebody else in the house could tell me. D-1d MAX ON THE PHONE MAX Nobody here can give you any information. You will please not call again. He hangs up. From off comes: NORMA'S VOICE Who was it, Max? What is it? D-1e PATIO, NORMA'S HOUSE It is a sunny day. The garden is in somewhat better shape. The old house looks less unkept. The pool is filled. Norma sits on a wicker chaise longue, her face shielded by an enormous straw hat, her eyes by dark glasses. Gillis, in bathing trunks, is on a rubber mattress in the pool. Max comes to the entrance door. MAX Nothing, Madame. Somebody Inqu- iring about a stray dog. We must have a number very similar to the pound. He starts to turn back. NORMA Wait a minute. I want you to get out the car. You're going to take the script over to Paramount and deliver it to Mr. De Mille in person. MAX Yes, Madame. He goes into the house. GILLIS (climbing out of the water) You're really going to send it to De Mille? NORMA This is the right day. She indicates a typewritten letter she is holding. NORMA (Cont'd) The chart from my astrologer. She read deMille's horoscope. She read mine. GILLIS Did she read the script? NORMA DeMille is Leo. I'm Scorpio. Mars has been transmitting Jupiter for weeks. Today is the day of greatest conjuction. Now turn around. Let me dry you. She puts the towel around his sholders and starts drying him. GILLIS I hope you realize, Norma, that scripts don't sell on astrologers' charts. NORMA I'm not just selling the script. I'm selling me. DeMille always said I was his greatest star. GILLIS When did he say it, Norma? NORMA So he said it quite a few years ago. So what? I never looked better in my life. Do you know why? Because I've never been as happy in my life. She kisses him. DISSOLVE TO: D-2 INT. THE ISOTTA, DRIVING DOWN SUNSET ABOUT 8:30 IN THE EVENING GILLIS' VOICE A few evenings later we Max is driving. In the were going to the house of tonneau sit Norma, in a one of the waxworks for chinchilla wrap, and some bridge. She'd taught Gillis in his tuxedo. me how to play bridge by Norma is rummaging then, just as she'd taught through her evening me some fancy tango steps, bag. She finds a and what wine to drink cigarette case, opens with what fish. it. It is empty. NORMA That idiot. He forgot to fill my cigarette case. GILLIS (Proffering his case) Have one of mine. NORMA They're awful. They make me cough. GILLIS (Pushing open the glass partition, to Max) Pull up at the drugstore, will you, Max. (To Norma) I'll get you some. NORMA You're a darling. She takes a dollar bill from her purse and gives it to him. D-3 EXT. SCHWAB'S DRUGSTORE The car drives up and Gillis hurries into the store. D-4 INT. SCHWAB'S DRUGSTORE Business is still rather lively. There are about a dozen shoppers, and the soda counter is half filled. Gillis enters and steps to the tobacco counter. GILLIS (To the salesgirl) Give me a pack of those Turkish cigarettes -- Melachrinos. The girl opens the glass showcase to locate the fancy brand. From OFF comes ARTIE'S VOICE Stick 'em up, Gillis, or I'll let you have it! Gillis turns. D-5 AT THE SODA FOUNTAIN Artie Green and Betty Schaefer sit having a sandwich and a milk shake. With his forefinger and a sound effect, Artie riddles Gillis' body. Gillis walks INTO THE SHOT. GILLIS Hello, Artie. Good evening, Miss Schaefer. BETTY (Excitedly) You don't know how glad I am to see youl ARTIE Walking out on the mob. What's the big idea? GILLIS I'm sorry about New Year's. Would you believe me if I said I had to be with a sick friend? ARTIE Someone in the formal set, no doubt, with a ten-carat kidney stone. BETTY Stop it, Artie, will you? (To Gillis) Where have you been keeping your- self? I've got the most wonderful news for you. GILLIS I haven't been keeping myself at all. Not lately. BETTY I called your agent. I called the Screen Writers Guild. Finally your old apartment gave me some Crestview number. There was always somebody with an accent growling at me. You were not there. You were not to be spoken to. They never heard of you. GILLIS Is that so? What's the wonderful news? BETTY Sheldrake likes that angle about the teacher. GILLIS What teacher? BETTY Dark Windows. I got him all hopped up about it. GILLIS You did? BETTY He thinks it could be made into something. GILLIS Into what? A lampshade? BETTY Into something for Barbara Stan- wyck. They have a commitment with Barbara Stanwyck. ARTIE Unless you'd rather have Sarah Bernhardt. BETTY This is on the level. Sheldrake really went for it. GILLIS O.K. Where's the cash? BETTY Where's the story? I bluffed it out with a few notions of my own. It's really just a springboard. It needs work. GILLIS I was afraid of that. BETTY I've got twenty pages of notes. I've got a pretty good character for the man. ARTIE Could you write in plenty of back- ground action, so they'll need an extra assistant director? BETTY Shut up, Artie. (To Gillis) Now if we could sit down for two weeks and get a story. GILLIS Sorry, Miss Schaefer, but I've given up writing on spec. BETTY I tell you this is half sold. GILLIS As a matter of fact. I've given up writing altogether. Max has appeared in the door. MAX Mr. Gillis, if you please. GILLIS Right with you. Max leaves. ARTIE The accent! I've got it: this guy is in the pay of a foreign government. Get those studs. Get those cuff-links. GILLIS I've got to run along. Thanks any- way for your interest in my career. BETTY It's not your career -- it's mine. I kind of hoped to get in on this deal. I don't want to be a reader all my life. I want to write. GILLIS Sorry if I crossed you up. BETTY You sure have. GILLIS So long. He leaves. ARTIE (Patting her hand) Babe, it's like that producer says: In life, you've got to take the bitter with the sour. D-6 THE ISOTTA, PARKED OUTSIDE Gillis comes from Schwab's, gets into the car. Max takes off. NORMA What on earth, darling? It took you hours. GILLIS I ran into some people I knew. NORMA Where are my cigarettes? GILLIS Where are your...? He realizes he's forgotten them, takes the dollar and hands it back to her. GILLIS Norma, you're smoking too much. DISSOLVE TO: D-7 LIVING ROOM, NORMA DESMOND'S HOUSE (EARLY AFTERNOON) Start on a tiny GILLIS' VOICE parasol being Whenever she suspected I twirled...Norma was getting bored, she peeks out from one would put on a live show side of the parasol, for me: the Norma Desmond a bandanna tied Follies. Her first number around her head with was always the Mack Sennett a rabbit's-ear bow. Bathing Beauty. She bats her eyes, winks roguishly. THE CAMERA PULLS BACK to reveal that Norma's black pyjama trousers are rolled up over her knees and her black stockings rolled down below them. The whole effect approximates a Mack Sennett bathing costume pretty effectively. She points at a leather pour. NORMA This is a rock. She climbs on it, pantomimes timidity, an attempted dive, then jumps off. Gillis lolls on a couch, watching the performance, very bored. NORMA I can still see myself in the line: Bebe Daniels, Marie Prevost, Mabel Normand ... Mabel was always stepping on my feet ...What's the matter with you, darling? Why are you so glum? GILLIS (Lighting a cigarette with a match) Nothing is the matter. I'm having a great time. Show me some more. NORMA (Taking the match) All right. Give me this. I need it for a moustache. Now close your eyes. She runs out of the GILLIS' VOICE picture. Gillis has Something was the matter, closed his eyes. all right. I was thinking THE CAMERA MOVES to about that girl of Artie's, his face. that Miss Schaefer. She was so like all us writers when we first hit Holly- wood -- itching with am- bition, panting to get your names up there: Screenplay by. Original Story by. Hmph! Audiences don't know somebody sits down and writes a picture. They think the actors make it up as they go along. NORMA'S VOICE Open your eyes. Gillis opens his eyes. Norma has equipped herselr with a derby hat, a cane, and blacked in a small moustache. She goes into a little Chaplin routine. While she is doing it, the telephone rings. After a moment Max comes to the living room door. MAX Madame is wanted on the telephone. NORMA You know better than to interrupt me. MAX Paramount is calling. NORMA Who? MAX Paramount studios. NORMA (To Gillis) Now, now do you belive me? I told you deMille would jump at it. MAX It is not Mr. deMille in person. It is someone by the name or Gordon Cole. He says it's very important. NORMA Certainly it's important. It's important enough for Mr. deMille to call me personally. The idea of having an assistant call me! MAX I myself was surprised at Mr. de Mille's manners. NORMA Say that I'm busy, and hang up. MAX Very good, Madam. He bows and exits. NORMA How do you like that? We've made twelve pictures together. His greatest successes. GILLIS Maybe deMille is shooting. NORMA I know that trick! He wants to belittle me. He's trying to get my price down. I've waited twenty years for this call. Now Mr. deMille can wait till I'm good and ready. DISSOLVE TO: D-8 NORMA, IN THE TONNEAU OF THE LIMOUSINE, DRIVING DOWN MELROSE She is in full makeup, GILLIS' VOICE with a veil, a daring About three days later she hat, a suit so stunning was good and ready. In- only she would venture credible as it may seem, to wear it. THE CAMERA there had been some more PULLS BACK. Beside her of those calls from sits Gillis in the glen Paramount. So she put on plaid suit. Max is about half a pound of driving. makeup, fixed it up with a veil, and set forth to see deMille in person. Norma is examining her face in the mirror of her vanity. Max, while driving, sees her in the rear view mirror. MAX If you will pardon me, Madame. The shadow over the left eye is not quite balanced. NORMA Thank you, Max. With a handkerchief, she corrects it. D-9 MAIN GATE, EXT. PARAMOUNT STUDIO The car drives down Bronson and stops smack in front of the iron gate. A young policeman is talking to an extra; an old policeman sits reading a newspaper. Max sounds the horn impatiently. YOUNG POLICEMAN Hold that noise! MAX To see Mr. de Mille. Open the gate. YOUNG POLICEMAN Mr. deMille is shooting. You got an appointment? MAX No appointment is necessary. I am bringing Norma Desmond. YOUNG POLICEMAN Norma Who? Norma has rolled down the window on her side. She calls to the old policeman. NORMA Jonesy! Come here, Jonesy! OLD POLICEMAN Yeah? (He comes forward slowly) Why, if it isn't Miss Desmond! How have you been, Miss Desmond? NORMA Fine, Jonesy. Now open that gate. OLD POLICEMAN Sure, Miss Desmond. (To the young policeman} Come on, Mac. YOUNG POLICEMAN They can't drive on the lot without a pass. OLD POLICEMAN Miss Desmond can. Come on. They fling open the gate. OLD POLICEMAN (As the car drives through) Stage eighteen, Miss Desmond. NORMA Thank you, Jonesy. And teach your friend some manners. Tell him without me he wouldn't have any job, because without me there wouldn't be any Paramount Studio. (To Max) Go on. They drive through the gates. The old policeman goes to wall phone beside the gate, dials a number. OLD POLICEMAN (Into phone) Norma Desmond coming in to see Mr. deMille. D-10 STAGE 18 A scene from SAMPSON AND DELILAH is being rehearsed in the background. The usual turbulent activity surrounds it: extras. makeup men, grips, assistants, etc., etc. In the dim foreground a stage hand is answering a stand telephone. He puts down the phone and moves (CAMERA WITH HIM) to a second assistant. STAGE HAND Norma Desmond is coming to see Mr. deMille. The second assistant walks (CAMERA WITH HIM) to the first assistant. 2nd ASSISTANT Norma Desmond coming in to see Mr. deMille. The first assistant (CAMERA WITH HIM) hurries to the set. Sitting with his back toward us is C.B. himself. He is rehearsing a scene with Hedy Lamarr. 1ST ASSISTANT Norma Desmond is coming in to see you, Mr. deMille. C. B. turns his head. DEMILLE Norma Desmond? lst ASSISTANT She must be a million years old. DEMILLE I hate to think where that puts me. I could be her father. 1ST ASSISTANT I'm terribly sorry, Mr. de Mille. By this time de Mille is on his feet. DEMILLE It must be about that appalling script of hers. What can I say to her? What can I say? 1ST ASSISTANT I can tell her you're all tied up in the projection room. I can give her the brush ... DEMILLE Listen, thirty million fans have given her the brush. Isn't that enough? 1ST ASSISTANT I didn't mean to -- DEMILLE Of course you didn't. You didn't know Norma Desmond as a plucky little girl of seventeen, with more courage and wit and heart than ever came together in one youngster. 1ST ASSISTANT I hear she was a terror to work with. DEMILLE She got to be. A dozen press agents working overtime can do terrible things to the human spirit. (to the set) Hold everything. He leaves, accompanied by his entourage. D-11 EXT. STAGE 18 Norma's limousine drives up. Max dismounts and opens the door. NORMA (taking Gillis's hand) Don't you want to come along, darling? GILLIS I don't think so. It's your script. It's your show. Good luck. NORMA Thank you, darling. She presses his hand against her cheek, descends from the car and walks toward - D-12 THE DOOR OF STAGE 18 The first assistant is holding it open. In the door- way stands Mr. deMille. Seeing Norma, he stretches out his arms. DE MILLE Hello, young fellow. NORMA Hello, Mr. deMille. She has reached him. They embrace. NORMA Last time I saw you was someplace very gay. I remember waving to you. I was dancing on a table. DE MILLE Lots of people were. Lindbergh had just landed in Paris. Come on in. He leads her into D-13 STAGE 18 During the ensuing dialogue, Mr. deMille walks Norma towards the set. DE MILLE Norma, I want to apologize for not calling you. NORMA You'd better. I'm very angry. DE MILLE I'm pretty busy, as you can see... NORMA That's no excuse. You read the script, didn't you? DE MILLE Yes, I did. NORMA Then you could have picked up the phone yourself instead of leaving it to one of your assistants. DE MILLE What assistant? NORMA Don't play innocent. Somebody named Gordon Cole. DE MILLE Gordon Cole? NORMA And if you hadn't been pretty darned interested in that script, he wouldn't have tried to get me on the phone ten times. DE MILLE Gordon Cole... Look, Norma, I'm in the middle of a rehearsal. (Indicating his own chair) Make yourself comfortable. He walks onto the set, accompanied by his assistants. DE MILLE (Sotto voce, to his first assistant) Get me Gordon Cole on the phone. Meanwhile, Norma starts to sit, sees the name MISS LAMARR on the chair and with a look of distaste changes and sits on the one marked C.B. DE MILLE. From somewhere comes A VOICE Hey, Miss Desmond! Miss Desmond! She looks around her. VOICE Up here! Norma looks up at the scaffolding. On the scaffold stands one of the electricians, next to his light. ELECTRICIAN It's met It's Hog-eyel Norma waves at him. NORMA Hello. Hog-eye points his light at her. HOG-EYE Let's get a look at you. The beam of the lamp moves toward Norma. It hits her. She sits bathed in light. A couple of old costume extras recognize her. EXTRAS Say, it's Norma! Norma Desmond! They rush over and start wringing her hand. Into the shot comes a middle-aged hairdresser. HAIRDRESSER Hello, Miss Desmond. It's Bessie. Some elderly electricians and stagehands move in. D-14 ANOTHER PART OF THE STAGE The first assistant brings the portable phone to deMille. DeMille lifts the receiver. DE MILLE Hello. D-15 GORDON COLE'S OFFICE IN THE PROPERTY DEPARTMENT, GORDON COLE ON THE PHONE. COLE Prop Department. Gordon Cole speaking. D-16 DE MILLE ON THE PHONE DE MILLE Cole, this is C. B. deMille. Have you been calling Norma Desmond?... What's it about? D-17 GORDON COLE, ON THE PHONE COLE It's that car of hers -- an old Isotta-Fraschini. Her chauffeur drove it on the lot the other day. It looks just right for the Crosby picture. We want to rent it for a couple of weeks. D-18 DE MILLE ON THE PHONE DE MILLE (Troubled) Oh. Well, thank you. He hangs up, walks back towards Norma. (CAMERA WITH HIM). Norma stills sits in the shaft of light, surrounded by about a dozen people who have come up to pay court. DeMille gestures up to Hog-eye and the light shifts away. The people about Norma disperse slowly with various ad-libs. DE MILLE Well, Norma ... (He sits down next to her) I got hold of Gordon Cole. Norma hasn't heard a word. NORMA Did you see them? Did you see how they came? DE MILLE You know, crazy things happen in this business. I hope you haven't lost your sense of humor ... Suddenly he realizes that she is crying. She takes the handkerchief from his pocket and puts it over her eyes. DEMILLE What's the matter, Norma? NORMA Nothing. I just didn't realize what it would be like to come back to the old studio. I had no idea how I'd missed it. DEMILLE We've missed you too, dear. NORMA We'll be working again, won't we, Chief? We'll make our greatest picture. DEMILLE That's what I want to talk to you about. NORMA It's a good script, isn't it? DEMILLE It's got a lot of good things. Of course, it would be an expensive picture... NORMA I don't care about the money. I just want to work again. You don't know what it means to know that you want me. DEMILLE Nothing would thrill me more -- if it were possible. NORMA But remember, darling -- I don't work before ten in the morning, and never after 4:30 in the afternoon. The first assistant comes up. 1ST ASSISTANT We're ready with the shot, Mr. deMille. DEMILLE You'll pardon me, Norma? Why don't you just sit and watch? (He steps onto the set) O.K. Here we go. 1ST ASSISTANT Roll 'em. DEMILLE Action! The scene starts. D-19 THE ISOTTA, PARKED OUTSIDE STAGE 18 Max stands talking to Gillis, who is seated in the car. MAX (Pointing to the row of offices in the building opposite) You see those offices there, Mr. Gillis? They used to be her dressing room, The whole row. GILLIS That didn't leave much for Wallace Reid. MAX He had a great big bungalow on wheels. I had the upstairs. See where it says 'Readers' Department'? I remember my walls were covered with black patent leather... The words "Readers' Department" have registered on Gillis' mind. He gets out of the car. GILLIS I'll be with you in a minute. He crosses the street towards the green staircase leading to the second floor. Meanwhile, two prop men walking down the street come into the SHOT. 1ST PROP MAN Hey, that's the comic car Cole was talking about! (To Max) Do you mind if we look inside? MAX Go away. Go away. D-2O CUBICLE IN THE READERS' DEPARTMENT Behind the desk sits Betty, typing the synopsis of a novel, a half-eaten apple marking her place. The door behind her opens and Gillis enters. GILLIS Just so you don't think I'm a complete swine -- if there's anything in Dark Windows you can use, take it. It's all yours. BETTY Well, for heaven's sake! She moves the book and the apple aside and points at the free space on the desk. BETTY Have a chair. Gillis sits on the desk. GILLIS I mean it. It's no good to me anyway. Help yourself. BETTY Why should you do that? GILLIS If you get a hundred thousand for it, you buy me a box of chocolate creams. If you get an Oscar, I get the left foot. BETTY You know, I'd take you up on that in a minute. I'm just not good enough to do it all by myself. GILLIS What about all those ideas you had? BETTY See if they make sense. To begin with, I think you should throw out all that psychological stuff -- exploring a killer's sick mind. GILLIS Psychopaths sell like hotcakes. BETTY This story is about teachers -- their threadbare lives, their struggles. Here are people doing the most important job in the world, and they have to wprry about getting enough money to re-sole their shoes. To me it can be as exciting as any chase, any gunplay. GILLIS Check. BETTY Now I see her teaching day classes while he teaches night school. The first time they meet ... From below comes the SOUND of the Isotta's horn. GILLIS Look, if you don't mind, I haven't got time to listen to the whole plot ... BETTY I'll make it short. GILLIS Sorry. It's your baby now. BETTY I'm not good enough to write it alone. We'll have to do it together. GILLIS I'm all tied up. I can't. BETTY Couldn't we work in the evenings? Six o'clock in the morning? This next month I'm completely at your disposal. Artie is out of town. GILLIS What has Artie to do with it. BETTY We're engaged. GILLIS Good for you. You've got yourself the best guy in town. BETTY I think so. They're on location in Arizona, shooting a Western. I'm free every evening, every week- end. If you want, we could work at your place. GILLIS It's just impossible. BETTY Nobody can be that busy. There is another honk: from down below. GILLIS Look, Betty, It can't be done. It's out. BETTY You're tough, all right. GILLIS You're on your own. Stop being chicken-hearted and write that story. BETTY Honest to goodness, I hate you. GILLIS (Turning 1n the open door) And don't make it too dreary. How about this for a situation: she teaches daytimes. He teaches at night. Right? They don't even know each other, but they share the same room. It's cheaper that way. As a matter of fact, they sleep in the same bed -- in shifts, of oourse. BETTY Are you kidding? Because I think it's good. GILLIS So do I. BETTY Came on back. Let me show you where it fits in. She reaches in a drawer for her notes on Dark Windows. GILLIS (At the door) So long. Betty picks up the apple and is about to throw it after him. BETTY Oh, you -- GILLIS And here's a title: AN APPLE FOR THE TEACHER. He ducks out quiokly, slamming the door behind him. Betty looks after him, then angrlly hurls the apple into the wastebasket. D-21 STAIRCASE OUTSIDE READERS' DEPARTMENT Max is rush1ng up the stairs toward the descending Gillis. GILLIS What's the matter, Max? MAX I just found out why all those tele- phone calls. It is not Miss Desmond they want. It is the car they want to rent. GILLIS What? Max has seen something off. MAX Ssh... With his head he indicates D-22 ENTRANCE TO STAGE 18 The first assistant has opened the door. DeMille is showing Norma out. DE MILLE Goodbye, young fellow. We'll see what we can do. NORMA (embracing him) I'm not worried. Everything will be fine. The old team together. Nothing can stop us. She turns and walks out of the shot. De Mille stands for a second watching her, then turns to his assistant. DE MILLE Get Gordon Cole. Tell him to forget about her car. He can find another old car. I'll buy him five old cars, if necessary. 1ST ASSISTANT Yes, Mr. De Mille. They turn back into Stage 18. D-23 THE ISOTTA Gillis seated in the rear. Max is helping Norma in and putting the robe over her. GILLIS (Apprehensively) How did it go? NORMA It couldn't have gone better. It's practically set. Of course, he has to finish this picture first, but mine will be his next. There is an exchange of looks between Max and Gillis. GILLIS He must be quite a guy. NORMA He'a a shrewd old fox. He can smell box office. Only I'm going to outfox him a litt1e. This isn't going to be C. B. deMille's Salome. It's going to be Norma Desmond's Salome, a Norma Desmond Production, starring Norma Desmond...Home, Max. MAX Yes, Miss Desmond. As he says the words, he and Gillis exchange a glance in the rear view mirror. SLOW DISSOLVE: END OF SEQUENCE "D" SEQUENCE "E" DISSOLVE IN ON: E-1 CLOSEUP OF NORMA'S FACE GILLIS' VOICE Absolutely no makeup. A After that, an army of hand with a strong small beauty experts invaded flashlight comes into the her house on Sunset picture. The beam of the Boulevard. She went flashlight travels over the through a merciless face, exploring it merci- series of treatments, lessly. While the light is massages, sweat cabinets, still on it, two pairs of mud baths, ice compres- creamed hands come into the ses, electric devices. shot and start to massage it. She lived on vegetable juices and went to bed DISSOLVE TO: at nine. She was deter- mined to be ready -- ready for those cameras E-2 A SHORT MONTAGE of various that would never turn. beauty treatments applied to Norma. DISSOLVE TO: E-3 NORMA BEFORE THE MIRROR IN HER BEDROOM It is nine o'clock in the evening. She is in night gown and negligee and has put triangular patches on the saddle of her nose and at the outer corner of each eye. She is rubbing lotion on her hands. She gets up and crosses to the door of Gillis' room and opens it a crack. NORMA Joe darling, are you there? E-4 GILLIS' ROOM It is dark except for a lamp over the chaise longue. Gillis lies on it, fully clothed, reading a book. GILLIS Yes, Norma. Through the slit in the door there is a suggestion of Norma. NORMA Don't turn around. Keep your eyes on the book. GILLIS Yes, Norma. Norma pushes the door open and comes in. NORMA I just came to say good night. I don't want you to see me -- I'm not very attractive. GILLIS Good night. NORMA I've lost half a pound since Tuesday. GILLIS Good. NORMA I was a little worried about the line of my throat. This woman has done wonders with it. GILLIS Good. NORMA You'd better get to bed yourself. GILLIS I think I'll read a little. NORMA You went out last night, didn't you, Joe? GILLIS Why do you say that? NORMA I just happen to know it. I had a nightmare and I screamed for you. You weren't here. Where were you? GILLIS I went for a walk. NORMA No you didn't. You took the car. GILLIS All right, I drove to the beach. Norma, you don't want me to feel I'm locked up in this house? NORMA Of course not, Joe. It's just that I don't want to be left alone. Not now, while I'm under this terrible strain. My nerves are being torn apart. All I ask is for you to be a little patient and a little kind. GILLIS I haven't done anything, Norma. NORMA Of course you haven't. I wouldn't let you. She bends and kisses the top of his head. NORMA Good night, my darling. She goes into her room, shutting the door behind her. Gillis puts his book down and looks at her door. E-5 THE DOOR TO NORMA'S ROOM The light can be seen through the gouged-out keyhole. It goes out. DISSOLVE TO: E-6 UPPER LANDING STAIRWAY AND HALL BELOW (NIGHT) GILLIS' VOICE Gillis, with his coat on by Yes, I was playing hooky now, comes cautiously to the upper railing and looks every evening along in down into the lighted hall below. there. It made me think I Max is just extinguishing of when I was twelve and the lights. Max exits in, the direction of the liv- used to sneak out on the ing room. folks to see a gangster After a moment Gillis starts silently down the stairs. picture. This time it wasn't to see a picture, E-7 LIVING ROOM it was to try and write (Lighted only by the last flicker of a fire on the one. That story of mine hearth). Max is putting a fire screen in front of Betty Schaerer had dug the fire. He hears some steps and the creak or the up kept going through main door being opened. He looks out and sees my head like a dozen locomotives... E-7a THE MAIN DOOR Gillis, in the moonlit porch, is closing the main door behind him. E-8 LIVING ROOM Max looks after Gillis, his face enigmatic as ever. DISSOLVE TO: E-9 GARAGE AND DRIVEWAY (MOONLIGHT) Gillis comes into the shot, gets into the Isotta, drives it out or the garage and down the driveway to Sunset, as quietly as possible. DISSOLVE TO: E-10 READERS' OFFICE BUILDING PARAMOUNT (NIGHT) Start on a LONG SHOT. THE GILLIS' VOICE BOOM MOVES FORWARD to the only So we'd started two lights. They are the door working on it, the and window of Betty Schaefer's two of us. Nights, cubicle. Betty sits at the when the studio was desk, typing. Gillis, his deserted, up in her coat off, his shirt-sleeves little cubby-hole rolled up, j.s pacing the floor, of an office. discussing the construction of a sentence. The discussion at a stalemate, Betty suggests some coffee. Gillis agrees. From the electric plate on the shelf beside her, Betty takes a glass coffee machine. Gillis seats himself in her chair and starts typing. Betty opens the door and comes out on the balcony to fill the coffee machine from the water cooler stand- ing beside the door. BETTY I got the funniest letter from Artie. It's rained every day since they got to Arizona. They re-wrote the whole picture for rain and shot half of it. Now the sun is out. Nobody knows when they'll get back. She moves back into the room. GILLIS Good. BETTY What's good about it? I miss him something fierce. GILLIS I mean this is good dialogue along in here. It'll play. BETTY It will? GILLIS Sure. Especially with lots of music underneath, drowning it out. BETTY Don't you sometimes hate yourself? GILLIS Constantly. No, in all serious- ness, it's really good. It's fun writing again. I'm happy here, honest I am. He resumes typing. Betty puts the water on. She picks up a pack of cigarettes on the desk, finds it's empty and throws it away, sees Gillis' open gold cigarette case and lighter on the table by the couch. Betty reaches for a cigarette. The inscription en- graved inside the case catches her eye. It reads: MAD ABOUT THE BOY -- Norma BETTY Who's Norma? GILLIS Who's who? BETTY I'm sorry. I don't usually read private cigarette cases. GILLIS Oh, that. It's from a friend of mine. A middle-aged lady, very foolish and very generous. BETTY I'll say. This is solid gold. GILLIS I gave her some advice on an idiotic script. BETTY It's that old familiar story, you help a timid little soul across a crowded street. She turns out to be a multimillionaire and leaves you all her money. GILLIS That's the trouble with you readers. You know all the plots. Now suppose you proof-read page ten while the water boils. DISSILVE TO: E-11 AN EMPTY STREET AT THE GILLIS' VOICE PARAMOUNT STUDIO (NIGHT) Sometimes when we got stuck we'd make a Gillis and Betty are walking litte tour of the down it. From a stage where drowsing lot, not talk- they are erecting a new set ing much, just wandering comes a great shaft of light. down alleys between the They stop at an apple-vending sound stages, or through machine in the foreground,buy the sets they were get- themselves a couple of apples ting ready for the next and walk on. day's shooting. As a matter of fact, it was DISSOLVE TO: on one of those walks when she first told me about her nose ... E-12 PARAMOUNT'S NEW YORK STREET (NIGHT) Betty and Gillis are walking down it, THE CAMERA AHEAD OF THEM. BETTY Look at this street. All card- board, all hollow, all phoney. All done with mirrors. I like it better than any street in the world. Maybe because I used to play here when I was a kid. GILLIS What were you -- a child actress? BETTY I was born just two blocks from this studio. Right on Lemon Grove Avenue. Father was head elec- trician here till he died. Mother still works in Wardrobe. GILLIS Second generation, huh? BETTY Third. Grandma did stunt work for Pearl White. I come from a picture family. Naturally they took it for granted I was to become a great star. So I had ten years of dramatic lessons, diction, dancing. Then the studio made a test. Well, they didn't like my nose -- it slanted this way a little. I went to a doctor and had it fixed. They made more tests, and they were crazy about my nose -- only they didn't like my acting. GILLIS (Examining her nose by the flame of his lighter) Nice job. BETTY Should be. It cost three hundred dollars. GILLIS Saddest thing I ever heard. BETTY Not at all. It taught me a little sense. I got me a job in the mail room, worked up to the Stenographic. Now I'm a reader... GILLIS Come clean, Betty. At night you weep for those lost closeups, those gala openings... BETTY Not once. What's wrong with being on the other side of the cameras? It's really more fun. GILLIS Three cheers for Betty Schaefer! I will now kiss that nose of yours. BETTY If you please. Gillis kisses her nose. As he stands there, his face close to hers - GILLIS May I say you smell real special. BETTY It must be my new shampoo. GILLIS That's no shampoo. It'smore like a pile of freehly laundred hand- kerchiefs, like a brand new auto- mobile. How old are you anyway? BETTY Twenty-two. GILLIS That's it -- there's nothing like being twenty-two. Now may I suggest that if we're ever to finish this story you keep at least two feet away from me ... Now back to the typewriter. They start walking in the direction of the office. DISSOLVE TO: E-13 THE GARAGE Gillis gets out. From the seat next him he takes a batch of script, folds it and puts it in his pocket. He suddenly becomes aware that he is watched, turns. Max stands in the moonlight, evidently waiting for him. GILLIS What is it, Max? Want to wash the car, or are you doing a little spying in your off hours? MAX You must be very careful as you cross the patio. Madame may be watching. GILLIS How about my going up the kitchen stairs and undressing in the dark. Will that do it? MAX I'm not inquiring where Mr. Gillis goes every night... GILLIS Why don't you? I'm writing a script and I'm dying to finish it, no matter what. MAX It's just that I'm very worried about Madame. GILLIS Sure you are. And we're not help- ing her any, feeding her lies and more lies. Getting herself ready for a pioture ... What happens when she finds out? MAX She never will. That is my job. It has been for a long time. You must understand I discovered her when she was eighteen. I made her a star. I cannot let her be destroyed. GILLIS You made her a star? MAX I directed all her early pictures. There were three young directors who showed promise in those days: D.W. Grirrith, C.B. deMille, and Max von Mayerling. GILLIS And she's turned you into a servant. MAX It was I who asked to come back, humiliating as it may seem. I could have gone on witn my career, only I found everything unendur- able arter she divorced me. You see, I was her rirst husband. DISSOLVE TO: E-14 NORMA DESMOND'S BEDROOM One lamp lit. Norma, in a white negligee, with the patches on her face, is pacing up and down -- a small, tormented, pitiable woman. Finally she opens the door to: E-15 GILLIS' ROOM (MOONLIGHT) Gillis lies in bed asleep, Norma in the doorway. NORMA You're here, Joe ... When did you come home? Where were you? Is it a woman? I know it's a woman ... Who is she? Oh Joe, why can't I ask you? I must know, I must! Her eyes fall on Gillis' coat, which hangs over a chair. In a pocket is part of the script. Norma takes it out, looks at it. She can't see it in the moonlight. She hurries with it into: E-16 NORMA'S BEDROOM Carrying the script Norma goes to the lamp and looks at it. On the first page she sees something which confirms all her suspicionso It reads: UNTITLED LOVE STORY by Joseph C. Gilliss and Betty Schaefer DISSOLVE: E-17 BETTY'S CUBICLE (NIGHT) Betty is typing. Gillis sits on the couch, proof- reading a scene. Betty stops typing and Gillis becomes aware of her eyes fixed on him. GILLIS Hey, what's the matter... Betty, wake up! (He whistles and catches her attention) Why are you staring at me like that? BETTY Was I? I'm sorry. GILLIS What's wrong with you tonight? What is it, Betty? BETTY Something came up. I don't want to talk about it. GILLIS Why not? BETTY I just don't. GILLIS What is it you've heard. Come on, let's have it. Betty gets up. GILLIS Is it about me? Betty doesn't answer, walks out on E-18 THE BALCONY She leans against a post, crying. Gillis comes out after her. GILLIS Betty, there's no use running out on it. Let's face it, what- ever it is. BETTY It's nothing. I got a telegram from Artie. GILLIS From Artie. What's wrong? BETTY He wants me to come on to Arizona. He says it only oosts two dollars to get married there. It would kind of save us a honeymoon. GILLIS Why don't you? We can finish the script by Thursday. Betty stands crying silently. GILLIS Stop crying. You're getting married. That's what you've always wanted. BETTY I don't want it now. GILLIS Why not? Don't you love Artie? BETTY Of course I love him. I always will. I'm just not in love with him any more. GILLIS What happened? BETTY You did. There is a moment's pause before he takes her in his arms. THE CAMERA MOVES AWAY. DISSOLVE TO: E-19 HALL AND STAIRCASE GILLIS' VOICE DESMOND HOME- (NIGHT) It wasn' t until I got back to that peculiar Gillis enters, closes prison of mine that I the door as quietly as started facing the facts. he can, and goes up There it was -- Betty the stairs. Schaefer's future right in the palm of my hand. E-20 GILLIS' ROOM Betty Schaefer engaged to Artie Green, as nice He enters and turns on the a guy as ever lived. light. He sinks down on And she was in love with the chaise longue,thinking. me. Me ! She was a fool His eyes wander to the not to sense that there door of Norma's room. was something phony in Through the gouged-out key- my set-up. And I was a hole he sees the light. heel not to have told her. But you just can't say those things to somebody you're crazy about. Maybe I'd never have to. Maybe I could get away with it, get away from Norma. Maybe I could wipe the whole nasty mess right out of my life... From Norma's room comes the sound of a telephone being dialled. Gillis enters the shot and stands listening. NORMA'S VOICE Is this Gladstone 0858? E-21 NORMA'S BEDROOM Norma lies in bed, dialing a number. She has the beauty patches at the corners of her eyes and over her nose. NORMA Can I speak to Miss Betty Schaefer? She must be home by now. E-22 A BEDROOM IN BETTY'S FLAT Connie, a girl of Betty's age with whom she shares the flat, is on the phone. Betty, in a dressing- gown, comes from the bathroom, toothbrush in hand. CONNIE (Hand over mouthpiece) Betty, here's that weird-sounding woman again. BETTY What is this anyway? (Taking the phone) This is Betty Schaefer. E-23 NORMA AT IHE PHONE NORMA Miss Schaefer, you must forgive me for calling you so late, but I really feel it's my duty. It's about Mr. Gillis. You do know Mr. Gillis? ...Exactly how much do you know about him? Do you know where he lives? Do you know how he lives? Do you know what he lives on? E-24 BETTY AT THE PHONE BETTY Who are you? What do you want? What business is it of yours anyway? E-25 NORMA ON THE PHONE NORMA Miss Schaefer, I'm trying to do you a favor. I'm trying to spare you a great deal of misery. Of course you may be too young to even suspect there are men of his sort... NORMA (Cont'd) I don't know what he's told you, but he does not live with relatives, nor with friends, in the usual sense of the word. Ask him ... Ask him again. During the latter part of her call, the doors from Gillis' room have been pushed open and Gillis has walked towards her. Suddenly Norma senses his pre- sence and turns around. The telephone freezes in her hand. She tries to hang it up. Very calmly Gillis takes the receiver from her hand. GILLIS (Into phone) That's right, Betty, ask me again. This is Joe. E-26 BETTY ON THE PHONE BETTY Joe, where are you? What's this all about? E-27 GILLIS ON THE PHONE Norma beside him. GILLIS Or maybe it would be a better idea if you came over and saw it for yourself. The address is 10086 . He hangs up. Norma looks up at him as he crosses to the other end of the room and stands staring at her. The silence becomes unbearable. NORMA Don't hate me, Joe. I did it because I need you. I need you as I never needed you. Look at me. Look at my hands, look at my face, look under my eyes. How can I go back to work if I'm wasting away under this torment? You don't know what I've been through these last weeks. I got myself a revolver. You don't believe me, but I did, I did! I stood in front of that mirror, only I couldn't make myself. It wouldn't be NORMA (Cont'd) fair to all those people who are waiting to see me back on the screen. I can't disappoint them. Only, if I'm to work, I need sleep, I need quiet, I need you! Don't just stand there hating me! Shout at me, strike me! But don't hate me, Joe. Don't you hear me, Joe? GILLIS Yes, I hear you. And I wish you'd keep still so I can hear the doorbell when she rings it. E-28 BETTY AND CONNIE, DRIVING IN A SMALL COUPE DOWN (NIGHT) E-29 INT. COUPE Connie is looking at the house numbers. CONNIE Here's ten thousand seventy-nine, Betty. It must be over there. Betty turns the car into the driveway of Norma's place, stops at the entrance steps. Betty gets out. CONNIE Betty, let me come along with you. Please. BETTY No, I'll be all right. She shuts the door of the car and goes up the steps. E-30 NORMA'S BEDROOM Norma lies on the bed. Gillis sits in a far corner of the room, motionless. NORMA (In a whimpering monotone) I love you, Joe. I love you, Joe. I love you, Joe. I love you, Joe. There is the sound of footsteps below and the ringing of a doorbell. Gillis rises. NORMA What are you going to do, Joe? Without a word, he leaves the room. Norma raises herself on the bed, reaching for a black negligee lying at the foot of it. As she does so, she dis- lodges her pillow a little, revealing a revolver hidden beneath it. E-31 DOWNSTAIRS HALL, THE DESMOND HOUSE (DARK) Max crosses the hall, putting on his alpaca jacket. He turns on the lights. Outside stands Betty. From the staircase comes - GILLIS' VOICE It's all right, Max. I'll take it. MAX Yes, sir. He stands back as Gillis opens the door. GILLIS Hello, Betty. BETTY (On the threshold) I don't know why I'm so scared, Joe. Is it something awful? GILLIS Come on in, Betty, Betty enters. As he leads her into the living room, Gillis puts his arm around her shoulders. GILLIS Ever been in one of these old Hollywood palazzos? That's from when they were making eighteen thou- sand a week, and no taxes. Careful of these tiles, they're slippery. Valentino used to dance here. BETTY This is where you live? GILLIS You bet. BETTY Whose house is it? They have reached E-32 THE LIVING ROOM Gillis leads Betty in. GILLIS Hers. BETTY Whose? GILLIS Just look around. There's a lot of her spread about. If you don't remember the face, you must have heard the name of Norma Desmond. BETTY That was Norma Desmond on the phone? GILLIS Want something to drink? There's always champagne on ice, and plenty of caviar. BETTY Why did she call me? GILLIS Jealous. Ever see so much junk? She had the ceiling brought from Portugal. Look at this. He pulls the rope, showing the projection screen under the picture. GILLIS Her own movie theatre. BETTY I didn't come here to see a house. What about Norma Desmond? GILLIS I'm trying to tell you. This is an enormous place. Eight master bedrooms. A sunken tub in every bathroom. There's a bowling alley in the cellar. It's lonely here, so she got herself a companion. A very simple set-up: An older woman who is well-to-do. A younger man who is not doing too well ... Can you figure it out yourself? BETTY No. GILLIS All right. I'll give you a few more clues. BETTY No, no! I haven't heard any of this. I never got those telephone calls. I've never been in this house ... Get your things together. Let's get out of here. GILLIS All my things? All the eighteen suits, all the custom-made shoes and the eighteen dozen shirts, and the cuff-links and the platinum key- chains, and the cigarette cases? BETTY Come on, Joe. GILLIS Come on where? Back to a one-room apartment that I can't pay for? Back to a story that may sell and very possibly will not? BETTY If you love me, Joe. GILLIS Look, sweetie -- be practical. l've got a good thing here. A long-term contract with no options. I like it that way. Maybe it's not very admirable. Well, you and Artie can be admirable. BETTY Joe, I can't look at you any more. GILLIS Nobody asked you to. Betty turns from him, to hide the fact that she is crying. GILLIS All right, baby. This way out. He leads her in the direction of the door. E-33 UPPER LANDING, DESMOND HOUSE Sitting crouched behind the balustrade is Norma, peering down into E-34 THE LOWER HALL Betty and Gillis have reached the entrance door. Gillis opens it. GILLIS Good luck to you, Betty. You can finish that story on the way to Arizona. When you and Artie get back, if the two of you ever feel like a swim, here's the pool ... He switches on the light. E-35 THE PATIO The lights go on in the pool, which shines brilliant- ly in the dark garden. E-36 BETTY She doesn't even look. Her eyes filled with tears, she runs down the entrance porch toward her car. E-37 THE ENTRANCE HALL Gillis looks after her, closes the door. From the upper landing comes the sound of soft sobbing. He looks up. E-38 NORMA, ON THE UPPER LANDING Gillis ascends the stairs. NORMA Thank you, Joe -- thank you, Joe. She tries to take his hand to kiss it as he passes. He doesn't stop. Norma catches his coat. Gillis moves right on into his room. Norma lies on the floor looking after him. She crawls toward a con- sole, pulls herself up by it, starts towards Gillis' door, passes a mirror, realizes how she looks, moves back to the mirror and takes the patches off her face and does a hasty job of removing the cream with her handkerchief, readjusts her expression to a poor travesty of a smile and goes to the door of Gillis' room. NORMA May I come in? I've stopped cry- ing. I'm all right again. Joe, tell me you're not cross -- tell me everything is just as it was, Joe. She opens the door. E-39 GILLIS' ROOM In the foreground, open on the bed, is a half-packed suitcase, Gillis just putting some of his old shirts in. Norma stands staring, speechless, for a second. Gillis moves out of the shot towards the closets. NORMA What are you doing, Joe? What are you doing? You're not leaving me? GILLIS Yes, I am, Norma. NORMA No, you're not. (Calling) Max! Max! GILLIS Max is a good idea. He can help with my luggage. (He gestures in the direction of the closet) Thanks for letting me wear the handsome wardrobe. And thanks for the use of all the trinkets. He takes the cigarette case and throws it on the chaise longue. Then he throws the lighter, the wrist watch, the platinum key-chain and the tie clip. GILLIS (Indicating the bureau) The rest of the jewelry is in the top drawer. NORMA It's yours, Joe. I gave it to you. GILLIS And I'd take it in a second, Norma -- only it's a little too dressy for sitting behind the copy desk in Dayton, Ohio. NORMA These are nothing. You can have anything you want if you'll only stay. What is it you want -- money? GILLIS Norma, you'd be throwing it away. I don't qualify for the job, not any more. NORMA You can't do this! Max! Max! ... I can't face life without you, and I'm not afraid to die, you know. GILLIS That's between you and yourself, Norma. NORMA You think I made that up about the gun... She rushes into her room. Gillis closes the suitcase calmly, notices that he is still wearing some cuff- links Norma gave him, takes them off. Norma reappears in the door, carrying the revolver. NORMA See, you didn't believe me!.. Now I suppose you don't think I have the courage! GILLIS Oh. sure -- if it would make a good scene. NORMA You don't care. do you? But hundreds of thousands of people will carel GILLIS Wake up, Norma. You'd be killing yourself to an empty house. The audience left twenty years ago. Now face it. During the preceding. Max has entered. He stands listening, paralyzed. NORMA That's a lie! They still want me! GILLIS No, they don't. NORMA What about the studio? What about De Mille? GILLIS He was trying to spare your feelings. The studio wanted to rent your car. NORMA Wanted what? GILLIS De Mille didn't have the heart to tell you. None of us has had the heart. NORMA That's a lie! They want me, they want me! I get letters every day! GILLIS You tell her, Max. Come on, do her that favor. Tell her there isn't going to be any picture -- there aren't any fan letters, except the ones you write yourself. NORMA That isn't true! Max? MAX Madame is the greatest star of them all... I will take Mr. Gillis' bags. He leaves. NORMA You heard him. I'm a star! GILLIS Norma, grow up. You're a woman of fifty. There's nothing tragic about being fifty - not unless you try to be twenty-five. NORMA I'm the greatest star of them all. GILLIS Goodbye. Norma. NORMA No one leaves a star. That makes one a star. Gillis picks up the typewriter and leaves. NORMA You're not leaving me! E-40 STAIRCASE Gillis descending with the typewriter. NORMA'S VOICE Joe! ...Joe! There is the SOUND OF A SHOT. The glass of the front door is shattered. Gillis at the door opens it and walks out, without looking back. Down the staircase rushes Norma. a disordered wild- ness in the way she moves. NORMA You're not leaving me! She hurries after Gillis. E-41 PATIO (NIGHT) Dark except for lights from the house and the luminousness of the lit pool. Gillis is crossing the patio towards the garage. He is carrying the typewriter. He doesn't accelerate his step, although he has heard the shot. Behind him Norma comes from the lighted house. NORMA You're not leaving me! She shoots twice in rapid succession. Gillis drops the typewriter. The shots have swung him around. He is now facing Norma. She shoots him. This shot hits him in the belly. He doubles up, instinctively backs away from her, plummets into the lit pool. Up the stone steps from the garage rushes Max. He sees the situation, hurries towards Norma, who stands exultant in the strange light from the pool. NORMA Stars are ageless, aren't they? DISSOLVE TO: E-42 THE PATIO Dawn is breaking. At the edge of the pool stand policemen, detectives and police photographers. Motorcycle policemen are holding off the mob which is trying to storm the house. A lietuenant from the Homicide Bureau leaves the crowd around the pool and goes into E-43 THE LOWER HALL, DESMOND HOUSE It is filled with a pandemonium of police officers, newspaper people, etc. who are kept from the upper floor by two policemen at the head of the stairs. The lieutenant from the Homicide Bureau goes through the crowd to the telephone at the foot of the stairs, picks up the phone and dials. LIEUTENANT Coroner's office? ... I want to speak to the Coroner ... Who's on this phone? E-44 THE WHITE TELEPHONE IN NORMA'S BEDROOM Standing talking into it is Hedda Hopper. MISS HOPPER I am! Now get off, this is more important ... Times City Desk? Hedda Hopper speaking. I'm talking from the bedroom of Norma Desmond. Don't bother with a rewrite man, take this direct. Ready? -- As day breaks over the murder house, Norma Desmond, famed star of yesteryear, is in a state of complete mental shock ... THE CAMERA PANS TO ANOTHER PART OF THE BEDROOM, where Norma sits at a mirror, staring at herself blankly. Firing questions at her are the Captain of the Holmby Hills Division and the L.A. Homicide Squad. Max stands by faithfully. HOLMBY HILLS CAPTAIN You do not deny having killed this man, Miss Desmond? HEAD OF HOMICIDE Did you intend to kill him? Just answer me that. HOLMBY HILLS CAPTAIN Was it a sudden quarrel? Had there been any trouble between you before? HEAD OF HOMICIDE If it was a quarrel, how come you had the gun right there? HOLMBY HILLS CAPTAIN This guy -- where did you meet him for the first time? Where did he come from? Who is he? HEAD OF HOMICIDE Did he have a wife? Did he had a girl friend? Did you know them? HOLMBY HILLS CAPTAIN Had he been trying to blackmail you? E-45 PATIO - (DAWN) GILLIS' VOICE The body of Gillis Well, this is where you came. being fished from Here's that pool again,the one the pool, put on a I always wanted. They must have stretcher, covered photographed me a hundred times. with an army blanket.Then they got a couple of prun- Two men from the ing hooks from the garden and Coroner's office fished me out ever so gently. carry it towards Funny how gentle people get with the Coroner's you once you're dead. They hearse, CAMERA beached me, like a harpooned PANNING with them. baby whale, and started to check the damage, just for the record ... By this time the whole joint was jumping -- cops,reporters, neighbors, passersby -- as much hoopdedoo as we get in Los Angeles when they open a Super Market. Even the newsreel guys came roaring in. Here was an item everybody could have some fun with, the heartless so-and- so's. What would they do to her? Even if she got away with it in court- crime of passion - tempo- rary insanity - those headlines would kill her: Forgotten Star a Slayer--Aging Actress-- Yesterday's Glamour Queen... E-46 NORMA'S BEDROOM The interrogators are still firing questions at Norma who sits lifeless, staring at herself. Max watches. HEAD OF HOMICIDE Did the deceased ever threaten you? Were you in fear of bodily injury? HOLMBY HILLS CAPTAIN Did you hate him? Had you ever thought of doing something like this before? HEAD OF HOMICIDE Was theft involved? Did you catch him trying to steal something, or find he had stolen something? A police lieutenant has entered, goes to the Head of Homicide. LIEUTENANT The newsreel guys have arrived with the cameras. HEAD OF HOMICIDE Tell them to go fly a kite. This is no time for cameras. A word has pierced the mists that surround Norma. NORMA Cameras? ...What is it, Max? MAX The cameras have arrived, Madame. NORMA They have? Thank you, Max. Tell Mr. DeMille I will be on the set at once. Max flashes a look at the Head of Homicide. HEAD OF HOMICIDE What is this? MAX Please ... HOLMBY HILLS CAPTAIN (sotto voce, to Head of Homicide) Well, it's one way to get her down stairs. HEAD OF HOMICIDE Okay. And let's have the car right outside. 7-1 NORMA You will pardon me, gentlemen. I have to get ready for my scene. She takes a comb and runs it through her hair, then starts applying some wild makeup. E-47 STAIRCASE AND LOWER HALL Max makes his way down the stairs through the crowd of newsmen to the newsreel cameras, which are being set up in the hall below. MAX Is everything set up, gentlemen? Are the lights ready? From the stairway comes a murnur. They look up. Norma has emerged from the bedroom and comes to the head of the stairs. There are golden spangles in her hair and in her hand she carries a golden scarf. The police clear a path for her to descend. Press cameras flash at her every step. Max stands at the cameras. MAX Is everything set up, gentlemen? CAMERAMAN Just about. The portable lights flare up and illuminate the staircase. MAX Are the lights ready? 2ND CAMERA MAN All set. MAX Quiet, everybody! Lights! Are you ready, Norma? NORMA (From the top of the stairs) What is the scene? Where am I? MAX This is the staircase of the palace. NORMA Oh, yes, yes. They're below, waiting for the Princess ... I'm ready. MAX All right. (To cameramen) Camera! (To Norma) Action! Norma arranges the golden GILLIS' VOICE scarf ebout her and proudy So they were grinding starts to descend the stair- after all, those cam- case. The cameras grind. eras. Life, which can Everyone watches in awe. be strangely merciful, had taken pity on Norma Desmond. The dream she had clung to so des- perately had enfolded her... At the foot of the stairs Norma stops, moved. NORMA I can't go on with the scene. I'm too happy. Do you mind, Mr. DeMille, if I say a few words? Thank you. I just want to tell you how happy I am to be back in the studio making a picture again. You don't know how much I've missed all of you. And I promise you I'll never desert you again, because after "Salome" we'll make another picture, and another and another. You see, this is my life. It always will be. There's nothing else - just us and the cameras and those wonderful people out there in the dark... All right, Mr. DeMille, I'm ready for my closeup. FADE OUT. THE END | 1 |
| STRANGERS ON A TRAIN by Raymond Chandler and Czenzi Ormonde FINAL DRAFT October 18, 1950 Converted to PDF by SCREENTALK FOR EDUCATIONAL PURPOSES ONLY www.screentalk.org FADE IN: EXT. UNION STATION, WASHINGTON, D.C. DAY LONG SHOT THE CAPITOL DOME IN THE B.G. AND THE AUTOMOBILE ENTRANCE TO THE STATION IN THE F.G. LOW CAMERA Activity of cars and taxis arriving and discharging passengers with luggage, busy redcaps, etcetera. We FOCUS on a taxi pulling up and stopping, The driver hands out modest looking luggage, including a bunch of tennis rackets in cases to a redcap. CAMERA PANS DOWN as the passenger gets out of the taxi so that we see only his shoes and the lower part of his trousers. He is wearing dark colored brogues and a conservative suit apparently. The feet move toward, the entrance to the station and out of scene. Immediately a chauffeur-driven limousine drives up and an expensive place of airplane luggage is handed out of this, and the passenger alighting from the back is seen to be wearing black and white sport shoes which, as before, are all we see of him. The sport shoes start off in the wake of the brogues. INT. STATION LOBBY CAMERA FOLLOWS the sport shoes and the brogues across the lobby into a passenger tunnel. There is the usual activity of passengers walking to and from, a loud-speaker announcing trains, etc. EXT. PASSENGER TUNNEL As the brogues and the sport shoes emerge to the train platform, CAMERA PANS them over to the steps of the train. INT. TRAIN The brogues and the sport shoes pass separately down the aisle, the sport shoes turning in at a compartment door and the brogues continuing toward the parlor car. DISSOLVE TO: INT. PARLOR CAR (PROCESS) The brogues come to rest before a chair as the owner sits down. A moment later the sport shoes come to rest. before in adjoining chair. Converted to PDF by www.screentalk.org 2. The legs belonging to the sport shoes stretch out, and one of the shoes touches one of the brogues. MAN'S VOICE (over scene) Oh, excuse Me! CAMERA PULLS BACK AND UP to SHOW two young men seated in two parlor car chairs. BRUN0 ANTHONY, the wearer of the sport shoes, is about twenty-five. He wears his expensive clothes with the tweedy nonchalance of a young man who has always had the best. The wearer of the brogues is a fine looking but, at the moment, a somewhat troubled young man. This is GUY HAINES. He, too, is in his middle twenties and is well dressed because he can now afford to be. He nods politely, acknowledging Bruno's apology, then turns away with the gesture implying he wants privacy. BRUNO (smiling with sudden recognition) I beg your pardon, but aren't you Guy Haines. Guy nods with a polite half smile. Being a well known tournament tennis player, he has had this sort of experience before. BRUNO (snapping his finger) Sure! I saw you blast Faraday right off the court in South Orange last season. What a backhand! Made the semi-finals, didn't you? Guy acknowledges this with a modest nod and turns to his magazine rolled up in is fist. BRUNO (with open admiration) I certainly admire people who do things. (smiling and introducing himself) I'm Bruno Anthony. Bruno. See Guy looks up. Bruno indicates his gold tie pin which bears his name in cut- out letters. Guy looks at it with the faintest expression of disdain. I suppose you think it's corny. But my mother gave it to me so of course I wear it to please her. Converted to PDF by www.screentalk.org 3. GUY (patiently)(a faint smile) How do you do. BRUNO (with an apologetic grin) I don't usually talk so much. Go Ahead and read. GUY (wryly) Thanks. Guy tries to read but is uneasily aware of Bruno's open appraisal. BRUNO It must be pretty exciting to be so important. GUY (fidgeting slightly) A tennis player isn't so important. BRUNO People who do things are important. I never seem to do anything. Not knowing how to answer this, Guy looks a little embarrassed. BRUNO (still insistent on being friendly) I suppose you're going to Southampton -- for the doubles. GUY (politely) You are a tennis fan. Bruno is inordinately pleased by this small tribute. BRUNO Wish I could see you play. But I've got to be back in Washington tomorrow. I live in Arlington, you know. He has taken out a cigarette case. Holds it out to Guy. Converted to PDF by www.screentalk.org 4. BRUNO Cigarette? GUY Not now, thanks. I don't smoke much. BRUNO I smoke too much. He fumbles for a match. Guy brings out a lighter and hands it to Bruno. BRUNO Thanks. (he stares at the lighter, impressed) Elegant. CLOSE SHOT OF THE LIGHTER Showing that it has the insignia of crossed rackets embossed on it, and underneath is engraved the inscription: "To G from A". BRUNO'S VOICE (reading) To G from A. Bet I can guess who A is. WIDER SHOT Guy reacts sharply. GUY (coldly) Yes? BRUNO Anne Burton. Sometimes I turn the sport page and look at the society news. And the pictures. She's very beautiful, Senator Burton's daughter. GUY You're quite a reader, Mr. Anthony. BRUNO Yes, I am. Ask me anything, from today's stock reports to Li'l Abner, and I got the answer. (MORE) Converted to PDF by www.screentalk.org 5. BRUNO (CONT'D) Even news about people I don't know. Like who'd like to marry whom when his wife gets her divorce. GUY (sharply) Perhaps you read too much. BRUNO (contritely) There I go again. Too friendly. I meet someone I' like and open my yap too wide. I'm sorry... At the appeal on Bruno's face, Guy slowly relents. GUY That's all right. Forget it. I guess I'm pretty jumpy. Bruno smiles with and signals a waiter. BRUNO There's a new cure for that. (to waiter) Scotch and plain water. A pair. Double. (to Guy with a chuckle) Only kind of doubles I play. GUY You'll have to drink both of them. BRUNO (grinning) And I can do it. (moving in) When's the wedding? GUY What? BRUNO The wedding. You and Anne Burton. (a gesture of explanation) It was in the papers. GUY It shouldn't have been. Unless they've legalized bigamy overnight. Converted to PDF by www.screentalk.org 6. BRUNO I have a theory about that. I'd like to tell you about it some time. But right now I suppose divorce Is still the simplest operation. The waiter has brought the drinks. Bruno slips the lighter into hip pocket to free his hands for the bills which he gives to the waiter, waving away the change. He offers a glass to Guy. Guy takes it. GUY (as if he needs it) I guess I will. BRUNO (happily) This is wonderful -- having your company all the way to New York. GUY (forced to explain) As a matter of fact, I'm not going direct. I'm stopping off. At Metcalf. BRUNO Metcalf? What would anybody want to go there for? GUY It's my home town. BRUNO Oh, I get it! A little talk with your wife to about the divorce! I suppose she was the girl next door. Held her hand in high school and before you knew it -- hooked! (proud of his perspicacity) Am I right? GUY (laconically) Close enough. BRUNO (raises his glass) Well, here's luck, Guy. Drink up -- then we'll have some lunch sent to my compartment. Converted to PDF by www.screentalk.org 7. GUY Thanks very much. But I think I'll go to the dining car. (he hails a waiter who is passing through with a food-laden tray) Do you know if there are any vacant seats in the dining car now? WAITER Not for about twenty minutes I'm afraid, Sir. BRUNO (pleased) See? You'll have to lunch with me. (motions the waiter back) Say, waiter, bring me some lamb chops and French fries and chocolate ice cream, Compartment D, Car 121. (turns to Guy) What'll you have, Guy? GUY Thanks just the same, but I really don't think -- BRUNO Oh, go on and order. The waiter is hovering impatiently. Guy gives in out of embarrassment. GUY Well, I'll Just have a hamburger and a cup of coffee. BRUNO (delighted, lifts his glass in another toast) To the next Mrs. Haines. Guy nods curtly. DISSOLVE TO: Converted to PDF by www.screentalk.org 8. INT. BRUNO'S COMPARTMENT ON TRAIN (PROCESS) Bruno and Guy are finishing lunch. Bruno has been drinking and his eyes are bright and feverish. An almost empty liquor bottle is near a couple of detective novels covered with gaudily Illustrated dust jackets. Bruno has in unlighted cigarette in his mouth. Guy's lighter is on the table. Bruno snaps it a couple of times, as though fascinated, lights his cigarette and puts the lighter on the table again. BRUNO Sure, I went to college. Three of them. Every time they kicked me out my father threw me back in. (bitterly) He finally gave up. He thinks I'm awfully small fry, not worth the bait. (wistfully) You my friend, Guy? GUY Sure. I'm your friend, Bruno. BRUNO (a little woozy) No, you're not, nobody thinks I'm anything special. Only my mother. (empties the bottle into his glass) My father hates me. Guy smiles this off as nonsense. GUY You must be imagining things. BRUNO (hitting the bottom of the bottle for the last drop) And I hate him. He thinks I ought to catch the eight-five bus every morning, punch a timeclock and work my way up selling paint or something. Him -- with all his money! GUY (amused by Bruno) Well, what do you want to do? BRUNO You mean before or after I kill him? Converted to PDF by www.screentalk.org 9. GUY (chuckling) Before, of course. BRUNO (leaning forward eagerly) I want to do everything. I got a theory you're supposed to do everything before you die. Have you ever driven a car, blindfolded, at a hundred and fifty miles an hour? GUY Not lately. BRUNO I did. I flew in a jet plans too. (his hand traces a swift streak through the air, and he adds sound effects) Zzzzzzzp! Man, that's a thrill! Almost blow the sawdust out of my head. I'm going to make a reservation on the first rocket to the moon... GUY (amused and curious) What are you trying prove? BRUNO I'm not like you, Guy. You're lucky. You're smart. Marrying the boss's daughter is a nice short cut to a career, isn't it? GUY (quickly) Marrying the senator's daughter has nothing to do with it. Can't a fellow look past a tennis not without being a goldbricker? BRUNO Take it easy, boy. I'm your friend, remember? I'd do anything for you. GUY (humoring Bruno) Sure, Bruno, sure. (MORE) Converted to PDF by www.screentalk.org 10. GUY (CONT'D) (glancing at his watch) We'll be pulling in soon. I've got to change trains. BRUNO What'd you say her name was -- your wife's? GUY Miriam. BRUNO That's it. Miriam Joyce Haines. Played around a lot, I suppose? GUY Let's not talk about it any more. BRUNO (almost hopefully) Maybe she'll make more trouble for you. GUY I don't think so. BRUNO You mean you got enough on her to get your divorce no matter what? GUY Let's change subject, Bruno, can't we? BRUNO Okay, Guy. Want me to tell you one of my ideas for murdering my father? GUY (indicating the detective novels) You've been reading too many of these. BRUNO (going right on) You want to hear about the busted light socket in the bathroom, or the carbon monoxide in the garage? GUY No. I may be old fashioned, but I thought murder was against the law. Converted to PDF by www.screentalk.org 11. BRUNO But not against the law of nature. My theory is that everybody is a potential murderer. Didn't you ever want to kill somebody? Say one of those useless fellows Miriam was running around with? GUY You can't go around killing people just because you think they're useless. BRUNO Oh, what's a life or two? Some people are bitter off dead, Guy. Take your -- wife and my father, for instance. It reminds me of a wonderful idea had once. I used to put myself to sleep at night -- figuring it out. Now, let's say you want to get rid of your wife. GUY Why? BRUNO Let's say she refuses to give you a divorce -- (raises a finger and stops Guy's protest) Let's say. You'd be afraid to kill her because you'd get caught. And what would trip you up? Motive. Now here's the plan... GUY I'm afraid I haven't time to listen. BRUNO (ignoring the remark) It's so simple, too. A couple of fellows meet accidentally, like you and me. No connection between them at all. Never saw each other before. Each of them has somebody he'd like to get rid of, but he can't murder the person he wants to get rid of. He'll get caught. So they swap murders. GUY Swap murders? Converted to PDF by www.screentalk.org 12. BRUNO Each fellow does the other fellow's murder. Then there is nothing to connect them. The one who had the motive isn't there. Each fellow murders a total stranger. Like you do my murder and I do yours. GUY (with relief) We're coming into my station. BRUNO For example, your wife, my father. Criss-cross. GUY (sharply) What? BRUNO (with a smile) We do talk the same language -- don't we, Guy? GUY (preparing to leave) Sure, we talk the same language. Thanks for the lunch. BRUNO (beaming) I'm glad you enjoyed it. I thought the lamb chops were a little overdone myself. He holds out his hand. Guy is in a hurry but he shakes hands. GUY Nice meeting you, Bruno. BRUNO (detaining him at the door) You think my theory is okay, Guy? You like it? GUY Sure, sure, Bruno. They're all okay. (he salutes a quick goodbye and hurries away) Converted to PDF by www.screentalk.org 13. Left alone, Bruno picks up Guy's lighter from the table, starts to call Guy back to hand It to him.Then he looks closer at the insignia of crossed tennis rackets. BRUNO (smiling) Criss-cross. DISSOLVE TO: A WIDE VIEW OF THE TOWN OF METCALF METCALF RAILROAD STATION as the train comes in. THE TRAIN STATION PLATFORM MED. SHOT As Guy gets off the with his suitcase and tennis rackets. A baggage man with baggage truck is passing. GUY Hi, Bill. BAGGAGE MAN (smiling) Guy Haines! Good to too you, boy. You be sure to win at Southampton tomorrow, hear me? I've got two dollars on your nose. GUY (indicating his suitcase and rackets) Then park these in a lucky spot for a few hours, will you? BAGGAGE MAN Sure thing. He loads them onto a truck. DISSOLVE TO: INT. METCALF STREET LONG SHOT Guy is walking up the main street. Converted to PDF by www.screentalk.org 14. EXT. MUSIC SHOP Typical music shop of a small town, with plate glass windows and displays of radios, records, sheet music, etc. Activity of a couple of customers and salespeople inside. Guy comes along the street and goes into the shop. INT. MUSIC SHOP As Guy enters. There are the usual counters and shelves, pianos and radios on display, and the sound of a piano being tuned in the back of the store. MIRIAM is finishing with a customer at a counter. MR. HARGREAVES, the manager, is busy at the shelves. Another girl clerk is serving a customer. In one of the glass cubicles where records are tried out, a customer is playing symphonic music; in a second glass cubicle another customer is listening to a record of popular music. A third cubicle is empty. Activity of the street is seen through the plate glass front. Guy walks straight to Miriam, just as she is finishing with her woman customer, handing over a small package. MIRIAM (taking money from customer) Even change. Thank you, Madam. (she looks up at Guy as the woman moves off) Well -- hello, Guy. GUY You're looking well, Miriam. Miriam's face is pretty because it is still young. She is self-centered and inclined to be vindictive. She wears harlequin glasses with myopic lenses which tend to make her eyes look small. MIRIAM So are you. You've got a nice tan, playing tennis with all your rich friends. GUY (ignoring the remark) What time do we meet your lawyer? MIRIAM (sly little smile) What's your hurry? Converted to PDF by www.screentalk.org 15. GUY My hurry? That's funny, coming from you! You're the one who's in a hurry, aren't you? MIRIAM (coyly) When you wouldn't give me the divorce right away, I sort of hoped it was because you were a little bit jealous. GUY (biting) I got over being jealous, a long time ago Miriam. Miriam's eyes slide toward the other girl clerk who has moved closer, within listening range. MIRIAM (indicating empty glass cubicle) Let's talk in there. Guy follows Miriam across to the empty room. Miriam has brought her purse along. They enter. INT. CUBICLE Once inside, the sounds of the music playing from other parts of the shop are heard but very faintly. The piano tuning still goes on, but less stridently. Miriam and Guy are cooped together in the close quarters. MIRIAM (intimately) Now this is cosier. Sort of like old times, isn't it, Guy? GUY (coldly) Oh, skip it, Miriam. It's pretty late to start flirting with a discarded husband. Especially when you're going to have another man's baby. MIRIAM Do you know, I think you're handsomer than ever? Converted to PDF by www.screentalk.org 16. GUY Let's see your lawyer and get this over with. MIRIAM Did you bring the money, Guy? Lawyers are expensive. GUY (taking money from his wallet) Here it is. MIRIAM (taking the money greedily) If I'd known what all that tennis nonsense of yours was going to lead to, I wouldn't have run out on you. GUY What are you trying to say, Miriam? Come out with it. MIRIAM (tucking the bills away) I'm not getting a divorce. GUY (tense and angry) Why, you little doublecrosser. I didn't want this divorce, you did. That's what you've been harping about for the past year. MIRIAM It's a woman's privilege to change her mind... Now I can shop for some pretty clothes. I wouldn't want you to be ashamed of me in Washington when we go to all those dinners and swanky parties. GUY And what do you mean by that? MIRIAM (Coyly) Don't look so mad, Guy. You always smile when your picture is being taken for the papers. (MORE) Converted to PDF by www.screentalk.org 17. MIRIAM (CONT'D) Especially when you have Anne Burton hanging on your arm. GUY Let's not talk about Anne Burton. MIRIAM So, it's really serious between you two? Well, you can throw your dreams about her into the ashcan. Guy, I'm coming to Washington. GUY What for? MIRIAM To have my baby and be with you. GUY Why me? It's not my baby. MIRIAM But people don't know that, Guy, do they? It would make a pretty story, wouldn't it -- the senator's daughter involved with a married man who's about to become a father. GUY (furiously) You black conniving little liar! A few people in the shop look around as Guy's voice rises above the sound of the record playing. MIRIAM Keep your voice down. GUY What happened? Did he run out on you? MIRIAM No man runs out on me. Not even you. GUY You're a liar and a cheat, Miriam. You've wanted to get rid of me long enough and now I'll go you one better -- I never want to see or hear of you again. Converted to PDF by www.screentalk.org 18. MIRIAM (demurely) I could be very pathetic as the deserted little mother in a courtroom, Guy. Think it over. Who would believe you? Guy seizes her angrily and in so doing, knocks the tone arm across the record with a loud screech. From outside we can see heads turn. Mr. Hargreaves, the manager, is very disturbed. MED. SHOT THROUGH GLASS PARTITION FROM HARGREAVES' VIEWPOINT We see Guy gripping Miriam's arms and apparently addressing her in a threatening manner, although we do not hear his words. The smile has faded from Miriam's face and something like cringing fear has taken its place. She is drawn and tense and seems to cower beneath Guy's rage. Mr. Hargreaves moves forward and opens Guy's tirade. GUY ...That's what should happen to people like you. And if I... HARGREAVES (interrupts) Break it up, folks. This isn't the place for a family quarrel. GUY (his eyes blazing) Sorry. I'm leaving. He starts to exit from the booth. Miriam grabs his arm and screams at him: MIRIAM (yelling like a fishwife) You heard what I said, Guy Haines. You can't throw me away like an old shoe. I'm coming to Washington to have my baby. Tell that to the senate! Guy strides out of the store, the manager and a few customers turning around in surprise. Converted to PDF by www.screentalk.org 19. The two customers in other booths, seeing the quarrel, open their doors simultaneously and Miriam's tirade is climaxed by a cacophony of noise, a big symphony, loud hot music, and the apparently unaware piano tuner. EXT. MAIN STREET METCALF SHOOTING TOWARDS STATION Guy is striding along angrily. He comes to the same intersection and the same cop. The officer makes a friendly gesture, is if he'd like to talk awhile, but Guy strides past him without noticing. EXT. METCALF STATION (PROCESS) Guy comes into the scene, crosses to a row of public telephone booths, enters one. Inside the telephone booth, he dumps some loose change on the shelf, sticks a nickel in the telephone, speaks into it. GUY Long distance. (a pause) I want Washington, D. C. The number is Republic 0800. Person to person. Miss Anne Burton. Another pause, very long. Guy is very restless. He digs a cigarette out of his pocket and sticks it in his mouth, then looks through his pockets for his lighter, doesn't find it. He looks puzzled, but about that time the operator speaks to him. GUY (continuing) Right. Guy picks coins up off the shelf and drops them into the telephone, then waits. He shifts the receiver and fumbles in his other jacket pocket, then turns to the phone. GUY (tautly, into phone) Anne, -- Anne darling. Yes, I'm in Metcalf -- (gets a grip on himself) No, everything didn't go smoothly. She doesn't want a divorce, not now.... Converted to PDF by www.screentalk.org 20. INT. BURTON LIVING ROOM ANNE BURTON is a beautiful, high-spirited and well-bred young woman. The smile on her face his faded to anxiety as she listens over the telephone which is on the desk. ANNE (after a pause then with unpleasant realization) Another man's child! But she can't do that to you, Guy -- it's unbelievable -- it's, it's evil! (she listens, then calmly) Yes, I know how you must feel. (pause) But you sound so savage. BACK TO GUY IN TELEPHONE BOOTH GUY (furiously) Sure I sound savage. I feel savage. I'd like to break her neck! (a pause, then raising his voice) I said I'd like to break her foul, poisonous, useless little neck! (the connection is bad and he strains to hear) What's that? Meantime the noise of a through train has been HEARD, and the horn on a streamliner locomotive. It has come up very fast, it is now almost to the station. Guy rises his voice and yells into the telephone. His voice fights the roar of the train: GUY I SAID I COULD STRANGLE HER! The expression on his face is frenzied and suggesting that he means exactly what he is saying. DISSOLVE TO: Converted to PDF by www.screentalk.org 21. INT. ANTHONY LIVING ROOM The scene opens on a CLOSEUP OF A MAN'S HANDS. One of them is semi-flexed and turning slowly, The other is receiving the final touches of a manicure. CAMERA PULLS BACK to reveal that these are Bruno's hands, and that, he is studying them moodily, CAMERA PULLS BACK FARTHER to reveal his mother, MRS. ANTHONY, sitting opposite him at a little table in the Anthony living room. She is working with scissors, file and nail buffer. Mrs. Anthony is a gentle, once pretty woman, whose pastel exterior harbors a tigress-like determination to protect her son, Bruno is in his robe and is unshaven. There is evidence of long established wealth in the heavy dark appointments of this room. MRS. ANTHONY Since you insisted on a manicure, dear, I do wish you'd keep your hands quiet. You're so restless lately. BRUNO (almost dreamily as he admires the free hand) I like them to look just right. Mrs. Anthony looks up, notices his moody expression. MRS. ANTHONY Did I file them too short? BRUNO No, Ma. They look fine. Thanks. MRS. ANTHONY Then what's the matter? BRUNO I'm all right, Ma. Don't worry about me. MRS. ANTHONY You look so Pale, dear. Are you out of vitamins? BRUNO I bought a bottle of them yesterday. A whole fifth. Converted to PDF by www.screentalk.org 22. MRS. ANTHONY (anxiously) But you have that 'look'. I can always tell. You haven't got into any more mischief, Bruno? He denies this with a slow, solemn shake of his head. MRS. ANTHONY I do hope you've forgotten about that silly little plan of yours? BRUNO (sharply) Which one? MRS. ANTHONY (smiling) About blowing up the White House? BRUNO (his eyes dancing) I was only kidding, Ma. Besides, what would the president say? MRS. ANTHONY (laughing gaily) You're a naughty boy, Bruno. But you can always make me laugh. (she rises) Now get shaved, dear, before your father gets home. Bruno's fist crashes down on the little table, upsetting it, as he gets to his feet. BRUNO I'm sick and tired of bowing and scraping to the king. MRS. ANTHONY (placating him) Now, now, Let's not lose control. Come see my painting, dear -- (she leads him toward an easel) I do wish you'd take up painting. It's such a soothing pastime. They look at the painting. Converted to PDF by www.screentalk.org 23. INSERT The painting is a horrible mess. Out of the violence of the pattern a man's face can be discerned, wild-eyed and distorted. We hear laughter from Bruno. BACK TO SCENE Bruno's roar of laughter puzzles Mrs. Anthony, but she is pleased to hear his good humor. He puts an arm around her. BRUNO You're wonderful, Ma! It's the old boy, all right. That's father! MRS. ANTHONY (bewildered) It is? I was trying to paint Saint Francis. At this moment there is the sound of the front door opening. Then immediately the telephone bell rings in the hall. Bruno is instantly alert, as if he had been expecting a call. He goes toward the door to the hall, as the butler enters. BUTLER (to Bruno) They are ready with your call to Southampton, Sir. Bruno's father MR. ANTHONY, purposefully enters the living room. He an impeccably dressed business man with an uncompromising eye. His entrance momentarily blocks Bruno's exit. MRS. ANTHONY (to her husband) How nice that you're early, Charles. I'll tell cook.... Bruno now exits into the hall, passing his father without speaking. MR. ANTHONY Just a minute, Eunice. (calls after Bruno) Bruno! Come here! I want to talk to you and your mother. Converted to PDF by www.screentalk.org 24. INT. HALL CLOSE SHOT BRUNO as he approaches the telephone. BRUNO (calls back to his father) Sorry father. Long distance. (he picks up the telephone) Hello... CAMERA MOVES IN TO A BIG HEAD CLOSEUP OF BRUNO at the telephone as the Voices of his mother and father can be heard from the other room. MR. ANTHONY'S VOICE Now it's hit and run driving! And you knew about it all the time! BRUNO (eagerly into phone) Guy? (pause) Bruno, Bruno Anthony. MR. ANTHONY'S VOICE You're going to protect him once too often. After all we do have a responsibility to society. Bruno gives a look in his father's direction, before he speaks into the telephone in a low voice. BRUNO I just wanted to ask how you made out with Miriam. INT. LOCKER ROOM OF TENNIS CLUB CLOSE SHOT GUY AT TELEPHONE GUY (puzzled) What? (listens) Metcalf? Who'd you say you were? Converted to PDF by www.screentalk.org 25. CLOSEUP BRUNO BRUNO (sotto voce) Bruno, Guy. Bruno Anthony. Don't you remember? On the train. The voices of Mr. and Mrs. Anthony can still be heard in dispute as Bruno listens at phone: MRS. ANTHONY I never permit it! Bruno gives a significant look in direction of the living room as he speaks into the phone. BRUNO (softly) Are you getting your divorce? MR. ANTHONY'S VOICE I tell you he should be sent somewhere for treatment before it's too late. BRUNO (into phone, with satisfaction) So she double-crossed you! Are you going to see her again? The phone clicks in Bruno's ear. He looks hurt for an instant, then replaces the receiver. Bruno listens to his father off scene and his expression becomes more enigmatic. MR. ANTHONY'S VOICE I tell you, Eunice, I'm going to have that boy put away if it's the last thing I do! Bruno looks off in direction of his farther's voice with an expression which says, "Crow while you can, you haven't much time." He reaches into his pocket, brings out Guy's cigarette lighter and as he flicks it on and off. DISSOLVE TO: EXT. METCALF STATION LONG SHOT DAY This is the same shot we saw when Guy arrived in Metcalf. We see the station and one of the main streets beyond the station. Converted to PDF by www.screentalk.org 26. LONG SHOT A NEARER VIEW We see the train come around the curve. Again this is just the same angle that we used for Guy. It comes to a stop in the foreground and we see Bruno alight onto the platform. He looks about him for a moment and then strolls away in the direction of the town. He approaches the row of telephone booths. EXT. STATION CLOSE SHOT We see Bruno enter the small booth and start to glance through the telephone directory. INSERT TELEPHONE DIRECTORY Bruno's finger runs down the names until it stops at: Joyce, Miriam Haines. 2420 Metcalf Avenue. A RESIDENTIAL STREET IN METCALF LONG SHOT It is now much later. It is beginning to get dark, and the street lights are on. In the far distance we see a local bus approaching. MED. SHOT SHOOTING DOWN onto a small seat by a bus stop, we see Bruno with an open newspaper in front of him. It is held up as he reads it. CLOSEUP Bruno is glancing over the top of the paper. LONG SHOT From his viewpoint we see a typical frame house. The upper windows are lit as are the lower ones as well. A woman is sitting in a rocker on the front porch. This is MRS. JOYCE, Miriam's mother. She has white hair. A woman comes along the street and pauses as she gets to Mrs. Joyce. Converted to PDF by www.screentalk.org 27. WOMAN (calls out as she passes) Hello Mrs. Joyce. Warm, ain't it? MRS. JOYCE That it is. WOMAN I've been reading where your son-in- law's been coming right along at tennis. MRS. JOYCE (sourly) We don't have any interest in tennis any more. The neighbor passes on. CLOSE UP Bruno, still glancing over the top of his paper. LONG SHOT Again from Bruno's viewpoint, we see Miriam's house. At this moment the front door swings open, emitting a long streak ot bright light. We see the silhouette of a woman emerge, followed by two other men. They're laughing and joking. Suddenly they look up the street. At this very moment the bus pulls up in front of Bruno's view, cutting off the sight of his quarry. The bus comes to a stop. CLOSE SHOT Bruno rises in alarm and moves around toward the end of the bus so that he shall not lose sight of the girl coming out of the house. SEMI-LONG SHOT From his viewpoint, the girl, whom we now see is Miriam, is running followed by the two young men. They are calling for the bus not to go - shouting, "Hi - stop!" Mrs. Joyce calls from the porch: MRS. JOYCE Don't you stay out too late, Miriam. Converted to PDF by www.screentalk.org 28. MIRIAM (calling back) Goodnight, Mother. See you later. CLOSE UP Bruno watches Miriam. MED. SHOT Miriam comes nearer and nearer to Bruno. With her two companions she brushes past him and jumps onto the bus. THE CAMERA PANS BRUNO AFTER THEM. EXT. AMUSEMENT PARK LONG SHOT We see the bus pull up outside the Amusement Park, and the various passengers alight. These include Miriam nd her companions, and Bruno. LONG SHOT NEARER VIEW OF THE AMUSEMENT PARK We see the usual midway with its various concessions on each side: in the distance the Ferris wheel, Merry-go-rounds, etc., and beyond that a lake. In the foreground we see people filling in and out. DISSOLVE TO: MED. LONG SHOT A GROUP BY A FROZEN CUSTARD STAND This group comprises Miriam and her two boy-friends. They lick their way out of the crowd and debate between themselves where to go next. CLOSE SHOT Miriam's eye catches the attention of something off screen. SEMI-LONG SHOT From her viewpoint we see Bruno standing and casually watching her. Other people pass around and in front of him, so that he is the only immobile figure. Converted to PDF by www.screentalk.org 29. SEMI-CLOSEUP Miriam, with a kind of coy consciousness, turns away with the others and they go on to some other concession. MED. SHOT As Bruno starts to advance in the direction of Miriam he is momentarily held up by a small boy in cowboy uniform carrying a gun and a balloon. The small boy points the gun at Bruno. SEMI-CLOSE UP The small boy pointing the gun fires it twice with a couple of 'bangs!' He then starts to move off. SEMI-CLOSE UP Bruno moves on past the boy. He casually touches the balloon with his cigarette end -- it goes off with a 'pop'. CLOSE UP The small boy turns and looks with dismay at his pricked balloon, wondering what happened. SEMI-CLOSE UP Bruno moves on, pleased with himself, returning his attention to Miriam who is somewhere ahead of him. MEDIUM SHOT Miriam and her two boy-friends by the sledge hammer concession where the aim is to swing the hammer hard enough down onto its target to ring the bell and register the 100 mark. Miriam is in the foreground of the shot. The first boy steps up to try his hand. As he swings, Miriam turns and glances about her, obviously looking for Bruno. LONG SHOT FROM MIRIAM VIEWPOINT The crowds milling, but no sign of Bruno. Converted to PDF by www.screentalk.org 30. MEDIUM SHOT The first boy having failed to ring the bell, the second stops up and slams the hammer down. CLOSE SHOT The register shooting up only to the hallway mark. CLOSE SHOT MIRIAM She looks a little disdainful and again glances around for Bruno. Looking first to her left where she sees nothing, she then looks to her right, and as she does THE CAMERA PANS to show Bruno standing right it her shoulder. Miriam gives a little start. Bruno smiles at her. With a smirk he walks over and after paying his fee, goes to take up the hammer. CLOSE UP MIRIAM She watches Bruno. CLOSE SHOT Bruno looks down at his hands. INSERT Bruno's two strong hands - as he holds them palms tilted upward and fingers curled in. CLOSE UP Bruno, as he smiles faintly, glancing across at Miriam. CLOSE UP MIRIAM She gives a faint smile in return. CLOSE SHOT With a studied movement, Bruno picks up the handle of the hammer and swings. Converted to PDF by www.screentalk.org 31. CLOSE SHOT The register shoots up to the 100 mark and rings the bell. MEDIUM SHOT Bruno drops the hammer and glances around at Miriam again. Her two boy-friends are calling for her from a little distance. BOY'S VOICE Come On, Miriam. Come On! CLOSE SHOT MIRIAM She turns away and is lost in the crowd. MEDIUM SHOT OVER BRUNO'S SHOULDER AT MERRY-GO ROUND IN BACKGROUND Bruno turns to follow Miriam, his manner casual. As he takes a few steps, WE PAN ACROSS with him until, over his shoulder, we see a merry-go-round in the background. Miriam and the two boys are aboard and climbing onto horses. As Bruno goes toward the merry-go-round, the CAMERA MOVES UP A LITTLE with him. The merry-go-round starts to move slowly round as Bruno hops on. MEDIUM SHOT ON MERRY-GO-ROUND Bruno begins to look around for Miriam, who is apparently on the other side of the merry-go-round. He starts to thread his way through the horses which are beginning to move up and down. CAMERA FOLLOWING HIM. He passes one or two of the oncoming heads before he reaches Miriam. She is on an outside mount which is high in the air when she sees Bruno facing her. Her laughter dies for a moment and she smiles at him coyly. Bruno passes her and gets on the horse directly behind her, Miriam glancing at him as her horse comes down. MEDIUM SHOT BRUNO ON HORSE With horse's head in foreground, as it is coming toward us. Converted to PDF by www.screentalk.org 32. SIDE VIEW MIRIAM Miriam on her horse, moving from left to right. Miriam, holding the reins, glances back with a gay laugh. SIDE VIEW BRUNO Bruno on his horse, as though he is chasing Miriam. He is a little more open now in his laughter. GROUP SHOT MIRIAM AND TWO BOYS Miriam and her boy friends begin to sing the song being played on the calliope. CLOSE UP MIRIAM As she starts to sing, she glances back. CLOSE UP BRUNO He is starting to join in the singing. MEDIUM SHOT The horses of the merry-go-round are filling the screen as they whizz by, and again we get the picture of Bruno chasing Miriam as they rush past the CAMERA, the music and tempo at a high speed. LAP DISSOLVE TO: EXTERIOR OF BOAT LANDING ON SHORE OF ARTIFICIAL LAKE Across the water may be seen a small wooded island. Between this and the boat landing there is an artificially constructed "Tunnel of Love". We see Miriam and her companions approach the boat concession and CAMERA FOLLOWS THEM onto the little landing stage. CAMERA MOVES UP SLOWLY over the boy's shoulders until we get MIRIAM IN CLOSE UP. She glances back. Her expression changes to a coy smile of satisfaction as she sees: Converted to PDF by www.screentalk.org 33. MEDIUM SHOT (FROM MIRIAM'S VIEWPOINT) Bruno is approaching the pay box. MEDIUM SHOT Miriam and her companions are escorted to a small boat with electric motor. Once they are seated the boat chugs away from the landing stage and off into the darkness. Bruno steps into the foreground and gets into the next boat which floats alongside. He, too, moves away into the darkness. ENTRANCE TO THE TUNNEL As Miriam's boat passes through, she gives another little glance over shoulder before her boat disappears into the darkness of the tunnel. After a brief moment Bruno's boat comes into the picture, and it, too, goes into the tunnel. INSIDE THE TUNNEL We see the silhouettes of the occupants of Miriam's boat on the wall of the tunnel, lit dimly from the light coming from the tunnel exit. The silhouette of Bruno in his boat, lit by the tunnel entrance, gradually approaches the other three. When the silhouettes are almost touching, we -- CUT TO: EXIT OF THE TUNNEL It is empty. There is a sudden piercing scream from inside, followed after a second or two by protestations and giggling as Miriam's boat emerges into the light. She is pushing one of the boys away from her. MIRIAM (squealing) George, stop it, I tell you! Their boat moves out of the picture, toward the island. Presently Bruno's boat comes smilingly following and he, too, moves on out of the picture. Converted to PDF by www.screentalk.org 34. MEDIUM SHOT ISLAND The group of Miriam and her companions are scrambling out of their boat and moving onto the island, one of the boys trying the boat on the shore. They disappear into the Woods of the island. Again Bruno's boat comes into the picture. He steps out, lift the prow of the boat a little onto the shore. LONG SHOT ISLAND We see the amusement park lighted beyond the lake. Silhouetted in the foreground, the trees and foliage of the island. Nearby we see the silhouetted figures of Miriam and her companions move across the scene, right to left. Miriam is pushing George away from her. MIRIAM (protesting perfunctorily) George, no! She backs away from him and the boys go on picture. Miriam goes in another direction, around, the bushes. George obviously misses her, for we hear his voice call out: GEORGE'S VOICE Miriam! Miriam backs out of the bushes until the back of her head is in CLOSEUP in the foreground of the shot. Suddenly she hears steps in back of her and turns her head toward CAMERA. Her face changes as she recognizes someone offscene. MIRIAM Oh! She gives a coy smile of recognition. CAMERA PULLS BACK to reveal the mad and shoulders of Bruno between Miriam and the camera. His hand holds Guy's lighter which he flicks on as he raises it above Miriam's face. 0f Bruno, we see only the back of his head and shoulders. BRUNO Is your name Miriam? MIRIAM (with surprise) Why yes. How did you -- Converted to PDF by www.screentalk.org 35. We see Bruno's gloved hands dart quickly to Miriam's throat. The lighter falls down out of picture, and as Bruno's hands grip her throat, his head moves slightly to blot out Miriam's face. His head moves a bit farther until Miriam's face is nearly uncovered at the other side of the screen, and we see her glasses fall off. CLOSE SHOT Miriam's glasses hit the ground. The shadows of their struggling figures over the shot. CLOSE UP The screen is filled with one of the lenses of the glasses. They are of the diminishing type. Against the moonlit sky we see reflected, the elongated struggling figures, as though we were shooting up at them. Suddenly one of the figures falls forward. CLOSE UP Miriam's head drops into the picture by the glasses. Bruno's hand comes into the picture and picks up the glasses. One of the lenses has been broken by Miriam's fall. As we see Bruno's sport shoes move away, the CAMERA MOVES PAST MIRIAM'S HEAD until it comes to Guy's lighter pressed into the earth. CLOSE UP BRUNO Bruno glances back over his shoulder. He looks down and goes back one or two steps. CLOSE UP BRUNO'S HAND Bruno's hands retrieve the lighter from the ground. LONG SHOT ISLAND We see a full view of the island again, with the amusement park beyond. The faint noise of the calliope continues in the distance. Bruno has been lost to view. Converted to PDF by www.screentalk.org 36. Miriam's companions are still searching for her. We hear their faint voices in the distance. VOICES Miriam! Miriam! Where are you? MEDIUM SHOT Bruno comes to the shore where his boat is moored. He gets in and is quickly chugging away. He moves calmly, matter-of- fact and not furtively. LONG SHOT LAKE Bruno's boat throbbing its way across toward the landing stage. MEDIUM SHOT LANDING STAGE There are two boats unloading. Bruno's boat is approaching. We hear a loud call from the island. Someone has found Miriam. VOICES Hey, here she is! What's the matter with her? Has she fainted? More shouts from the island cause the people at the landing stage to look back. The boatman's attention is also attracted. Suddenly, as Bruno is getting out of boat, there is a loud scream from the island. VOICE (crying out) She is dead! OTHER VOICE (from island) Help! Help! Bruno by this time has stopped onto the landing stage, and in company with the other people, is looking back as if to see what's wrong on the island. Then he moves away, starting off of the landing stage. The boatman turns and glances at Bruno, but quickly returns his attention to the disturbance across on the island. He hurries forward and with a couple of men passengers jumps into one of the boats. He calls to his assistant as he gets into the boat. Converted to PDF by www.screentalk.org 37. BOATMAN Got a cop! The assistant runs off out of the pictures MEDIUM SHOT BRUNO As Bruno calmly threads his way along the midway, we hear above the noise of the various concessions, a shrill police whistle in the distance. Presently a couple of policemen comes running from direction of the main entrance and past Bruno. He glances at them over his shoulder, then strolls on toward the main entrance to the park. ENTRANCE TO AMUSEMENT PARK EXTERIOR As Bruno comes out through the turnstile, he stands for a moment on the street. At this moment a man hesitates at the curbstone. He is blind and tapping the sidewalk with his white cane. He takes one step into the roadway, then hesitates. Bruno steps forward and takes the blind man's arm. CAMERA PULLS BACK as Bruno escorts the blind man across the road. With a sweeping gesture he holds back a couple of cars to lot them pass. Once on the other side of the road, the blind man utters his thanks. BLIND MAN Thanks. He goes off. Bruno looks back toward the park, then glances down at his wristwatch. INSERT BRUNO'S WRISTWATCH The time is 9.30. LAP DISSOLVE TO: INT. OBSERVATION CAR OF A TRAIN NIGHT Through the rear window we see the tracks rushing away from us. Seated in the foreground are Guy Haines and a rather professorial type opposite him, a bespectacled man around forty-five or fifty who is extremely drunk. Converted to PDF by www.screentalk.org 38. MEDIUM SHOT GUY He is reading an evening newspaper. CLOSE SHOT The feet opposite Guy stretch out and touch Guy's feet. CLOSEUP GUY He lowers his paper and looks across. MED. SHOT The drunk opposite Guy looks down at his feet and then up to Guy resentfully as though Guy had kicked him. He eyes Guy up and down, then suddenly, without warning, bursts into song, to the tune of the Barber Shop Chord. COLLINS There was a man, now please take note. There was a man who had a goat. He loved that goat, Indeed he did. He loved that goat, just like a kid. (He stops singing abruptly and addresses Guy) What is your opinion? GUY (amused) You'll never make the Metropolitan. COLLINS (fuzzily -- pumping Guy's hand) Name's Collins. On sabbatical - Delaware Tech. Glad to meet you. I jus' gave a speech in New York. On integration. In the differential calculus a function is given and its differential is obtained. Understand? GUY (solemnly) Sure, I understand. Converted to PDF by www.screentalk.org 39. COLLINS (resentfully) Y'do? Again he bursts into loud song. LAP DISSOLVE TO: LONG SHOT WASHINGTON EXTERIOR ABOUT 1 A.M. MOONLIGHT A solitary taxi is seen driving past the Capitol Building. LAP DISSOLVE TO: The taxi comes to a side street and stops outside a small apartment house. MED. SHOT Guy gets out of the taxi with his rackets and bag, pays the driver and goes up the steps to the front door of his apartment. CLOSE SHOT As Guy is about to enter the front door and we see his name posted on a small card as one of the several tenants, he hears a soft call from across the street. VOICE (softly) Guy! Guy turns his head and looks across the street. MED. LONG SHOT (FROM GUY'S VIEWPOINT) We see a small space between two houses across the street. Out of the darkness the voice repeats. VOICE Over here, Guy. MED. SHOT GUY He turns, and with a slightly bewildered and wary expression, goes out of the picture to cross the street. Converted to PDF by www.screentalk.org 40. MED. SHOT Guy reaches the other side of the street and still puzzled and cautious, approaches the dark alleyway. MED. SHOT After a moment a figure steps out of the darkness. It is Bruno. He steps back into the darkness again as Guy comes up to him. TWO SHOT Guy frowning in puzzlement as he looks at Bruno. BRUNO (cheerfully) Hello, Guy. GUY (recognizes Bruno -- not pleased) What are you doing here? At this time of night? BRUNO (a little sadly) You don't seem very pleased to see me, Guy. Guy stands without answering. BRUNO (pleased again) I brought you a little present. GUY What do you mean? Bruno's hand comes out of his pocket and he hands Miriam's glasses to Guy. INSERT Guy's hands taking Miriam's glasses from Bruno. One of the lenses is broken. Converted to PDF by www.screentalk.org 41. TWO SHOT As Guy takes the glasses he looks at Bruno in bewilderment. GUY What's this all about? BRUNO Recognize them? CLOSEUP GUY He looks down at the glasses, mystified. He looks up again to Bruno. CLOSEUP BRUNO BRUNO It was very quick, Guy. She wasn't hurt in any way. It was all over in no time. CLOSEUP GUY He is horrified. He looks swiftly down at the glasses in his hand, then back to Bruno. BRUNO'S VOICE (bragging) I know you'd be surprised. Nothing for us to worry about. Nobody saw me, only Miriam. TWO SHOT Guy can hardly believe what he is hearing. BRUNO I was very careful. Even when I dropped your lighter there, I went right back to it up. If It'd been found, it would have ruined our whole scheme, wouldn't it? GUY Are you trying to tell me you've -- Why, you maniac! Converted to PDF by www.screentalk.org 42. BRUNO (looks at Guy with astonishment) But, Guy, you wanted it! We planned it on the train together, remember? Guy suddenly starts to go. Bruno grabs his arm. BRUNO Where are you going? GUY Where do you think I'm going? I'm going to call the police, of course. BRUNO But you can't, Guy. We'd both be arrested for murder. Guy turns back slowly and faces him. GUY We'd both be arrested for murder? BRUNO You're is much in it as I am. We planned it together. Criss-cross. I do your murder -- GUY (suddenly angry) You crazy fool! You think you can get away with that? BRUNO (a little hurt) Oh, come now, Guy. Why should I go to Metcalf and kill a total stranger, unless it was part of the plan and you were in on it? You're the one that benefits, Guy. You're a free man. I didn't even know the girl. Guy makes a move to leave, but Bruno holds on tight. GUY Let me go, Bruno. I had nothing to do with this and the police will believe me. Converted to PDF by www.screentalk.org 43. BRUNO (concerned) If you go to the police now, you'll just be turning yourself in as in accessory. You see, you have the motive. At this moment both turn at a sound across the street. LONG SHOT (FROM THEIR VIEWPOINT) We hear the sound of a telephone ringing in Guy's apartment. The top of one of his windows is open. BRUNO What is it? GUY My telephone. BRUNO (amused) Someone has some news for you, Guy. Guy still stares across the street. LONG SHOT (FROM HIS VIEWPOINT) We see a police car pull up outside Guy's apartment. TWO SHOT Bruno pulls Guy back further into the shadows. Guy instinctively flattens himself against the wall. He looks across the street again. LONG SHOT (FROM HIS VIEWPOINT) We see the two policemen go into his apartment building. TWO SHOT Guy is still flattened against the wall to keep out of light. BRUNO Tell them you know about it already, Guy. Converted to PDF by www.screentalk.org 44. CLOSEUP GUY He looks across at the police, then down at himself with some surprise and disgust, then over at Bruno, suddenly conscious he is behaving like a criminal and that Bruno is responsible for his predicament. GUY (muttering) You've got me acting, like a criminal, you crazy fool! Bruno for a moment looks menacingly at Guy. BRUNO Don't you call me that. Bruno's flare of anger dies. They both look again across the street. LONG SHOT (FROM THEIR VIEWPOINT) The two policemen come out of the house, get into their car and drive off. Guy's telephone is still ringing. TWO SHOT BRUNO You must be tired, Guy. I know I am. I've sure had a strenuous evening. Guy looks at him, almost numb. BRUNO Now look, Guy, about my father. I have the plans made. Two plans. A plan of the grounds and a plan of the house. I have in old Luger I bought at a pawn shop in San Francisco. My father -- Guy turns and starts to move in across the street. TWO SHOT Bruno follows Guy and we FOLLOW them across the street. CAMERA ON THEIR BACKS. Guy strides ahead to the house. Converted to PDF by www.screentalk.org 45. BRUNO Wait a minute, Guy. To have to talk. We have to arrange things. Guy turns at the door to his apartment building. GUY (furiously) Get away before I give you what you gave Miriam. BRUNO (sadly) You're not yourself, Guy. You're tired. When you think things over, you'll see I'm right. Tomorrow -- Guy opens his door, turns on Bruno. GUY (with finality) I don't know you. I never saw you before. I never want to see you again. He goes in and slams the door in Bruno's face. BRUNO (to the closed door) But we have to -- He realizes there is no use in trying to talk to Guy any further. He turns and faces the CAMERA IN CLOSE UP as he moves away, looking sad almost to the point of tears. INT. GUY'S APARTMENT Guy is standing at the telephone which is still ringing. He has Miriam's glasses in his hand. He looks down at them for a moment, then picks up the receiver. He hesitates, then speaks into the phone. GUY (hoarsely, into phone) Yes? (Pause) Yes, Anne. I'm sorry, darling. I just got in. (pause) Of course I'm all right. (MORE) Converted to PDF by www.screentalk.org 46. GUY (CONT'D) (forcing his voice to sound normal) But you sound upset. Is anything wrong? (Pause) All right. I'll come over. Right away. He hangs up but keeps his hand on the telephone, deliberating. He starts to dial, then suddenly hangs up and starts out. DISSOLVE TO: EXT. A RESIDENTIAL STREET, WASHINGTON LONG SHOT NIGHT A taxi drives up and stops in front of a handsome residence. It is the Burton home. Guy gets out of the taxi and goes up the steps. MED. SHOT OVER GUY'S SHOULDER His figure tense, he rings the bell. After a moment's wait, the door is opened from inside and Anne Burton stands in the lighted hallway. She looks at Guy with an anxious, taut expression, searches his face hastily, then as he takes a step inside she is suddenly in his arms. They embrace with wordless fervor. GUY (holding her close) Anne darling, you're trembling. Anne draws back and looks into his face as if searching for an answer to some question in her mind. ANNE Guy -- (her fingers gently touch his face) I wonder if you know how much I love you. Guy takes her hand from hIs face, caresses it with his lips. GUY (forcing a smile) Brazen woman. I'm the one to say that. Converted to PDF by www.screentalk.org 47. ANNE (tensely) But I wanted you to know, before... (forcing herself to be calm) Before we go into the living room. Father wants to see you. CLOSEUP GUY He looks apprehensively in direction of the living room, conscious of what the news is to be, but covering up. LONG SHOT LIVING ROOM FROM GUY'S VIEWPOINT SENATOR BURTON and BARBARA BURTON are seated near a desk on the farthest side of the room. Senator Burton is a distinguished fifty, a man with great pride in tradition, his family and his career. Barbara, Anne's younger sister, is a lively seventeen who loves excitement, says exactly what she thinks and rarely thinks before she says it. Superficially, in height and figure, she resembles Miriam. She also weirs glasses. By her gestures we gather she is speaking urgently, but softly, to her father, who lifts a weary hand to quiet her as she looks toward Guy in the hallway, Barbara keeps quiet and also looks toward Guy. They both wait for him to enter. CLOSEUP GUY He steels himself for the long walk across the hall and the living room. CLOSEUP ANNE Watching Guy closely. MED. SHOT As Guy starts to make the long trek across the living room, with Anne behind him -- GUY (stiffly) Good evening, sir. Hello, Babs. Converted to PDF by www.screentalk.org 48. Barbara has been squirming in her seat, then as if jet propelled she catapults out of it and runs to Guy, giving him a big hug and a smack on the cheek. BARBARA Something awful has happened, Guy. SENATOR (firmly) Sit down, Barbara. Subdued, she sits down. But Guy remains standing. SENATOR (finding it difficult to begin) There seems to be no way of diplomatically breaking tragic news. I'm sorry, Guy, to be the one to tell you. It concerns your wife. She's been murdered. Guy stares woodenly at the Senator, is if hypnotized. BARBARA The police have been using everything but radar to locate you. SENATOR You're to call Headquarters at Metcalf. The full impact of what has happened hits Guy once more. GUY Miriam...murdered. ANNE (with inner tension) She was...strangled. Slowly Guy's eyes meet hers. They are remembering what he said on the phone: "I could strangle her." He sinks into a chair. The Senator is quite distressed. During the following scene Barbara quietly goes about the business of pouring drinks and serving them. She knows everyone's preference. Converted to PDF by www.screentalk.org 49. SENATOR (wrylt, to Guy) It happened on an island in an amusement park. It was sort of a lovers lane, I believe. A rather sordid atmosphere. BARBARA (quickly, to Guy) Miriam went there with two boys. They were the ones who found her. So they're not suspects. But you probably will be. SENATOR Young lady, we can't overlook the fact that murder is at our doorsteps. But I forbid you to drag it into the living room! BARBARA (wide-eyed) Let's not fool ourselves. The police will say Guy wanted Miriam out of the way so he could marry Anne. In a crime of this sort the police first go after the husband, and Guy had every motive. SENATOR (aghast) Motive? GUY (quietly) She's right. Whichever way you look at it...I'm in a spot. SENATOR (disconcerted but whistling in dark) Oh come now, my boy. I'm sure you have nothing to worry about. BARBARA (flatly) If he hasn't an alibi for nine-thirty tonight he has plenty to worry about. Converted to PDF by www.screentalk.org 50. ANNE (who hasn't taken anxious eyes off Guy) You can tell them where you were, can't you, Guy? GUY (wearily) At nine-thirty I was on the train from New York to Washington. SENATOR (relieved) There you are. BARBARA Who saw you? Did you speak to anyone? You'll need a Witness, you know. GUY (as if it didn't matter) Yes, I spoke to someone. SENATOR (hopefully) Anyone you know? GUY No. His name was Collins. He is a professor. SENATOR (brightening) Harvard. GUY University of Virginia. The Senator's expression says: "Well, that's not too bad." CLOSEUP ANNE Her face shows her relief that Guy can account for his time. ANNE Then everything's's all right. Converted to PDF by www.screentalk.org 51. BACK TO SCENE BARBARA Not quite. Detectives play a game called Motive, Motive, Who'd got the Motive. ANNE (near the breaking point) I'm sick of hearing that word! BARBARA He'll still have to answer questions. SENATOR Routine. Pure routine. GUY I'm afraid there'll be a lot of reporters at your front door in the morning. BARBARA Daddy doesn't mind a little scandal. He's a senator. ANNE (answering Guy's look) It can't be helped, darling. It is not your fault. It's not as though anyone can say you had something to do with it. GUY Someone might say it...I'd do anything to keep you all out of this mess. SENATOR Profit by my experience, Guy. Never lose any sleep over accusations. (an afterthought) Unless they can be proved, of course. We'll help all we can. Dreadful business, dreadful. That poor unfortunate girl. BARBARA (flatly) She was a tramp. Converted to PDF by www.screentalk.org 52. SENATOR (pontificially) She was a human being. let me remind you that even the most unworthy of us has the right to life and the pursuit of happiness. BARBARA (unimpressed) From what I hear, she pursued it in all directions. SENATOR Barbara! ANNE Father, it's getting terribly late, and Guy looks so tired... SENATOR (quickly) Of course, of course. Back to bed, Barbara. BARBARA (ignoring this - to Anne and Guy) Well, you two. Nothing stands in your way now. You can be married right away. Think of it -- you're free! CLOSE TWO ANNE AND GUY look at one another with a growing realization of what Miriam's death actually means to their happiness -- they are free. BACK TO SCENE The Senator firmly urges Barbara to the door. SENATOR (to Barbara) One doesn't always have to say what one thinks! BARBARA (sweetly) Father, I'm not a politician. Converted to PDF by www.screentalk.org 53. The Senator gives her a gentle but firm push out of sight. SENATOR You won't forget that call, Guy? Captain Turley. GUY Yes sir. Goodnight. Barbara pokes her head quickly around the door. BARBARA I still think it would be wonderful to have a man love you so much he'd kill for you. (she ducks out) TWO SHOT Left alone, Guy and Anne embrace. Anne's nervous tension comes to the surface in a flood of relief. ANNE I told myself over and over I was being silly, but there was one horrible moment tonight when the news came through. I kept remembering what you shouted telephone from Metcalf. GUY That I could strang... Anne quickly puts her fingers over his mouth. ANNE Don't even say it. Forget you ever said it. Even more terrifying than the murder itself, Guy, was the awful thought that if you had anything to do with it we'd be separated, -perhaps forever. I'd never see you again. I couldn't bear it. DISSOLVE TO: LONG SHOT MAIN STREET OF METCALF DAY with its customary mid-afternoon activity. LAP DISSOLVE TO: Converted to PDF by www.screentalk.org 54. EXT. METCALF POLICE HEADQUARTERS DAY A knot of people are hanging around the entrance, including a few newspaper photographers. There is a rush of interest when a taxi pulls up and Guy steps out of it. Guy pushes his way through the people. Two or three bulbs flash. There is a murmur from the crowd and we hear Guy's name. He passes into the entrance. INT. CORRIDOR OF POLICE HEADQUARTERS Guy comes into the corridor from the street and approaches two policemen who are standing nearby. GUY Captain Turley's office? One of the policemen gestures to a door at the right. Guy crosses and enters. INT. RECEPTION ROOM OUTSIDE CAPTAIN TURLEY'S OFFICE At one side of the room is a young police sergeant seated at a typewriter. A group of people are seated in chairs lined against the opposite wall. Guy enters, crosses to the sergeant at the desk. GUY Captain Turley is expecting me. Guy Haines. SERGEANT Just a moment, Mr. Haines. He rises and goes into an adjoining room. CLOSEUP GUY He now has time to take stock of the waiting people. He catches his breath when he sees: CLOSEUP MRS. JOYCE Miriam's mother, dressed all in black, is seated in one of the chairs. She has been staring at the floor, but brings her eyes up slowly to glare at Guy with a look of burning hatred. Converted to PDF by www.screentalk.org 55. MRS. JOYCE (a fierce whisper) You'll pay for this! CLOSEUP MR. HARGREAVES Mr. Hargreaves from the music shop looks across at Guy, attempts in awkward nod but is very embarrassed. CLOSEUP GUY Guy nods in returns. MED. SHOT The two boys who were with Miriam at the amusement park. They look at Guy with interest. MED. SHOT GUY He looks about him uncomfortably, then turns suddenly as he sees: MED. SHOT Seated behind Guy, apart from the others who are waiting, is Professor Collins, Guy's drunken companion on the train of the night before. The professor is completely sober now, dignified and erect. He has removed his glasses to polish them and does not react to Guy's presence. CLOSEUP GUY Guy starts with a smile of recognition to say, "How do you do?" but at that moment he hears the door open and his name called: SERGEANT'S VOICE Will you come in, please, Mr. Haines? MED. SHOT Guy breaks away from his uncompleted greeting to the professor and goes through the door to Captain Turley's office, followed by the eyes of the waiting people. Converted to PDF by www.screentalk.org 56. INT. CAPTAIN TURLEY'S OFFICE CAPTAIN TURLEY is conscientious, methodical and always polite. He puts aside photographs and records and rises from behind his desk as Guy comes in. A detective lieutenant, CAMPBELL, is attending a coffee maker. Their expressions are grave by contrast with Guy's confident attitude after seeing the professor in the waiting room. CAPTAIN TURLEY Good of you to be so prompt, Mr. Haines. This is Lieutenant Campbell. (the two nod to each other) Won't you sit down? GUY Thank you, sir. (he sits) CAPTAIN TURLEY I know you're a busy man, so we won't detain you any longer than necessary...Now you already been good enough to tell us where you were last evening, and we've managed to locate the gentleman you spoke with on the train. Turley signals to Campbell to call the professor in. GUY (brightening) Yes. I saw him outside. CAMPBELL (at open door) Will you come in please, professor? CLOSEUP GUY He looks up eagerly. MED. SHOT Professor Collins comes in and sits in a chair opposite Guy. TURLEY Professor Collins, this is Mr. Haines. He was with you on the train last night. Converted to PDF by www.screentalk.org 57. The professor studies Guy for a moment, then awkwardly turns to Turley. COLLINS I'm terribly sorry, but I really don't remember meeting this gentleman. CLOSEUP GUY Surprised. His confident expression fades. CLOSEUP PROFESSOR COLLINS He turns from the captain to Guy. COLLINS (apologetically) Unfortunately, I remember very little about the journey from New York...You see, there had been a little celebration -- MED. SHOT GROUP Guy interrupts with a slight note of impatience. GUY But we were sitting opposite each other in the observation car! You were singing a song about a goat -- COLLINS (incredulously) A goat? GUY (urgently) And calculus. You were going over a speech you'd made. Turley and Campbell are watching closely. COLLINS I was? I'm sorry, Mr. Halnes. (shakes his head) I certainly must have celebrated! I can't remember you at all. Converted to PDF by www.screentalk.org 58. CLOSEUP GUY Momentarily Guy is frustrated, then he turns quietly to Turley. GUY (calmly, logically) Captain, is it so important whether or not Professor Collins remember me? Surely, the important thing is that I've been able to name a man who was on the train with me. You've been able to find him. Isn't that proof of where I was at nine-thirty last night? Guy asks this question with a look of near triumph that he has clearly established his alibi. DISSOLVE TO: INT. BURTON LIVING ROOM EVENING The Burtons are having coffee. Barbara has been glancing through a new murder mystery with a lurid cover. As Guy enters, Anne rises to greet him. ANNE Hello, darling. Have you had your dinner? GUY On the train. ANNE You weren't in Metcalf all this time? We expected you hours ago. BARBARA (flatly) I didn't. They sometimes throw a suspect in the can and keep him there all night. SENATOR (after a disapproving glance at Barbara) Sit down, Guy. Sit down. Give him some coffee, Anne. (MORE) Converted to PDF by www.screentalk.org 59. SENATOR (CONT'D) (back to Guy) You had no trouble with the police of course, once they verified your alibi? GUY (morosely) When an alibi is full of bourbon, sir, it can't stand up. BARBARA You mean the professor was boiled? GUY Completely. He didn't remember me. ANNE But, you knew he was on the train! Wasn't that enough to prove you were on it, too? GUY Apparently not at the right time. They suggested I could have caught the train at Baltimore after Miriam was murdered. They had it all worked out -- (taps his head) in their timetables. ANNE (growing indignant and increasingly nervous) That's ridiculous. They're acting as if you were guilty. BARBARA (somewhat subdued and trying to be comforting) Everything will be all right, Anne. The police were just being thorough -- (she's unsure of herself, and defers to the senator) Weren't they, daddy? SENATOR I certainly hope so. (to Guy) What is your next step? Converted to PDF by www.screentalk.org 60. GUY (wryly) Whatever it is, the police will know it. They gave me a present -- come take a look. He crosses to the window, lifts the curtain slightly, then turns back to the others. GUY (continuing) My guardian angel. The group move to look out the window, the senator with reluctance. LONG SHOT EXT. STREET FROM THEIR VIEWPOINT Through the window we see the figure of a man across the street. He is lighting a cigarette and strolling up and down. BACK TO GROUP BARBARA (impressed) You're being tailed! GUY (turning to them) That's Leslie Hennessy. He works sixteen hours a day. Somebody else takes over for the next eight. (drops the curtain, turns back into room) As a matter of fact, Hennessy's a very nice fellow. BARBARA Shouldn't we ask him in for Coffee -- or something? Nobody bothers to answer her. The Senator is disturbed, but confident of his own prestige as he goes back to his coffee. SENATOR I'll have him called off immediately of course. Converted to PDF by www.screentalk.org 61. GUY (calmly) I'm afraid where I go, Hennessy goes. Even to the Senate. SENATOR (Pausing with his cup hallway to his mouth) Is he likely to -- picket my office? GUY Very likely. The Senator's cup is suddenly back on its saucer and he is on his feet, pacing nervously. SENATOR I would suggest, Guy, for your own peace of mind, of course, that you work here at the house for a few days. (a pause) It would be less embarrassing for you. Guy has been looking at Anne and is concerned at the worry on her face. He nods in assent to the Senator's suggestion, but puts his hand over Anne's. GUY (hopelessly) Then what about practicing? Perhaps I'd better forget Forest Hills? SENATOR My dear boy, wouldn't it look rather -- awkward -- if you suddenly canceled all your plans. ANNE He's right, Guy. You mustn't do anything that would look suspicious. You've got to carry on as though nothing has happened. BARBARA (pointing out the window) Escorted by Mr. Hennessy. The are crestfallen again. RANDALL, the manservant, has entered with the telephone. Converted to PDF by www.screentalk.org 62. RANDALL A call for you, Mr. Haines. They say it is urgent. The phone is plugged in to a connection and Guy crosses the room and picks up the receiver. The Burtons watch him. GUY Hello -- INT. TELEPHONE BOOTH BIG HEAD CLOSEUP OF BRUNO His face wears the most affable expression. BRUNO Hello, Guy. I tried your apartment, but -- (pause) Why, Guy, this is Bruno! INT. BURTON LIVING ROOM Guy hangs up the telephone quickly. He looks at the others, awkwardly tries to explain: GUY Must be some mistake. It wasn't for me. His embarrassment grows as Anne looks at him with a puzzled expression. FADE OUT. FADE IN EXT. WASHINGTON STREET APPROACHING JEFFERSONS MEMORIAL DAY Guy and HENNESSY are walking along the street together, CAMERA MOVING WITH THEM. Their relationship is most friendly. Guy carries a briefcase. Hennessy is an amiable but not gullible young man in his early thirties. He knows his job, is well groomed, well educated, and well liked. GUY Well, I suppose I was pretty lucky to be seeded fifth, really. Converted to PDF by www.screentalk.org 63. HENNESSY I've never seen the Forest Hillss tournament before. I'm looking forward to it. GUY (wryly) Do you mean we'll be going there together, Hennessy? HENNESSY Oh, don't worry. This thing will be cleared up by that time. (changes the subject) Ever thought of turning professional, Guy? GUY I won't have to do that. When I'm through with tennis. I'll be going into politics, I hope. HENNESSY (aghast) Politics! It's a good thing for you I don't report that to the chief. He turns to light a cigarette. As he does, Guy gives a barely perceptible start at what he sees offscene. LONG SHOT JEFFERSON MEMORIAL FROM GUYS VIEWPOINT The tiny figure of a man is standing at the base of the tall white column. The figure lifts in arm and waves. Instinct tells us that this is Bruno. Hennessy is still mumbling his opinion of politics. HENNESSY'S VOICE If he knew you were getting into that rat-race -- TWO SHOT GUY AND HENNESSY Guy turns his back on Bruno's figure and looks frantically toward to street, wanting to get away. HENNESSY -- He'd put ten men on your trail. He says -- Converted to PDF by www.screentalk.org 64. GUY (interrupts) Let's take this cab. It's getting late. He hails a taxi which is cruising by, and they start to get in. Guy directs the driver. GUY Pentagon Building, please. HENNESSY Oh, no, not there! I always get lost. INT. TAXI CLOSE SHOT Guy turns and looks out of the window. LONG SHOT JEFFERSON MEMORIAL from Guy's viewpoint, shot through the cab window. Again we see the solitary figure of Bruno looking after Guy and beginning to recede with the background as the cab starts off. DISSOLVE TO: INT. GUY'S APARTMENT NIGHT As Guy comes in from outside, there is a note on the floor that has been pushed under the door. Guy picks it up, stares at it for a minute before he opens it. He takes out a handwritten note and reads it with an expression of disgust. INSERT NOTE (IN GUY'S HANDS) IT READS: Dear Guy: We have to meet and make plans. Call me at Arlington ----. Time's getting short. Bruno The handwriting is sprawling and erratic, embellished with conceited flourishes. Converted to PDF by www.screentalk.org 65. MEDIUM SHOT Guy looks off for a moment with set face, then tearing the note into shreds, crosses to a small desk, lights a match and holds it to the fragments, letting them burn and fall into an ash tray. Guy looks off for a moment with set face, then tearing the note into shreds, crosses to a small desk, lights a match and holds it to the fragments, letting them burn and fall into an ash tray. DISSOLVE TO: LONG SHOT EXT. MELLON GALLERY LATE AFTERNOON CAMERA is in a low setup, to take in the sign across the doorway which identifies the gallery. Hennessy stands in the foreground in front of the building, on duty. LAP DISSOLVE TO: INT. MELLON GALLERY Guy and Anne are walking slowly through a more or less deserted room of the gallery. Their manner is relaxed and intimate. ANNE Well, we'd better be getting back. GUY We've actually been alone for an hour. Seems almost indecent. You like? ANNE (softly) I like. GUY I was beginning to feel like a goldfish. ANNE So was I. When we build our house, darling, we won't even have glass windows. No doorbells, no newspapers, no telephone -- Converted to PDF by www.screentalk.org 66. GUY No Hennessy. ANNE (suddenly serious) How long can it go on? GUY I don't know. I suppose until they find out who did it. ANNE We'll be happier then, won't we? GUY I suppose so. Anne looks it him, surprised at his lack of enthusiasm. They walk on out of the picture. A figure steps out from behind a pillar in the main hall of the gallery, near the spot from which they have disappeared. It is Bruno. He calls. BRUNO (softly) Guy! Anne stops and looks back. Guy knows who it is and would not turn but that he is forced to by Anne's action. He takes a few steps towards Bruno. CLOSEUP Anne watches Guy approach this stranger. She looks downward at Bruno's tie pin. CLOSEUP Bruno's tie pin, bearing his name, gleams in the light. CLOSEUP Anne reads the name on the tie pin. TWO SHOT Guy comes up to Bruno, steps in front of him. Converted to PDF by www.screentalk.org 67. GUY (muttering harshly) Will you stop pestering me! BRUNO But Guy, you haven't called me. My father's leaving for Florida the end of this week -- GUY (interrupts) You crazy fool! There's a detective outside. He'll see us together! BRUNO (brushing this off) Oh, they can't have anything on you. (looking past Guy) Isn't that Anne Burton? Slight improvement over Miriam -- eh, Guy? GUY Stay away from me, I tell you! He leaves Bruno abruptly to rejoin Anne. Bruno looks after him, a little hurt. TWO SHOT Guy rejoins Anne and they start to walk away. ANNE Who was it, Guy? GUY (unnerved) I never saw him before. Just some tennis fan. Anne looks at him a little oddly. He seems unduly concerned about a casual stranger. CLOSEUP ANNE Her face is troubled. FADE OUT. Converted to PDF by www.screentalk.org 68. FADE IN INT. MORTON STUDY MED. SHOT Guy and a secretary have set up office in the Morton study. As the scene opens the secretary is handing Guy a large envelope. SECRETARY Here's a special delivery, Mr. Haines. It's marked personal. As Guy is opening the envelope, Barbara speaks to him from atop a library ladder. She is getting a book from one of the top shelves of a bookcase, which is next to a window. BARBARA Are you getting in any practice today, Guy? GUY (as he takes out a large folded sheet of paper and glances at it, mystified) Yes, if I can get a court at the club. As Guy's hands unfold the paper and hold it for moment, we see that it is a diagrammed plan of the grounds and the Interior of the Anthony house. There are dotted lines along the upper hall, with an arrow which points to one room and where Bruno has indicated in his handwriting, "My father's room." Over this we hear the voices of Barbara and the secretary: SECRETARY'S VOICE Barbara, who are you waving at? BARBARA'S VOICE Mr. Hennessy. I think it is a shame Daddy won't let us have him in the house to sit down. Have you met him yet, Louise? SECRETARY'S VOICE No. BARBARA'S VOICE He is awfully cute. Converted to PDF by www.screentalk.org 69. MED. SHOT Guy frowns, quickly folds the paper up and stuffs it into his pocket. He looks off abstractedly. CLOSEUP SECRETARY She looks at Guy sympathetically. SECRETARY Is anything wrong, Mr. Haines? CLOSEUP GUY Her voice breaks his reverie. He answers her with a forced smile. GUY No, thank you, Louise. FADE OUT. FADE IN TENNIS COURT AT WASHINGTON COUNTRY CLUB There are twenty or thirty people sitting in the bleacher seats opposite the umpire's chair. A game of mixed doubles is in progress. MED. SHOT AT THE ENTRANCE TO THE COURT Guy appears, carrying his racquets. His partner for the forthcoming game, and one or two other players, are close by. CLOSER SHOT Guy looks about him. Several people are looking at him awkwardly or avoiding his eyes. He moves self-consciously away, and the CAMERA PANS HIM around the court to the umpire's chair. MED. SHOT A couple of women players whisper something about Guy as he goes past them. Converted to PDF by www.screentalk.org 70. FIRST WOMAN I didn't think he'd show up after what happened. SECOND WOMAN And miss all the publicity? MED. SHOT As Guy stands at the umpire's chair, the umpire glances down and gives him a rather embarrassed greeting. CLOSEUP GUY He looks across at the watching crowd. MED. SHOT FROM GUY'S VIEWPOINT The heads of the people in the bleachers move from side to side, to follow the play on the court. One head is not moving. It is staring at Guy. It is Bruno. At this moment, we hear the umpire calling, "Game, set and match" to the winning mixed doubles pair. CLOSEUP GUY His expression becomes set. LONG SHOT The mixed doubles couples complete their handshaking at the net and move off the court. We see Guy move up to the base line while the other player takes his position for the preliminary knock-up. MED. SHOT As Guy casually knocks the ball across the net, he glances again toward Bruno. MED. SHOT FROM GUY'S VIEWPOINT Bruno is making his way out of the small stand. Converted to PDF by www.screentalk.org 71. CLOSEUP GUY Perplexed and apprehensive as to what Bruno may be up to. He hears his opponent's voice. PLAYER'S VOICE Ready, Guy? Guy shakes off his abstraction and poises himself to receive the ball. LAP DISSOLVE TO: MED. SHOT PASSAGEWAY LEADING TO TERRACE We see Guy coming alone, having fInIshed his game. He is carrying his rackets, wears a towel around his neck, etcetera. He walks into foreground, into CLOSEUP, and suddenly stops short at what he sees: MED. SHOT FROM GUY'S VIEWPOINT The group at the table comprising Bruno, Anne and the two French people. Bruno is preening himself as the others laugh uproariously, obviously at something Bruno has said. Anne catches sight of Guy and smiles at him. CLOSE SHOT GUY CAMERA MOVES WITH HIM as he comes forward toward the table. MED. SHOT GROUP AT TABLE As Guy comes into the scene. He stands staring. ANNE Guy, darling -- this is Mr. Antony -- a friend of Monsieur and Madame Darville... (to Bruno) Guy Haines. CLOSEUP GUY He gives a weak acknowledgment in Bruno's direction, realizing that Bruno has wormed his way into the group and that he must accept the introduction. Converted to PDF by www.screentalk.org 72. MEDIUM SHOT Bruno half rises, smiles affably at Guy, reaches out his hand. Guy is forced to shake hands with him BRUNO I've been a fan of yours for a long time, Mr. Haines. In fact, I follow everything you do. MME. DARVILLE Mr. Antony has been telling us such charming stories... Very funny. CLOSEUP GUY He gives another weak little smile. MED. SHOT In response to the Frenchwoman's attentive and eager expression, Bruno leans forward on the table and starts saying something more in extremely fluent French. CLOSEUP ANNE She is staring at Bruno with a new expression. CLOSEUP FROM ANNE'S VIEWPOINT Bruno's coat has spread open a bit, and his tie pin bearing the name "Bruno" is resting on the edge of the table. CLOSEUP ANNE She becomes aware that this is the man she has seen call to Guy in the art museum, that they have met before. Her eyes turn a little in Guy's direction, though she does not look at him. CLOSEUP GUY He is still watching Bruno talk to the French couple. Guy is unaware of Anne's looks. Suddenly his attention is arrested by the sound of Barbara's voice calling him. Converted to PDF by www.screentalk.org 73. BARBARA'S VOICE Guy! He turn his head and CAMERA PANS him to Barbara, who is standing a few steps from the table beckoning to him. BARBARA (Sotto voce) I've just been talking to your shadow. (very impressed) Guy, did you know Mr. Hennessy helped crack that axe murder I was reading about? You know, the one where the body was cut up and hidden in the butcher shop? He was locked in the ice box with the left leg for six hours! GUY He pulls those yarns right out of his hat, Babs. CLOSEUP GUY He gives a sharp look back toward Bruno. There is more laughter coming from the French couple at the table. CLOSE SHOT GROUP AT TABLE FROM GUY'S VIEWPOINT Bruno is occupied with his French joke, but Anne is looking at Guy strangely. TWO SHOT GUY AND BARBARA Guy turns back to Barbara. Barbara looks with interest toward Bruno. BARBARA Who's the nice looking Frenchman with the Darvilles? GUY He's not French. His name's Antony. Barbara steps toward the table. MED. SHOT AT TABLE as Barbara joins the group. Converted to PDF by www.screentalk.org 74. BARBARA How do you do, Madame Darville. Monsieur. They looks up. CLOSEUP BRUNO Bruno stops in the middle of some French to stare at Barbara. Her voice continues. BARBARA'S VOICE How are you? FRENCH COUPLES' VOICES Delightful to see you. How sweet you look, Miss Barbara. CLOSE SHOT BARBARA FROM BRUNO VIEWPOINT BARBARA I hope you aren't forgetting our little party on Thursday, Madame. From Bruno's viewpoint, as Barbara speaks,CAMERA MOVES IN CLOSER until to faintest impression of the merry-go-round fills the screen with the effect of whirling around Barbara's head. Her glasses seem to glint until her eyes are obliterated by the glare. MED. SHOT THE GROUP MME. DARVILLE We are planning on it? M. DARVILLE But of course. All talk dies out as all eyes turn to Bruno, who is staring at Barbara. Except Anne's, who is saying quietly to Bruno: ANNE This is my sister Barbara. Barbara, this is Mr. Antony. CLOSEUP BRUNO He does not acknowledge the introduction immediately. He is still staring at Barbara. Then he nods abstractedly. Converted to PDF by www.screentalk.org 75. CLOSEUP ANNE She is looking at Bruno, wondering what mystery lies behind this strange individual and why he and Guy have disclaimed any previous acquaintance. FADE OUT. FADE IN INT. GUY'S APARTMENT NIGHT CLOSEUP A LUGER PISTOL HELD IN GUY'S HANDS CAMERA PULLS BACK TO SHOW Guy staring down at it. He is partially dressed for an evening party, in black bow tie but without his jacket. He leans forward to take up a letter from among brown paper wrappings on the table. INSERT: LETTER Dear Guy -- Just two more days left. We must get together for final details. The note, in Bruno's handwriting, is unsigned. CLOSEUP GUY He stares down at the note. At this moment there is a knock at the door. MED. SHOT Guy hastily gather together the gun, the note and the wrappings and puts them in a dresser drawer. He crosses to the door and opens it. Hennessy enters, carrying a topcoat. GUY Hiya, Hennessy. Won't keep you out late tonight. (getting into his dinner jacket) With Forest Hills coming up tomorrow, I've got to get some sleep. Converted to PDF by www.screentalk.org 76. HENNESSY (helping himself to a cigarette) That's too bad. Hammond takes over in a couple of hours. I'd like to see him earn his salary. Guy turns to the dresser drawer in which he has put the note and the gun, maneuvering his body between the dresser and Hennessy's view. He takes out a handkerchief, closes the drawer, sticks the handkerchief in his pocket, speaking as he does so. GUY Doesn't that bloodhound over relax? He sticks so close he's beginning to grow on me -- like a fungus. HENNESSY (mildly) He thinks you're a very suspicious character. He doesn't trust anybody! Not even himself. Guy is eager to get out of the room, and Hennessy is maddeningly slow in his movements. GUY Come on. (indicating at Hennessy overcoat) Don't forget your sleeping bag. HENNESSY (taking his time) Yeah, If I have to wait too long on the sidewalk my feet get cold. And if I sit too long on those stone steps, my -- Guy has the door open and eases Hennessy toward the haIl. GUY (quickly) Don't worry. Since you told Barbara Burton about the icebox, you're her favorite charity. She'll send the butler out with something to defrost you. HENNESSY (grinning) Cute kid. Converted to PDF by www.screentalk.org 77. He's gone, and with a last glance at the dresser, Guy goes out and closes the door. LAP DISSOLVE TO: EXT. BURTON HOUSE LONG SHOT NIGHT The street outside the Burton house is lined with cars and limousines. Various guests are arriving. MED. SHOT On the opposite side of the street we see Hennessy, now wearing his topcoat. He looks bored as he glances across the street to the house. LAP DISSOLVE TO: INT. BURTON HOUSE BIG HEAD CLOSEUP OF ANNE Her face is troubled. CAMERA BEGINS TO PULL BACK. We see now that the reception is in progress and that Anne stands beside her father to greet the arriving guests. CAMERA PULLS BACK FURTHER to show us a full view of a very crowded Washington gathering Many white ties and tails and decollete in evidence. Many accents. Even some foreign languages are being spoken. Music and chatter in the b.g. CLOSE SHOT Anne and the Senator are still greeting new arrivals. Anne's manner is somewhat preoccupied. She glances around as she speaks, as though looking for someone. ANNE (to new arrival) Thank you so much, Mr. Lindsay. We'll look forward to it. PANNING SHOT FROM ANNE'S VIEWPOINT THE CAMERA PASSES various groups of guests in conversation including Guy and Barbara who are together. From this distance we cannot hear what they are saying. CAMERA CONTINUES TO the front door. It opens to admit a new arrival. It is Bruno. He wears white tie and tails, looking very elegant. Converted to PDF by www.screentalk.org 78. We see Guy excuse himself from Barbara, cross to Bruno and speak to him angrily, obviously asking, "What are you doing here?" Bruno, however, greets Guy with a smile then turns from him, unperturbed and bland. He sees Anne and moves toward her, smiling. CLOSEUP ANNE As Bruno comes in her direction, Anne's expression shows her mystification and concern about Bruno's presence and about Guy's attitude toward him. MED. SHOT Bruno comes up to Anne and the Senator. He gives a slight bow to the Senator; then puts his hand out to Anne. BRUNO Good evening, Miss Burton. The Senator looks inquiringly. Anne makes the introduction. ANNE This is Mr. Antony, father. SENATOR How do you do, sir. BRUNO I'd like to talk to you sometime, Senator, about my idea of harnessing the life force. It will make atomic power look like the horse and buggy. (the Senator and Anne are beginning to look at him in amazement) I'm already developing my faculty for seeing millions of miles. And, Senator, can you imagine being able to smell a flower on the planet of Mars? I'd like to lunch with you some day soon and tell you more about it. Interrupted by new arrivals, Bruno moves away out of the picture, with a charming smile to Anne. The Senator greets the new guests with open mouth and simply shakes their hands while glancing off in direction of the departing Bruno. Converted to PDF by www.screentalk.org 79. DOWAGER (to Senator) So nice to see you, my dear Senator. SENATOR Ah yes, indeed -- I beg your pardon? She realizes he hasn't heard a word she's said and haughtily moves on. The Senator turns to Anne. SENATOR (still looking after Bruno) I don't remember inviting that young man. Who is he? ANNE A friend of the Darvilles. SENATOR He has an unusual personality. Provocative. CLOSEUP ANNE She looks off in Bruno's direction extremely disturbed at this new aspect of the mysterious stranger. CLOSEUP GUY He is watching Bruno. MED. SHOT Guy sees Bruno join a group of several ladies who are seated on a settee and a couple of older men who are standing by. A waiter comes along with a tray of drinks. Bruno takes one. CLOSEUP BARBARA She comes from the same direction that Guy came. She stops short as she sees: MED. SHOT FROM BARBARA'S VIEWPOINT Bruno is now heartily joining in conversation with one of the elderly gentlemen. Converted to PDF by www.screentalk.org 80. CLOSE SHOT BRUNO AND GROUP Bruno talking to an elderly, dignified gentlemen. BRUNO But tell me, Judge, after you've sentenced a man to the chair, isn't it difficult to go and eat your dinner after that? JUDGE Young man, when a murderer is caught, he must be tried. When he is convicted, he must be sentenced. When he is sentenced to death, he must be executed. BRUNO Quite impersonal, isn't it, sir? JUDGE So it is. Besides, it doesn't happen every day. At this moment, Anne comes into the scene. She hesitates as she hears Bruno's answer. BRUNO So few murderers are caught. The Judge moves out of the way. Bruno smiles blandly at the ladies. One of them speaks to him. MRS. CUNNINGHAM Well, Mr. Antony, you seem very interested in the subject of murder. Anne looks more troubled, then moves on out of the scene. BRUNO No more than anyone else. No more than you, for instance. MRS. CUNNINGHAM Me? I'm not interested in murder. Bruno pulls up a chair to face the two woman on the settee, sits down, straddling the seat, to look at them over the back of the chair and settle down for a nice conversation. Converted to PDF by www.screentalk.org 81. BRUNO (his tone is teasing) Oh, come now, everyone's interested in that. Everyone would like to put someone out of the way. Now surely, Madame, you're not going to tell me that there hasn't been a time when you wanted to dispose of someone. Your husband, for instance. MRS. CUNNINGHAM (laughs) Good heavens, no! BRUNO (playfully) Ah ah! (shaking a finger at her) Are you sure? Do you mean to tell me there wasn't a tiny moment - when you'd been made really angry? And what did you say? MRS. CUNNINGHAM (squirms, giggling) Well... BRUNO There you are, you see! There you are! All right, now you're going -- to do a murder. How are you going to do it? This is the fascinating part -- how are you going to do it...I didn't get your name? MRS. CUNNINGHAM Mrs. Cunningham. BRUNO Mrs. Cunningham, how are you going to do it? MRS. CUNNINGHAM (entering into the spirit of the play) Well, I suppose I'll have to get a gun from somewhere. BRUNO (shakes his head) Tssk, tssk. Oh no, Mrs. Cunningham. (MORE) Converted to PDF by www.screentalk.org 82. BRUNO (CONT'D) Bang, bang, all over the place. Blood everywhere? The other woman joins in: MRS. ANDERSON What about a little poison? BRUNO Ah! That's better, that's better. Mrs.....? MRS. ANDERSON Anderson. BRUNO (he is thoroughly enjoying himself) That's better, Mrs. Anderson. But Mrs. Cunningham is in a dreadful hurry. Poison could take...let's see...ten to twelve weeks, if poor Mr. Cunningham is to die from natural causes. MRS. CUNNINGHAM I have a wonderful Idea! I can take him out in the car and when I get to a lonely spot, knock him on the head with a hammer, pour gasoline over him and over the car and start the whole thing ablaze. BRUNO (looks at her deprecatingly) And then have to walk all that way home? Mrs. Anderson laughs. BRUNO No, I have the best way, and the best tools. (he holds out his hands and shows them) Simple, silent, and quick. The silent part being the most important. Let me show you what I mean. (MORE) Converted to PDF by www.screentalk.org 83. BRUNO (CONT'D) (he raises his hands toward Mrs. Cunningham's throat, then stops a moment to ask) You don't mind if I borrow your neck for a moment do you? MRS. CUNNINGHAM (giggles) Well, it's not for long. BRUNO Oh! no. (he takes a drink and puts his glass down) Now, when I nod my head, just see if you can cry out, and I bet you can't. (he places his hands around Mrs. Cunningham's neck) Now with my two thumbs...you see that's where I'll be able to prevent any sound coming from you. Now, just wait for the nod of my head. CLOSEUP BRUNO As he starts to Press her neck, his eyes wander from the face of his "victim" to someone else off scene. MED. SHOT BARBARA She is watching this rather unorthodox demonstration. The CAMERA MOVES UP until her head fills the screen. Her glasses glint in the light. CLOSEUP BRUNO He is now transfixed. His breathing becomes heavy. A strange expression comes over his face. He still stares off at Barbara. MED. SHOT BARBARA We see the whirling merry-go-round spinning around her head. Converted to PDF by www.screentalk.org 84. BIG HEAD CLOSEUP BRUNO He now seem to have almost gone into a trance. Over the shot we begin to HEAR a strangled cry, and a broken exclamation, then Mrs. Anderson's voice. MRS. ANDERSON'S VOICE Mr. Antony! Mr. Antony! ANOTHER VOICE Stop him! Stop him! CLOSEUP Bruno's wrists and hands and the neck of his victim. We can just see Mrs. Cunningham's chin at the top of the screen. Her head is tossing from side to side. Her hands are clutching at Bruno's wrists. The hands of the other two women, also in the picture, are pulling at Bruno's wrists. Mrs. Cunningham's hands begin to slide off. Her head drops back. Over this we HEAR cries of: VOICES Stop him! Help, somebody! Pull him off! Mr. Antony! Mr. Antony! CLOSEUP BRUNO His body is swaying slightly at the various efforts to drag him away from Mrs. Cunningham. His eyes begin to close, and slowly he falls away from the picture in a dead faint on the floor. MEDIUM SHOT There is a rush of people around Mrs. Cunningham, who is breathing frantically, her eyes opening and closing. A couple of women are feebly slapping her hands, someone else is fanning her face. MEDIUM SHOT The Senator and Guy rush into the picture. They look at the fallen Bruno. They search around for an explanation. Other man come in ad they start to pick Bruno up. Converted to PDF by www.screentalk.org 85. GUY Bring him this way. Guy gives a quick look in direction of Mrs. Cunningham, sees that she is being attended to. MEDIUM SHOT Anne rushes into the picture. She sees Bruno being helped to his feet; then turns her attention to Mrs. Cunningham, who has now somewhat recovered. Mrs. Cunningham is helped to the settee. There is a babble of women's voices trying to explain what has happened. ANNE (thru the babble) Bring her upstairs. As the two groups pass off in different directions, the few people who ran into the scene late are asking the others what the disturbance is. "What's wrong?" "Did she faint?" "I didn't see anything." "What happened to him?" "Somebody hurt?" But one small figure stands in the clear. It is Barbara, She is still transfixed by what she has seen. Her hands are trembling. CAMERA MOVES SLOWLY IN ON HER. We see that her lips are trembling, too, and in her eyes frightened tears are welling. Her breath is heavy. INT. STUDY Bruno is stretched out on a settee. He is completely out. His collar and tie are open. Two or three of the male guests are just leaving the room. The Senator remains behind for a moment with Guy. SENATOR I thought he was a bit weird when he arrived. Who is he? GUY I hardly know him, sir. SENATOR Get him out of here as soon as you decently can -- will you. This is a nice item for the gossips. First thing you know, they'll be talking about orgies. I'd better get back... GUY Yes, sir. Converted to PDF by www.screentalk.org 86. The Senator leaves. Guy stands over Bruno's outstretched figure. MEDIUM SHOT Bruno is now half awake. Almost without seeing Guy, he staggers to his feet and begins to make his way to the door. Guy advances, and with a sharp thrust, pushes Bruno back on the settee. Bruno looks and sees Guy clearly for the first time. BRUNO What happened? I was on a merry-go- round somewhere. It made me dizzy. Guy moves forward, and thrusting his hand in Bruno's open shirt, pulls him to his feet. Bruno ignores Guy's violence and remain puzzled. GUY (disgusted) You're a mad, crazy maniac, and you ought to be locked-up! Now will you get out of here and let me alone? BRUNO But, Guy -- Guy smashes Bruno in the jaw, in utter disgust, and knocks him back onto the settee. Bruno looks up from his sprawled position, a dull look in his eye. BRUNO You shouldn't have done that, Guy. GUY (subsiding) Come on -- pull yourself together. Do your tie up. Bruno staggers to his feet. He fumbles at his collar. As he crosses to him, CAMERA MOVES IN to a CLOSER SHOT. GUY Here -- let me. He fixes Bruno's shirt and collar together and quickly ties his white bow. Bruno stands swaying like a small boy as Guy does this. Converted to PDF by www.screentalk.org 87. CAMERA PANS WITH THEM as Guy starts to escort Bruno from the room. GUY Have you got a car here? BRUNO (mumbling) Driver's outside. They pass trough door into the hallway. INT. HALL MEDIUM SHOT One or two of the guests turn their heads as Guy takes Bruno across to the front door. CLOSE SHOT Barbara appears in the hallway, coming from the crowded sitting room. She watches the two men go out the front door. MEDIUM SHOT Bruno and Guy going out the front door. The man-servant does not close it immediately, so we are able to HEAR the call for Mr. Antony's car. CLOSEUP BARBARA She turns her head and looks up the stairs. Barbara has not quite recovered from her ordeal. She hurries forward to greet Anne who is hurrying down the stairs. TWO SHOT CAMERA PANS DOWN with Anne as she descends the last few steps. Barbara enters the picture and the two girls meet at the foot of the stairs. ANNE What's the matter, Barbara? Did you see it happen? Did you see it -- all? Converted to PDF by www.screentalk.org 88. CLOSEUP BARBARA BARBARA (still shaken) He looked at me! His hands were on her throat, but, he was strangling me! CLOSEUP ANNE ANNE (aghast) How do you mean? TWO SHOT BARBARA He was looking at her first. Then he looked over at me. He went into a sort of trance (shudders) He looked horrible! (reflectively) He thought he was murdering me. CLOSEUP ANNE She looks away, with growing consciousness of the situation TWO SHOT BARBARA Anne, why me? Why me? What did I have to do with it? Anne is extremely concerned and thoughtful. Suddenly she gets an idea and with a pat on Barbara's arm, asks hurriedly: ANNE Do you know where Guy is? BARBARA He went out with that man! Anne hurries to the front door and passes through. Converted to PDF by www.screentalk.org 89. EXT. HOUSE Anne comes out onto the steps and looks around. She stops short as she sees: LONG SHOT EXT. STREET FROM ANNE'S VIEWPOINT There are cars lined up outside on the street. One limousine is pulling up in the center, two figures at the passenger door. One is climbing in. The other is Guy. CLOSEUP ANNE She calls out urgently: ANNE Guy! CLOSE SHOT Guy turns and closes the door. MEDIUM SHOT FROM ANNE'S VIEWPOINT The limousine moves off and Guy comes toward her. MEDIUM SHOT Anne comes down the steps and intercepts Guy on the sidewalk. She leads him along a few paces and then stops and faces him. CLOSE TWO SHOT Anne nods off in the direction of the departed Bruno and speaks in a desperate, low voice. ANNE You didn't meet him for the first time the other day, did you, Guy? Guy stares at her for a moment. GUY You mean when you introduced us at the club? Converted to PDF by www.screentalk.org 90. ANNE Yes. Did you notice how he stared at Barbara that day? GUY (awkwardly) Well, I didn't -- particularly -- ANNE (breaks in) He stared at her again tonight -- while his hands were around Mrs. Cunningham's throat. Guy looks at Anne with an expression of growing fear and alarm. She goes on inexorably: ANNE What did Miriam look like, Guy. GUY (awkwardly) Well, why do you ask me? You've seen her pictures in the paper. ANNE Go on, I want you to tell me. GUY (haltingly) Well, she was dark, not too tall, rather pretty -- ANNE What else? GUY What else is there? ANNE She wore glasses, didn't she? GUY Yes. ANNE She looked a lot like Barbara, didn't she? Guy suddenly begins to realize what Anne is getting at. Anne lowers her head, deliberately avoids looking at Guy, as she asks: Converted to PDF by www.screentalk.org 91. ANNE How did you get him to do it, Guy. GUY I get him to do it? ANNE He killed Miriam, didn't he? Tell me, Guy! GUY Yes. (suddenly bursting out) He's a maniac. I met him on the train going to Metcalf. He had a crazy scheme about exchanging murders. I do his murder and he do mine. ANNE (quietly) What do you mean -- your murder, Guy? GUY Well, he'd read about me in the paper. He knew about Miriam -- and about you. He suggested that if he got rid of Miriam for me, I should kill his father. ANNE You must have realized he was talking a lot of nonsense! GUY Of course! I didn't give it another thought. And now a lunatic wants me to kill his father. ANNE (beginning to believe) It's too fantastic! GUY (grimly) Yes, isn't it? ANNE You mean you've known about Miriam all this time? Converted to PDF by www.screentalk.org 92. GUY Since the first night. He gave me her glasses. ANNE Why didn't you call the police? GUY (bitterly) And have them say what you did -- "Mr. Haines, how did you get him to do it?" And Bruno would say we'd planed it together. ANNE Oh, Guy -- what can we do? GUY I don't know, Anne...I don't know. ANNE (With an anxious look across the street) Guy, hadn't we better go inside? Your friend Hennessy's watching us. (she Shudders) GUY (sadly) You see, Anne, that's why I didn't want you to know anything about this. I wanted to protect all of you -- your father, Barbara. And now that you know, you're acting guilty, too. ANNE (desperately) Oh, if we could only talk to father or someone about it. GUY No, that's no good, Anne. I mustn't drag anyone else into this mess. Come on. Let's go in. They go toward the house. CUT TO: Converted to PDF by www.screentalk.org 93. TWO SHOT ACROSS THE STREET As Hennessy watches Anne and Guy go toward the house, his relief, HAMMOND, comes up. Hammond's a zealous, hard-eyed sleuth. HENNESSY (a little glum) Hello, Hammond. HAMMOND You look worried. What's the matter? HENNESSY You'd better keep on your toes. Something funny's going on. DISSOLVE TO: INT. GUY'S APARTMENT LATER THAT NIGHT Still in his dinner clothes, Guy is seated in deep thought near the telephone, wrestling with his problem. There is an open telephone directory in front of him. He comes to a decision, picks up the telephone and dials a number. He waits for the answer, then: GUY Bruno? Yes, yes, it's Guy...I've decided to do what you want. I'll make that little visit to father.... (listens a moment) Tonight. (listens another moment) Yes, I want to get this thing over with, can you leave the house again, Bruno? (pause) You'd better stay out till daylight. Guy hangs up, rises and starts to move with purpose for his night's activities. DISSOLVE TO: INT. GUY'S APARTMENT NIGHT Guy is sitting at the table. He is dressed differently, having changed from his dinner clothes to a sack suit. There is only one lamp lighted in the room. Guy presents a grim picture. Converted to PDF by www.screentalk.org 94. He is studying the plan of Bruno's house, and he picks up the key Bruno sent along with it. Finally he looks at his watch, then folds the plan and puts it in his pocket with the key. He rises, crosses to the chest of drawers, opens the top drawer. INSERT: THE OPEN DRAWER Guy's hands take out the Luger. His hand then picks up Miriam's glasses from the drawer, holds them a moment. He is about to put them back, then decides to take them along, puts them into his pocket. MED. SHOT CAMERA PANS GUY across to the window. He parts the curtains slightly and looks out. MED. SHOT ON STREET (FROM GUY'S VIEWPOINT) Hammond is lighting a cigarette as he strolls in front of the house. INT. GUY'S APARTMENT Guy crosses to his door, which he opens surreptitiously. MED. SHOT CORRIDOR Guy glances down the stairs, then closes the door behind him quietly and moves away to a window at the turn of the stairs. EXT. FIRE ESCAPE Guy comes out of the window onto the second floor fire escape. He creeps stealthily down and emerges into a narrow alleyway. He steps back into the shadows for a moment when he sees: LONG SHOT FROM GUY'S VIEWPOINT (PROCESS) The strolling figure of Hammond on the far side of the street. Converted to PDF by www.screentalk.org 95. MED. SHOT Guy turns away and is soon lost in the darkness of the street. LAP DISSOLVE TO: EXT. A TALL PAIR OF ELABORATE IRON GATES NIGHT We are on the inside of the gates. We see them swing open slightly and the figure of Guy edges through them. CLOSE SHOT Guy leaves the gates ajar and then, taking the plan of Bruno's house from his pocket, and the key, he looks toward the house. EXT. STEPS LONG SHOT NIGHT This is a long flight of steps. Moonlit. They are lined with tall black cypress trees which throw their shadows across the steps. Guy moves out of one shadow, into another and carefully starts up the stairs. AT THE DOOR He pauses, looks about for a moment and listens. Then he puts the key into the lock, finding it with his flashlight. The door opens a few inches. He turns off the flash, and enters. INT. ANTONY HOME ENTRANCE HALL As Guy moves in soundlessly and closes the door. He looks toward the stairs which are in shadow. MED. SHOT Guy starts up the stairs slowly. He carries his flashlight and the plan. AT THE TOP OF THE STAIRS THE DOG A huge shadow lies it the head of the stairs. As Guy comes slowly up the stairs, the Great Dane looks down at him. Converted to PDF by www.screentalk.org 96. GUY ON THE STAIRS He reacts to the sight of the dog, stops an instant, and turn on his flashlight. The heavy massive face of the dog looks straight down at him. Guy turns off the flashlight and after a moment of indecision starts slowly up the stairs once more, the dog watching every step he takes. UPPER HALLWAY Guy comes up the last few stairs and still the dog hasn't moved. Guy slowly edges past him and the Great Dane's head turns to watch him. GUY moving quietly along the hallway, approaches two doors. He takes out his flash and identifies the door with his plan. INSERT: The plan shows two doors in relation to the stairway. The first one is clearly marked: "MY room." The adjoining door is marked: "My FATHER'S room." CLOSE SHOT GUY He pauses at the first door, then passes it quietly, walking on to the next one. He turns the knob soundlessly and passes through into the room. INT. ANTONY BEDROOM LONG SHOT The room is in darkness except for the dim outline of the recumbent figure in the bed. We hear Guy's voice, in a loud whisper: GUY Mr. Antony! The figure stirs. ANOTHER ANGLE Guy takes a stop closer to the bed. Converted to PDF by www.screentalk.org 97. GUY (urgently) Mr. Antony! Don't be alarmed -- but I must talk to you about your son. About Bruno. Mr. Antony! The figure on the bed turns and a hand stretches out toward a bedside light. The light goes on with a sudden glare. CLOSEUP FACE OF BRUNO IN THE LIGHT (LOW CAMERA) The low CAMERA throws a vast shadow up on the wall behind him, creating a grimace of his smile. BRUNO Yes, Mr. Haines? CLOSEUP GUY His face is dead. MED. SHOT Bruno rises from the bed and sits on the and of it. He is fully dressed, just as he was at the party, in white tie and tails. BRUNO (politely) My father isn't home tonight, Mr. Haines. (smiles grimly at Guy's surprise) I was about to tell you that over the phone. But you came to such a sudden decision. I wondered why. GUY (recovering quickly) Since you sent me a key to your house, I decided to use it -- to make a little social call on your father. I thought he'd be Interested to know he his a lunatic son. The faintest flicker of Bruno's eyes indicates the intensity of his reaction. He stares hard at Guy. Converted to PDF by www.screentalk.org 98. BRUNO Then a I correct, Mr. Haines, in assuming that you have no intention of going ahead with our arrangement? GUY No intention whatsoever. I never had. BRUNO I see. You won't have any further use for the key, then, Mr. Haines. (he holds out his hand and Guy gives him the key) Thank you very such. As Bruno continues to stare at him, Guy takes out the Luger. For a moment a look of fear comes into Bruno's face as he thinks Guy will probably shoot him. After a pause, Guy tosses the gun on the bed. GUY Or this. Bruno's relief turns again to menace. He picks up the gun and fingers it nervously. GUY (kindly) Look, Bruno. You're terribly sick. (haltingly) I don't know whether it's possible for you to realize it or not. I don't know much about these things, Bruno. But why don't you go someplace where you can get some treatment? Not only for your own sake, Bruno, but you can't go on causing more and more destruction to anyone you happen to meet. Bruno pays no attention. He rises. TWO SHOT Guy's arguments have made no impression on Bruno whatsoever. He fingers the gun. BRUNO I don't like to be doublecrossed. (MORE) Converted to PDF by www.screentalk.org 99. BRUNO (CONT'D) I have a murder on my conscience, but it's not my murder, Mr. Haines -- it's yours. And as you're the one to profit, I think you should be the one to pay for it. For an instant his nervous hands seem to be struggling with the urge to kill Guy. GUY (gives up) Well, I guess it's no use, Bruno. We sees to have nothing further to discuss. Bruno goes to the door in silent acquiescence and opens it for Guy to pass through. INT. HALLWAY MED. SHOT Guy walks toward the stairs, tense and apprehensive. Bruno is following him, still holding the gun. When the Great Dane sees Bruno it gets to its feet, as if waiting for a command. Guy starts down the stairs but Bruno stays where he is, the dog beside him. Gay turns and looks back it this tableaux of menace. BRUNO Don't worry. I'm not going to shoot you, Mr. Haines. It might disturb mother. (with a feeling of power) I'm a very clever follow. I'll think of something better than that. Much better. LONG SHOT Bruno remains in the foreground of the scene as Guy proceeds on down the stairs. We see him open the front door and pass through. LAP DISSOLVE TO: Converted to PDF by www.screentalk.org 100. EXT. STREET ACROSS GUY'S APARTMENT EARLY CLOSE SHOT HENNESSY AND HAMMOND MORNING Hennessy is relieving Hammond who has kept watch on Guy's apartment night. HAMMOND (in the middle of his story) He came back at three twenty-five. I didn't even know he'd given me the slip until his 'phone kept ringing for about half an hour. Nobody sleeps that sound. So I got the janitor to let me in. No Haines. HENNESSY (to himself) Wonder where he went? HAMMOND We'll probably hear of another dame murdered. HENNESSY (puzzled) Shut up. I'd better contact Metcalf. I should think this calls for more questioning of Mr. Haines. HAMMOND Questioning? Nuts! Let's take him in. HENNESSY My dear Mr. Hamond, how many times do I have to tell you that we have nothing conclusive on Haines? There's no evidence that he was ever at the scene of the crime. Can't you get that into your thick head? (quietly) Now stay put till I get back. As he starts away -- FADE OUT. Converted to PDF by www.screentalk.org 101. FADE IN INT. ANTONY LIVING ROOM LATE MORNING Anne and Mrs. Antony are in the middle of a conversation. Anne's manner is tense and purposeful, Mrs. Antony's much less serious. MRS. ANTONY Oh, now, Miss Burton, really! I know Bruno's been in some very awkward scrapes, but nothing so ridiculous as a murder. (she gives a short little laugh) ANNE (desperately) But, Mrs. Antony, you've got to make him do something about this. Don't you see that just one word from him would extricate Guy from this dreadful situation? MRS. ANTONY (lightly) Oh, but Miss Burton, I'm sure this thing must be some practical joke. You know, Bruno sometimes goes too far. (girl to girl) Of course I shouldn't be saying this to an outsider, but sometimes he's terribly irresponsible and gets into all kinds of escapades. ANNE But don't you understand, Mrs. Antony -- your son is responsible for a woman's death. MRS. ANTONY (drawing herself up with some hauteur) Did Bruno tell you this? ANNE Of course not, Mrs. Antony. MRS. ANTONY (that settles it) Well, there you are. (MORE) Converted to PDF by www.screentalk.org 102. MRS. ANTONY (CONT'D) (she sighs and rises, winding it up) Well, Miss Burton, it was very nice of you to call. You must excuse me now. I must get back to my painting. Do you care for painting, Miss Burton? I find it so soothing. (shakes Anne's hand) You must come again sometime. She goes out. Anne is left helpless, standing in the middle of the room. She picks up her purse and is about to go when she hears a voice: BRUNO'S VOICE Oh, Miss Burton! Anne turns back in direction of the voice. CAMERA PULLS BACK until we can see the feet of Bruno protruding from behind a chair in which he is sitting. He has obviously heard the entire conversation between Anne and his mother. Bruno rises. He is in dressing gown and pajamas. BRUNO I'm afraid mother wasn't very helpful, was she? (he strolls toward Anne) You know she hasn't been well for a long time. She's a little -- how shall I say -- confused. (shakes his head commiseratingly) Poor mother. Anne is too stunned to speak. BRUNO You know, I'm very upset with Guy. He shouldn't have sent you on an errand like this. ANNE Guy doesn't know I'm here, Mr. Antony. BRUNO He's been leading you up the garden path, I'm afraid. He must be very desperate to try to involve me. I've been protecting him ever since we had that conversation on the train and he told me how he hated his wife. Converted to PDF by www.screentalk.org 103. Bruno is now standing near the window a little apart from Anne, with his back to him. He takes something out of the pocket of his dressing gown and looks down at it in his hand. It is Guy's lighter. Suddenly he stuffs it back his pocket and turn back to Anne. BRUNO Why, do you know, Miss Burton, he tried to get me to go back to the island one night after dark and pick up his lighter so the police wouldn't find it? He dropped it there, you know, when -- well, that night. Anne's horror is growing. BRUNO The whole thing's been worrying me so much. But of course I couldn't do it, Miss Burton. It would have been too risky. And besides, it would have made me an accessory. Anne stares at this insane man and sinks on the settee. She starts to cry in sheer frustration. Bruno goes to her sympathetically. BRUNO Miss Burton, I know how you feel. He puts his hand on her shoulder. Anne flings it off. There is an awkward pause as Bruno looks down at her. Then he begins to look around restlessly. BRUNO Miss Burton, you must excuse me. I have an urgent appointment. (looks it his watch) I must go up and change. Now, I really must go...if you'll excuse me... He turns, starts out of the room and up the stairs in the hall. Anne watches him. STAIRWAY FROM ANNE'S VIEWPOINT Bruno turns and waves to Anne from the landing, then goes on up the stairs. Converted to PDF by www.screentalk.org 104. INT. LIVING ROOM MED. SHOT Anne slowly rises, a lonely figure in the large room, and makes her way out. DISSOLVE TO: LONG SHOT FOREST HILL STADIUM Grouped. A game is in progress. MED. SHOT A TERRACE NEAR THE MAIN STADIUM (PROCESS) where people get refreshments. There are various table with umbrellas. MED. SHOT AT ONE OF TABLE (PROCESS) Anne and Guy are seated at the table. ANNE ...And he said that if the police found your lighter there, that's all they'd need -- something to prove you were at the scene of the murder. GUY (grimly) That big lie about my wanting him to get it back means he's going to put my lighter on that island! ANNE (urgently) Guy, you'll have to get there before he does. You won't have time to play, You'd better tell them. (she nods her head in the direction of the center court) GUY Darling, if that loudspeaker announces that I'm not going to play, Hennessy bound to be suspicious He'd keep me from ever getting near Metcalf. ANNE Then I'll go. Converted to PDF by www.screentalk.org 105. GUY (quickly) No, darling. (he puts his hand on hers and speaks firmly, with concern for her safety as well as for his own situation) You stay right here and help me give Hennessy the slip after the match. ANNE But, Guy, that'll be too late! GUY (getting a thought) Didn't Bruno say that I wanted him together there one night after dark? ANNE Yes. GUY Well, that's what's in his mind now. He's not going to expose himself in broad daylight, If I can finish off this match in three sets, I'll still get there in time. REYNOLDS, Guy's opponent, enters scene behind Guy's chair. REYNOLDS We're on in a few minutes, Guy. (to Anne) How are you, Miss Morton. Anne acknowledges his greeting with a nod. GUY Okay, Tim. Be right with you. Reynolds leaves Anne and Guy rise, and as they walk toward the stadium, we can see Guy start to speak to Anne in a whisper. ENTRANCE TO COVERED STAND ALREADY SHOT Hennessy and Hammond are standing by. Converted to PDF by www.screentalk.org 106. HAMMOND Well, if Turley said to pick him up for questioning, let's pick him up! HENNESSY Let him have his game first, Hammond. HAMMOND (sourly) This is the first time I ever waited for a murder suspect to play tennis before I pulled him in. When the boys it headquarters heir about this they'll send me orchids. Guy and Anne come into the scene just as the players from the previous match emerge. They pass through, nodding to Hennessy. HENNESSY Good luck, Guy. Guy is so preoccupied with his grim doesn't nod to Hennessy until Anne nudges him. INSIDE THE STAND MED. SHOT Anne is reluctant to leave Guy who must now join his opponent, Reynolds. GUY You got it straight? (ANNE nods) Just make sure Barbara has everything ready as soon as the third set starts. He goes on to the court, and Anna goes to her box. MED. SHOT Anne joins Barbara in the box. She starts to whisper something to her. LONG SHOT Guy and Reynolds complete their warm-up as the umpire announces that Guy is to serve. The game starts. Converted to PDF by www.screentalk.org 107. EXT. ANTONY HOME A taxi is at the front door. Bruno is descending the steps. He gets into the cab, which moves off. FOREST HILLS MED. SHOT ANNOUNCER'S BOOTH (PROCESS) Over the shoulder of the announcer WE SEE the game in progress through the window of his booth. ANNOUNCER --It looks like an interesting match with Haines constantly charging the net -- not like Haines at all -- to press so early in the game... MED. SHOT TEN COURT Guy and his opponent, Reynolds, in play. Guy scores a point. CLOSEUP THE UMPIRE He announces game to Haines. MED. LONG SHOT We see the two men change ends and come toward the Umpiri's chair. Reynolds stops to take a drink of water. Guy, with an impatient glance at him, moves over to the passing line and waits, the CAMERA going with him. EXT. WASHINGTON STREET A taxicab is seen coming along. MED. SHOT INSIDE CAB (PROCESS) Bruno is sitting with in unlighted cigarette in his mouth. CAMERA MOVES IN until he is in big CLOSEUP. His eyes look down. There is the SOUND of a click, then, Guy's lighter comes up into the picture held against the cigarette. LAP DISSOLVE TO: Converted to PDF by www.screentalk.org 108. FADE IN LONG SHOT FOREST HILLS STADIUM Grouped. A game is in progress. MED. SHOT A terrace where people get refreshments. There are various tables with umbrellas. MED. SHOT AT ONE OF TABLE (PROCESS) Guy is seated. He has his rackets with him and is waiting his turn to start his match. An official is talking to him but Guy keeps looking around as if expecting someone. OFFICIAL Well, at least there'd be a trip to Australia, if you made it. GUY (absently) We'll know more about that by the end of the week... (his face brightens as he sees Anne) Anne hurries in, nods briefly to the official who has started to leave, and sits down. OFFICIAL They're close to the finish, Guy GUY Be right there. (turns to Anne) I was afraid you wouldn't get here. Wish me luck, darling. He makes a move as if to follow official toward the stadium, but Anne puts hand on his arm. ANNE (quickly and urgently) Guy, listen to me, If I sound all mixed up I can't help it. I -- I'm scared. GUY What about? Converted to PDF by www.screentalk.org 109. ANNE That's just it. I don't know. It's Bruno. I talked to him, Guy -- Guy stares at her, takes a quick look toward the stadium, then gives Anne his full attention. ANNE He acted peculiar -- as if he could put the murder right in your lap, and not involve himself at all. GUY (shaking his head) He'd drag himself into it, -- and Bruno loves Bruno. I'm all right so long as he thinks I have an alibi for that night. (noticing the stricken look on Anne is face) He knows? Anne nods slowly. GUY (grimly) Then he'll think of something. He said he would. ANNE Guy, has he anything that the police could trace to you -- (quoting Bruno) Any little thing. GUY My cigarette lighter. He said once he could have left it on the islands as evidence (a pause) But he wouldn't do that. Not in broad day light. ANNE (trying to think) But he's going somewhere, Guy. He told his mother -- GUY (tensely) Metcalf -- did he say Metcalf? Converted to PDF by www.screentalk.org 110. ANNE No, -- I don't think so. Oh, why can't I remember -- he said such crazy things! GUY (tensely) Try to think, Anne! VOICE (OFFSCENE) Guy Haines! -- Reynolds! While Anne is frantically trying to remember, Guy turns toward, the stadium and gives a signal of "Be right there." ANNE Something about the moon -- he said he had an appointment with the moon. Guy's shoulders droop with disappointment. GUY That's no help. But I can't take any chances. I've got to get that lighter -- somehow. REYNOLDS, Guy's opponent, ENTERS SCENE behind Guy's chair. REYNOLDS Okay, Guy. We're on. He walks away. Anne and Guy rise, following him. GUY I'll have to default. ANNE And have Hennessy and that other one right at your heels? Guy's expression says she's right, as they walk toward the stadium. ENTRANCE TO COVERED STAND Hennessy and Hammond, the two detectives, are standing by. HAMMOND First time I ever waited for a killer to play tennis before I nabbed him! (MORE) Converted to PDF by www.screentalk.org 111. HAMMOND (CONT'D) When the boys at headquarters hear about this they'll send me an orchid! HENNESSY We got our orders. We take him in -- after the match. Guy and Anne come INTO THE SCENE just as the players from the previous match emerge. They pass through, nodding to Hennessy. HENNESSY (a little sadly) Good luck, Guy! Guy gives him a thank-you nod. Hammond rolls his eyes in disgust at Hennessy's politeness. INSIDE THE STAND MED. SHOT Anne is about to turn to her box but she is reluctant to leave Guy, who must now join his opponent, Reynolds. As their eyes hold, in mutual helplessness, Guy suddenly stares at her with realization. GUY The moon! You said he had an appointment -- Anne looks puzzled as Guy looks up at the sun, then at his watch. GUY Then he is going to Metcalf. But he has to wait until it gets dark -- (with frantic haste, he thinks quickly, then murmurs to Anne) Listen, Anne, as soon as the third set starts, tell Barbara -- MED. CLOSE SHOT REYNOLDS waiting at the bottom of steps to the stand. Guy joins his opponent, and Anne goes to her box. Guy and Reynolds move onto the court amid the rounds of applause that greet them. Converted to PDF by www.screentalk.org 112. MEDIUM SHOT ANNE JOINS BARBARA In the box. She starts to whisper something to her. LONG SHOT Guy and Reynolds complete their warm-up as the umpire announces that Guy is to serve. The game starts. EXT. ANTONY HOME A taxi is at the front door. Bruno is descending the steps. He gets into the cab, which moves off. FOREST HILLS MED. SHOT ANNOUNCER'S BOOTH Over the shoulder of the announcer WE SEE the game in progress through the window of his booth. ANNOUNCER It looks like an interesting match -- with Haines out to blast Reynolds into a fast fight, -- not like Haines at all -- to press so early in the game... MED. SHOT THE COURT Guy and his opponent, Reynolds, in play. Guy scores a point. CLOSEUP THE UMPIRE He announces game to Haines. MED. LONG SHOT We see the two men change ends and come toward the Umpire's chair. Reynolds stops to take a drink of Water. Guy, with an impatient glance it him, moves over to the passing line and waits, the CAMERA going with him. EXT. WASHINGTON STREET A taxicab is seen coming along. Converted to PDF by www.screentalk.org 113. MED. SHOT INSIDE TAXI CAB Bruno is sitting with an unlighted cigarette in his mouth. CAMERA MOVES IN until he is in big CLOSEUP. His eyes look down. There is the SOUND of a click, then Guy's lighter comes up into the picture held against the cigarette. LAP DISSOLVE TO: INT. ANNOUNCER'S BOOTH FOREST HILL The announcer is broadcasting the progress of the match and we learn from him that this first set is nearly finished. LONG SHOT THE COURT Guy and Reynolds in play. MED. SHOT Anne and Barbara sitting in their box watching the play anxiously. MED. SHOT At the entrance to the covered stand. The two detectives Hennessy and Hammond, are watching. Hammond is bored by this game. HAMMOND Stupid game. You'd never get me into them short pants. I'd feel naked. HENNESSY (his eyes intent on the game) You'd feel naked in an Eskimo suit -- if you weren't wearing your badge. MED. SHOT Guy playing hard but holding his own. MED. SHOT Reynolds, his opponent, playing back. Converted to PDF by www.screentalk.org 114. LONG SHOT The big crowd watching. MED. SHOT Guy scores point over Reynolds. MED. SHOT There is general applause from the crowd in the covered stand as we HEAR the Umpire's announcement. UMPIRE'S VOICE (O.S.) Mr. Haines wins the first set. EXT. UNION STATION WASHINGTON D.C. We see Bruno get out of a cab and pass into the depot. LONG SHOT FOREST HILLS The game in process. MED. SHOT A nearer view of the game. CLOSE SHOT GUY IN PLAY volleying with Reynolds. CLOSE SHOT Reynolds playing the covered stand people are concentrating. MED. SHOT Guy misses a point and the game. He and Reynolds make for the Umpire's chair. We HEAR the Umpire announce. UMPIRE'S VOICE Game to Mr. Reynolds. Games are two all...Second set. Converted to PDF by www.screentalk.org 115. INT. UNION STATION WASHINGTON, D.C. Bruno is casually waiting for the train. He stands near a news-stand reading a paper. INSERT: We see that the paper is open at the sports page. There is a picture of Guy among other tennis players. WITH A DISSOLVE the whole character of this page changes with the exception of Guy's picture, which becomes surrounded with large type, announcing the arrest of Guy Haines for the murder of his wife Miriam. A sub-heading tells of Guy's cigarette lighter found at the scene of the crime. All this DISSOLVES AWAY and the page becomes once more the sports section. CLOSEUP Bruno looks up with satisfaction. LONG SHOT FOREST HILLS The crowd watching. MED. SHOT Guy and Reynolds in play. MED. SHOT Guy playing hard. MED. SHOT Reynolds playing back. CLOSEUP The Umpire watching the game. Suddenly he announces: UMPIRE Game to Mr. Reynolds. Games are three all... second set. Converted to PDF by www.screentalk.org 116. INT. CLUB CAR ON TRAIN Bruno is now seated in his accustomed place in the club car. His gloved fingers are quietly toying with Guy's lighter. A passenger next to him asks: PASSENGER May I have a light, please? Bruno looks at him for a moment and then at the lighter. With great deliberation he puts the lighter away in his pocket and takes out book-matches. Lighting a match, he holds it to his fellow passenger's cigarette. LONG SHOT FOREST HILLS The game as seen from under the covered stand. MED. SHOT Anne and Barbara very tense. CLOSEUP GUY about to serve, looks anxiously across the court. CLOSEUP THE CLOCK CLOSEUP GUY as he serves. CLOSEUP REYNOLDS returns. CLOSEUP BALL hits the net. CLOSEUP UMPIRE announces. Converted to PDF by www.screentalk.org 117. UMPIRE Second set to Haines. Haines leads two sets to love. There is a round of applause. We see the heads of the two players reach the Umpire's chair. Guy, very anxious still, as he wipes his neck with a towel. INT. COVERED STAND CLOSE SHOT ANNE BARBARA Anne is speaking. ANNE If he wins this next set -- you'd better have everything ready. (takes bill from her purse and hands it to Barbara) Here -- give the driver this ten dollars. BARBARA (puzzled) I wish understood what this is all about! ANNE (urgently) You don't have to understand, just do it. And for heaven's sake, act natural. Barbara nods and goes along. ENTRANCE TO COVERED STAND Barbara smiles winningly at Hennessy as she goes through. Her interpretation of "acting natural" is exaggerated and rather comical. Hammond's eyes narrow as he looks after her suspiciously. LONG SHOT The game in progress. Guy starts the next set. He serves. MED. SHOT Reynolds returns. Converted to PDF by www.screentalk.org 118. MED. SHOT Guy volleys. MED. SHOT Reynolds puts the ball in the air. CLOSE SHOT Guy smashes. CLOSE SHOT The ball hits the net. CLOSEUP UMPIRE UMPIRE Love fifteen. LONG SHOT THE CROWD We HEAR the smash of the ball and the voice of the Umpire. UMPIRE'S VOICE (O.S.) Love thirty. CLOSEUP ANNE looking very worried. Again the call of the Umpire. UMPIRE'S VOICE (O.S.) Double fault. Love forty. INT. THE ANNOUNCER'S BOOTH The announcer telling his listeners that Guy Haines seems to be a little reckless. ANNOUNCER -- Haines hasn't let up his terrific pace for an instant, smashing every (MORE) Converted to PDF by www.screentalk.org 119. ANNOUNCER (CONT'D) ball with a recklessness we've never seen in his playing. It's beginning to look as if he doesn't care whether he wins or loses because he's in a hurry - an awfully big hurry --- LAP DISSOLVE TO: EXT. METCALF STATION We see Bruno alight from the train. He makes his way in the direction of the town. MED. SHOT METCALF STATION As Bruno comes toward us, he stands on the sidewalk and then takes the lighter from his pocket once more. At this moment a hurrying passenger on his way to the depot accidentally jogs Bruno's elbow. The lighter flies from his hand. CLOSE SHOT We see it fall through the bars of a grating by the sidewalk. CLOSEUP BRUNO looks down in dismay. FOREST HILLS MED. SHOT The game in progress. Guy and his opponent playing hard. Guy misses a point. We HEAR the Umpire's call. UMPIRE'S VOICE (O.S) Game to Mr. Reynolds. Mr. Reynolds leads five games to three in the third set. EXT. METCALF STATION Bruno is leading a porter toward the grating, pulling him by the arm. They reach the drain. Converted to PDF by www.screentalk.org 120. BRUNO Down there -- my -- my cigarette -- (catches himself -- not wanting to say "cigarette lighter") case. It's very valuable. PORTER (peering down) Down here? BRUNO You've got to get this grating up right away. Two passersby enter. FIRST PASSERBY What's the trouble? BRUNO (yelling) Can't we do something...! (to passerby) I dropped my cigarette case. PORTER (looking down) Mightn't be any good, mister. Probably gone down the storm drain. BRUNO (horrified) Storm drain? FIRST PASSERBY On the other hand, it might have lodged on the edge. SECOND PASSERBY Don't they have a trap down there -- like under a sink? BRUNO (excited) Don't just stand here -- do something! PORTER (calmly) Guess we could phone the city engineer, all right. (MORE) Converted to PDF by www.screentalk.org 121. PORTER (CONT'D) Worst he could do would be to tell me to take a running jump and -- (Bruno grabs his arm. Porter shakes Bruno off) Relax, mister. BRUNO I don't want to relax. He goes on his knees and forces his arm down the drain. INT. THE ANNOUNCER'S BOOTH FOREST HILLS ANNOUNCER (with great excitement) This is more than a tennis game, ladies and gentlemen -- it's a desperate fight with Guy Haines playing as if his life depended on it! MED. SHOT Guy is volleying. MED. SHOT Reynolds lobs. CLOSEUP Guy smashes. CLOSE SHOT Reynolds lobs again. CLOSE SHOT Guy smashes. CLOSE SHOT Reynolds misses and the ball hits inside the line. Converted to PDF by www.screentalk.org 122. CLOSEUP The Umpire calling. UMPIRE Game to Mr. Haines. Mr. Reynolds leads five games to four...third set. EXT. METCALF STATION MED. SHOT A few more passersby have stopped to watch Bruno, whose arm is pushed through the grating. CLOSEUP Bruno's face -- straining. CLOSEUP Under the grating Bruno's hand is groping. His fingers are a long way from the lighter. LONG SHOT FOREST HILLS with the game in progress. MED. SHOT EXT. CLUB A taxi has pulled up. Barbara gets out. CLOSE SHOT She takes the ten dollar bill from her purse and passes it to the driver. She gives a final look inside the cab. CLOSEUP On the seat are Guy's everyday pants, laid out. MED. SHOT Barbara hurries out of the picture toward the club. Converted to PDF by www.screentalk.org 123. LONG SHOT The crowd watching. CLOSEUP The tense face of Anne. CLOSEUP The Umpire is somewhat impressed. INT. THE ANNOUNCER'S BOOTH CLOSEUP The announcer is telling his listeners that the score is now six-five in favor of Haines. That he has pulled up wonderfully and only needs one more game to win the match. EXT. COVERED STAND ENTRANCE Barbara, very nervous but trying to "act natural", passes Hennessy and Hammond. Hammond's eyes again follow her, but Hennessy is intent on the game. MED. SHOT FEATURING BOX As Barbara joins Anne, she gives her a surreptitious signal by ringing her thumb and forefinger, indicating everything is set. CLOSE SHOT Guy now playing hard. CLOSEUP His racket smashing at the ball. CLOSEUP Reynolds and his racket hitting the ball back. Converted to PDF by www.screentalk.org 124. CLOSEUP THE UMPIRE CALLING UMPIRE Advantages, Mr. Haines. CLOSEUP Guy serving. CLOSEUP His ball hitting the racket. CLOSEUP The ball in the net. CLOSEUP A second ball hitting the net. The Umpire's voice calling: UMPIRE'S VOICE (O.S) Duece! EXT. METCALF STATION A LOW SHOT ON Bruno bent over the grating and the onlookers behind him. BIG HEAD CLOSEUP BRUNO straining and panicky. CLOSEUP Under the grating, Bruno's fingers get near the lighter, and in their groping, they knock the lighter off the ledge, onto the ledge below. CLOSEUP Bruno's horror-stricken face. Converted to PDF by www.screentalk.org 125. FOREST HILLS MED. SHOT Guy still playing. CLOSE SHOT Barbara standing with Hennessy, watching. We HEAR the score. UMPIRE'S VOICE (O.S) Advantage, Mr. Reynolds. CLOSEUP ANNE unable to bear the suspense. She glances O.S. MED. SHOT The waiting cab. CLOSE SHOT Guy and Reynolds in play. UMPIRE'S VOICE (O.S) Score is deuce. CLOSE SHOT Reynolds serves. CLOSE SHOT Guy volleys. He waits for the return ball. He misses it. UMPIRE'S VOICE (O.S) Advantage, Mr. Reynolds. EXT. METCALF STATION ANGLE SHOOTING THROUGH the grating at CLOSEUP BRUNO'S HEAD AND SHOULDERS staining. Converted to PDF by www.screentalk.org 126. CLOSEUP Under the grating, Bruno's fingers go lower and lower, straining to reach the lighter, which is still a few inches out of reach. FOREST HILLS MED. SHOT Guy is volleying with Reynolds. INT. ANNOUNCER'S BOOTH He is very excited. ANNOUNCER -- Haines hasn't let up for a moment. If he wins this set, he wins the whole match! CLOSEUP ANNE AND BARBARA in their box. They are extremely tense. MED. SHOT Guy slams hard a shot that wins him the game. LONG SHOT CROWD applauding and shouting. CLOSE SHOT ANNE AND BARBARA At an urgent signal from Anne, Barbara hurries out as if she knew what she had to do. LONG SHOT Guy shakes hands with his opponent, and then hurries across to Anne in the stand. He leans over the front of the box. While congratulating him outwardly, she whispers something to him. He leaves his racket with her and hurries away. Converted to PDF by www.screentalk.org 127. MED. SHOT STAND ENTRANCE A block of people leaving cut off Hennessy's view. Barbara tries desperately to keep his attention off Guy. BARBARA (breathlessly) Isn't it wonderful, Mr. Hennessy? He won! It calls for a celebration. Anne says you must have dinner with us. Just the family, and you, and Guy. HENNESSY (awkwardly) Sorry I can't make it. Business. BARBARA But Guy is your business. You'll be with him, won't you? HENNESSY (a little grimly) Yeah -- I'll be with Guy. MED. SHOT Guy moving along the front of the stand making for another exit. CLOSE SHOT Barbara takes it for granted that Hennessy will accept her invitation. BARBARA Guy says you love steak -- rare, Medium, or well-done? HENNESSY I sure wish I could -- SEMI CLOSEUP Hammond is looking off. He calls into the stand. HAMMOND Hennessy! He points off toward Guy. Converted to PDF by www.screentalk.org 128. MED. SHOT Guy is hurrying toward the public entrance of the stand. SEMI CLOSEUP Hennessy and Hammond move off, leaving a dismayed Barbara. SEMI LONG SHOT Guy hurrying under the stand toward the waiting cab. MED. SHOT The two men hurrying after him. EXT. CLUB Guy goes to the waiting cab and gets in. The cab moves off. MED. SHOT The two men hurry out of the club and stand helplessly looking after the departing cab. They hurry out of the picture. CLOSE SHOT We see them grab another car. It is a chauffeur-driven limousine. Hammond jumps in front and seats himself beside the driver. Hennessy hops in the back. The car moves off. INT. LIMOUSINE TWO SHOT Hennessy finds himself seated by an old dowager about seventy- five years of age. She looks startled for a moment and almost recoils from him. He shows her his badge. HENNESSY If you'll pardon us, madam, we need your help. We're chasing a man. The old lady's eyes light up. DOWAGER How exciting. (MORE) Converted to PDF by www.screentalk.org 129. DOWAGER (CONT'D) (she leans forward and calls to the chauffeur) Hurry, O'Toole! Hurry! She leans back and maintains her air of excitement as she looks across at Hennessy. CLOSE SHOT INSIDE THE TAXI Guy is busy changing his pants. He glances over his shoulder. INT. CAR The two men looking ahead toward Guy. EXT. METCALF STATION CLOSEUP BRUN0'S FACE - ANGLE SHOOTING UP to get the peering faces behind him. Bruno still frantically trying to reach the lighter. CLOSEUP Under the grating Bruno's fingers slowly closing in on the lighter. They barely manage to grasp it. CLOSEUP BRUN0'S FACE -- triumphant. CLOSEUP Bruno's fist, holding the lighter, comes through the grating. CLOSE SHOT Bruno straightens up. CAMERA BACK as all the onlookers turn their heads in his direction. ONLOOKER You sure must think a lot of that -- Whatever it is. Converted to PDF by www.screentalk.org 130. Bruno doesn't answer. With the lighter in his closed fist, he darts through the crowd, the people looking after him. LONG SHOT The sun is much lower. INT. CLUB CAR Guy is now glancing at his watch. The sun is behind him and very much lower. EXT. AMUSEMENT PARK Bruno is looking at his watch and then across at the sky. LONG SHOT FROM HIS VIEWPOINT The last trace of the setting sun has gone. EXT. METCALF STATION MED. SHOT Guy is stepping off the train. He crosses to a waiting taxi, CAMERA FOLLOWING him. CLOSE SHOT GUY (to the driver) The amusement park, quick. As he gets in the Cab, we go to -- CLOSE SHOT MAN watching Guy get into taxi. As we hear the taxi drive away, the man hurries across to a waiting police car. CLOSE SHOT He puts his head in the side window and tells the two waiting detectives where Guy has gone. MAN Amusement park. Converted to PDF by www.screentalk.org 131. We see one of the detectives take up a microphone as the car drives off. EXT. AMUSEMENT PARK It is now getting dark. MED. SHOT Bruno leaves his spot at the side of the tent and ambles over toward the queue of people waiting for boats. CLOSE SHOT BRUNO joining the queue. He glances ahead of him. MED. SHOT FROM HIS VIEWPOINT We see the light above the pay booth go on, shedding a downward glare. CLOSE SHOT BRUNO pulls his hat a little further over his, eyes. Some new arrivals join the queue behind him. INT. TAXI Guy looking anxiously ahead on his way to the amusement park. AMUSEMENT PARK ENTRANCE We see a police car arrive. One uniformed man and two detectives get out of the car and make their way toward the entrance. One of to detectives stands at the entrance while the other two hurry into the grounds. MED. SHOT Guy's taxi arrives. MED. SHOT Across the street, another police car arrives. Converted to PDF by www.screentalk.org 132. MED. SHOT At Guy is paying his cab fare, he glances around him. MED. SHOT FROM HIS VIEWPOINT He sees one police car. CLOSE SHOT GUY gives a furtive glance around while waiting for his change. MED. SHOT ANOTHER POLICE CAR MED. SHOT Guy cautiously makes his way toward the entrance to the Amusement Park. MED. SHOT Guy passes the waiting detective and looks off. From his viewpoint we see: MED. SHOT THE TWO DETECTIVES who were at the station indicate Guy is the man. MED. SHOT One detective turns away and starts to follow Guy. CLOSE SHOT BRUNO in the queue of people. He is edging slowly along. He is about ten people away from the entrance. He suddenly looks ahead and sees. FROM HIS VIEWPOINT The uniformed man and the detective are talking casually to the boat men in charge of the concession. Converted to PDF by www.screentalk.org 133. DETECTIVE (to boatman) The killer is here tonight. So keep your eyes open and the minute you see him, let us know. CLOSE SHOT The boatman looks at them with an expression of alarm. CLOSE SHOT BRUNO begins to look a little uneasy. We see him begin to mentally deliberate. MED. SHOT Guy, threading his way through the crowds, conscious that he is being followed, but nevertheless, on the lookout for Bruno. CLOSE SHOT BRUNO moving along the line. CAMERA MOVES IN until his head and shoulders fill the screen. He is now coming within range of the flood-lit pay-box. The light seems to creep up across his chest and slowly reveal his face. He lowers his head. MED. SHOT The boatman begins to look along the queue. There is an expression of growing recognition on his face. MED. SHOT Bruno sees this, makes a decision and casually deserts the queue of people. MED. SHOT The boatman hurries across to the uniformed man and begins to talk to him excitedly, looking in Bruno's direction. MED. SHOT GUY Coming along and looking for Bruno. His eyes light up. Converted to PDF by www.screentalk.org 134. SEMI-LONG SHOT FROM HIS VIEWPOINT We see Bruno making his way from the queue of people. CLOSE SHOT GUY calls to Bruno. GUY Hey, Bruno. CLOSE SHOT BRUNO gives a quick glance back, sees Guy then he turns and looks off in another direction. SEMI-LONG SHOT The uniformed man and the boatman approaching him. CLOSE SHOT Bruno hurries on. He stop short as he sees. SEMI-LONG SHOT FROM HIS VIEWPOINT Another uniformed man. MED. SHOT Bruno starts to run. MED. SHOT Guy starts to run after him. MED. SHOT Bruno is seen to jump on a merry-go-round, which is just starting. Its pace is already pretty fast. MED. SHOT Guy runs toward Bruno. Converted to PDF by www.screentalk.org 135. CLOSE SHOT DETECTIVE Haines! Hold it! Hold it! The detective pulls out his gun and starts to run after Guy. SEMI-LONG SHOT Guy jumps on the merry-go-round after Bruno. Its speed is so great that he nearly gets flung off. CLOSE SHOT The detective fires at Guy. CLOSE SHOT The man running the machine in the center of the merry-go- round is suddenly hit in the shoulder. CLOSE SHOT His hand, which is on the starting lever, jerks it down. MED. SHOT The detective, after Guy, jumps on the machine but is flung off on his back. FULL SHOT The merry-go-round has now started to increase the speed. CLOSE SHOT Bruno at the far side is trying to jump off, but it's going too fast. LONG SHOT FROM HIS VIEWPOINT We see the hard ground whizzing past him. Everything seems to be a blur. We get a glimpse of screaming women and the crowds rushing up from the midway. Converted to PDF by www.screentalk.org 136. CLOSE SHOT BRUNO He turns and glances over his shoulder. MED. SHOT FROM HIS VIEWPOINT Guy is threading his way between the rising and falling horses. Guy gets right up close to him. TWO SHOT As Guy comes near to Bruno, the latter turns on him and starts to attack him. BRUNO I want to get off of here! Let me off of here! It makes me dizzy. GUY Stop it, Bruno. Give me my lighter, Bruno! MED.SHOT Against the whirling background of the merry-go-round, Turley and Campbell rush up as the detective struggles to his feet, slightly hurt. The noise from the calliope is very loud. CAMPBELL (to Turley, puzzled; indicating the merry- go-round) Who's the man he's fighting with on there? At this moment the boatman rushes up. BOATMAN (excited) There he is! That's the one! That's the one who killed her! TURLEY Of course he is. We know that. CLOSE SHOT ON MERRY-GO-ROUND Guy and Bruno in a struggle. Guy has to protect himself from a madman whose hands attempt to reach his throat. Converted to PDF by www.screentalk.org 137. They are staggering across between the rising and falling horses. MED. SHOT OUTSIDE MERRY-GO-ROUND A detective turns to the group around hIm. DETECTIVE Get somebody to come and stop that thing! An elderly man in soiled work clothes speaks up. WORKMAN I'll handle it. Immediately the workman heads straight for the merry-go-round and starts to crawl under it on his stomach. DETECTIVE (calls after him) Hey! Be careful! Stop! A second detective speaks to him quizzically. 2ND DETECTIVE Well, do you want to do it yourself? The first detective leans over and looks off toward the workman who is continuing his slithering way under the machine, then straightens up. 1ST DETECTIVE (changing his mind) No. I think he'll make it all right. MED. SHOT GUY AND BRUNO Bruno swings around till his back is to us. He pushes Guy toward the edge, but Guy manages to grab the rein of the nearest horse. The momentum of the machine swings Guy around against the horse, whose big head towers in the f.g. Bruno, on this side of the horse pushes forward and tries to grab the reins from Guy's hand. He tries to slash at Guy's face. The back of Bruno's head is toward us during this. Guy suddenly leans out across the horse and smashes his fist against Bruno's face. Bruno's head goes back until it is in the f.g. in a upside-down position. Converted to PDF by www.screentalk.org 138. MED. SHOT THE CAMERA IS LOW so that we get the effect of Bruno falling into the CAMERA from Guy's blow! MED. SHOT In the f.g. is a young boy of four years. He is excited by the speed of the ride and laughs at the fight with great enjoyment. He sees this by suddenly glancing over his shoulder. In the b.g. Guy and Bruno are continuing their fight. Bruno rises. Guy staggers after him. Bruno again leaps upon Guy. The two men sway toward the CAMERA until Bruno gets alongside the little boy. The boy now shows some anxiety. The three figures now fill the screen with the horses' heads in the f.g. Bruno is forced against the little boy, who now, alarmed, beats Bruno on the cheek with one hand, the other holding onto the brass rail in front. Bruno stops and with a sweep of his arm, knocks the little boy off the horse onto the floor below. The little boy, in falling, grabs the horse's rein or stirrup. CLOSE SHOT Guy breaks away from Bruno and dives around the back of the horse to grab the little boy. CLOSE SHOT As Guy grabs the boy, he staggers forward with him to a small gondola. Bruno leaps onto his back but Guy manages to put the boy in the gondola. CLOSE SHOT UNDERNEATH THE WHIRLING MERRY-GO-ROUND The boat man is making slow progress. FROM HIS VIEW POINT We see his goal. It is the wounded mechanic in the center, who is slightly stirring. All during this the base of the merry-go-round is skimming above the back and head of the boat man. Converted to PDF by www.screentalk.org 139. BACK ON MERRY-GO-ROUND The two men are now in a clinch. Guy tries to fight off the maddened Bruno. They are flung between the horses, bouncing one against the other, almost half way around the merry-go- round. CLOSE SHOT BRUNO AND GUY Again they struggles between two horses. On each side of them are two young screaming girls. The two bounce from one horse to the other. CLOSE SHOT The calliope has little figures and these boat away on their cymbals almost as though they are applauding what's going on. CLOSE SHOT Underneath the merry-go-round, the boat man has made further progress. He is creeping inch by inch. His nose starts to run. He starts to fumble for a dirty piece of handkerchief. He blows his nose and then moves on. CLOSE SHOT Back above the two men swinging past the two girls on their horse and they both crash to the floor underneath another horse, upon which is riding side-saddle, a mother and her three-year-old little girl. CLOSEUP The two big heads of the men, battling. The two men roll underneath the horse's hoofs, which are seen rising and falling. They get right underneath one horse. CLOSEUP Guy has turned over on his back and his eyes look up. Converted to PDF by www.screentalk.org 140. CLOSEUP FROM HIS VIEWPOINT We see the big horse's head above and its hoofs coming down toward the CAMERA and filling the screen. We get a faint impression of the screaming mother hugging her child to her breast, above. BIG CLOSEUP THE HORSE'S HOOFS striking Guy's head. CLOSE SHOT Guy wrenches himself out of this position. He rolls away from the CAMERA right to the edge of the merry-go-round. He manages to grab a rail. MED. SHOT Guy's body is flung out horizontally. We see the crowd behind back-up for fear of being knocked over. The screw of tension increase. Over this comes the sound of an approaching ambulance siren. CLOSE SHOT Bruno edges himself toward Guy. He is hanging on to the reins of a horse. His feet manage to roach Guy's knuckles. CLOSEUP BRUNO'S VICIOUS EXPRESSION CLOSEUP BRUNO'S FEET kicking at Guy's knuckles. CLOSEUP GUY'S AGONIZED EXPRESSION MED. SHOT A flash of the horror-stricken faces of the spectators seen through the whirling machine. Converted to PDF by www.screentalk.org 141. CLOSEUP Machinery and the lever that was pulled on too fast. The Boatman's hand comes up into the picture and pulls the lever over. LONG SHOT The sudden braking causes the whole merry-go-round to topple over with a grinding roar. LONG SHOT FROM HIGH ANGLE The merry-go-round his keeled over. For a moment we don't know who has survived. There is a surge of people milling and shouting. Those who have jumped back out of the way when the merry-go-round toppled, now rush forward again as the cloud of dust settles. From the midway in the background others are running forward. MED. LONG SHOT Distraught parents try to force their way to their children who were on the merry-go-round, but are hold back from the wreckage by police. MED. CLOSE SHOT Guy is somewhat stunned from his fall. He is helped to his feet by some men in the crowd. His knuckles are bleeding. In the background people are rushing about. The crowd is in uproar as women and children are helped from the wreckage. Officials and uniformed policemen pushing back the surge of the crowd. AD LIBS Get back. Get back there. Give us room here. Turley and Campbell rush in to Guy. TURLEY Are you all right, Haines? GUY Yes, I think so. Converted to PDF by www.screentalk.org 142. Guy is surrounded by police and Campbell stands at his elbow. At this moment the boatman runs in. One of the detectives is with him. DETECTIVE Mr. Turley! Mr. Turley! (indicating boatman) He says this isn't the man we want. (with a nod in Guy's direction) It's the other one -- the one he was fighting with. TURLEY (stops to give his full attention to this unexpected bit of information) What do you mean, this isn't the -- (turns to Guy, not quite taking it in) Not Haines? (back to boatman) But you said he was. You pointed him out. BOATMAN No, I didn't, sir. I've never seen this man before in my life. I meant the other one. The detective who was holding Guy instinctively relax his hold on Guy's arm. Turley turns to Guy, puzzled. TURLEY What is this all about, Haines? Did you know he killed your wife? GUY (nods) He has my cigarette lighter and wanted to plant it there on the island to pin the whole thing on me. (urgently) Let me talk to him. Let me show you. Where is he? ANOTHER DETECTIVE Over here. He leads the way. They follow. Converted to PDF by www.screentalk.org 143. MED. CLOSE SHOT as Guy and Turley enter to the spot where Bruno is pinned under the overturned machine. He is caught between two of the horses, the head of one of them across his chest. Bruno's head sags back somewhat, but is resting on pieces of debris. A uniformed policeman looks up from Bruno to Turley: POLICEMAN This one's in a pretty bad way, Mr. Turley. Guy is shocked at the sight of Bruno. GUY (looking down at Bruno) Can't you get that stuff off him? POLICEMAN No, they've done everything they can until the crane comes. Bruno opens his eyes and sees Guy. BRUNO Hello, Guy. Turley has leaned forward to look at the helpless Bruno. BRUNO (weakly nodding at Turley) Who's that? GUY This is Mr. Turley, Chief of Police. BRUNO (with a half smile) So they got you at last, eh, Guy? Guy looks around desperately, frustrated for a moment as Turley eyes him stonily. Then he turns again to Bruno. GUY (rather gently) Can you talk a little? Can you tell the chief you have my lighter? Converted to PDF by www.screentalk.org 144. BRUNO (with a faint, quizzical smile) I haven't got it. It's still on the island where you left it. Guy looks around helplessly to Turley, who looks back at him suspiciously. DETECTIVE (looking down at Bruno) I think he's going. Turley leans over to look. CLOSE SHOT BRUNO'S FIST FROM TURLEY'S VIEWPOINT As Bruno is dying, his closed fist slowly starts to open. DETECTIVE'S VOICE He's finished. Guy's lighter is now revealed in Bruno's open hand. MED. SHOT GROUP Turley takes the lighter from the dead Bruno's hand. Guy is watching him. Turley straightens up and holds the lighter out to him. TURLEY Is this your lighter, Haines? Guy nods without speaking, and with a half look in Bruno's direction. TURLEY Well, you were right. (sticks the lighter into his own pocket) I'd better keep this for the time being. (in a friendly tone) We can clear the whole thing out the morning. How about staying in town over night, Haines? I imagine you have a lot to tell me. Nine o'clock, all right? Converted to PDF by www.screentalk.org 145. GUY (nods) Okay, Mr. Turley. Thanks. Turley turns back to the group around Bruno. Guy looks down for a moment at Bruno, then speaks to the boatman, who is standing nearby. GUY Can you tell me where there's a telephone? BOATMAN (indicating) There's one up near the entrance. (with a look back to the dead Bruno) Who was he, Bud? Guy looks back sympathetically in Bruno's direction, speaks without looking at the boatman. GUY Bruno. Bruno Antony. (reminiscently and a little compassionately, remembering what Bruno had said of himself) A very clever fellow. He moves off through the crowd. DISSOLVE TO: INT. BURTON STUDY NIGHT Anne, Barbara and the Senator are sitting silently in the attitudes of waiting. The telephone rings. Anne is instantly on her feet. Barbara and the Senator watch her anxiously as she goes to answer it. ANNE (into phone) Hello... (impatiently) Yes, operator, yes! (waits a moment, then eagerly:) Guy? (MORE) Converted to PDF by www.screentalk.org 146. ANNE (CONT'D) (a pause, then she closes her eyes with heartfelt relief. Another pause, then:) Yes, darling, yes. Of course I'll be there...Goodbye. She hangs up, turns slowly, to face Barbara and her father. Her expression is one of intense relief. ANNE Guy'll be back tomorrow. (overcome with emotion she has difficulty in speaking) He wants me to take him some things. With a sob, Barbara flings herself into Anne's arms. As she cries, Anne strokes her head comfortingly. Then with a half- choked sobs Anne, too, begins to cry. She speaks through her tears, looking over Barbara's shoulder at her father. ANNE He says he looks silly in his tennis clothes. The Senator eyes them a moment, then speaks a little wryly: SENATOR I presume from all those tears that you have had good news. DISSOLVE TO: INT. PARLOR OF TRAIN NEXT DAY Anne and Guy are sitting quietly together. Opposite them is a man in a clerical collar who is reading a sports magazine. On the cover is a picture of a tennis player in action. The man looks over the top of his magazine at Guy, with recognition. He leans forward. CLERIC I beg your pardon, but aren't you Guy Haines? GUY (uncomfortably) Yes. Converted to PDF by www.screentalk.org 147. Guy and Anne exchange a brief look, rise hurriedly and start to walk away before the conversation can go any farther. The cleric looks after them with a frown and a puzzled shrug of his shoulders, as if to say, "Did I say something wrong?" FADE OUT. THE END | 1 |
| FROZEN Written by Jennifer Lee Final Shooting Draft 9/23/13 OPEN ON: ICE. We're underwater looking up at it. A saw cuts through, heading right for us. EXT. SNOW-CAPPED MOUNTAINS -- DUSK ICE HARVESTERS, dressed in traditional Sami clothing, score a frozen lake. They SING. "The Frozen Heart (Ice Worker's Song)" ICE HARVESTERS BORN OF COLD AND WINTER AIR AND MOUNTAIN RAIN COMBINING, THIS ICY FORCE BOTH FOUL AND FAIR HAS A FROZEN HEART WORTH MINING. The men drag giant ice blocks through channels of water. ICE HARVESTERS (CONT'D) CUT THROUGH THE HEART, COLD AND CLEAR. STRIKE FOR LOVE AND STRIKE FOR FEAR. SEE THE BEAUTY SHARP AND SHEER. SPLIT THE ICE APART! AND BREAK THE FROZEN HEART. Hup! Ho! Watch your step! Let it go! A young Sami boy, KRISTOFF (8), and his reindeer calf, SVEN, share a carrot as they try to keep up with the men. ICE HARVESTERS (CONT'D) Hup! Ho! Watch your step! Let it go! Young Kristoff struggles to get a block of ice out of the water. He fails, ends up soaked. Sven licks his wet cheek. ICE HARVESTERS (CONT'D) BEAUTIFUL! POWERFUL! DANGEROUS! COLD! ICE HAS A MAGIC CAN'T BE CONTROLLED. A sharp ice floe overtakes the workers, threateningly. They fight it back. ICE HARVESTERS (CONT'D) STRONGER THAN ONE, STRONGER THAN TEN STRONGER THAN A HUNDRED MEN! Massive fjord horses drag heavy ice plows. 2 FROZEN - J. Lee ICE HARVESTERS (CONT'D) BORN OF COLD AND WINTER AIR AND MOUNTAIN RAIN COMBINING The sun sets. Lanterns are lit. ICE HARVESTERS (CONT'D) THIS ICY FORCE BOTH FOUL AND FAIR HAS A FROZEN HEART WORTH MINING. CUT THROUGH THE HEART, COLD AND CLEAR. In the dark, Kristoff and Sven finally manage to get a single block of ice out of the water. ICE HARVESTERS (CONT'D) STRIKE FOR LOVE AND STRIKE FOR FEAR. THERE'S BEAUTY AND THERE'S DANGER HERE. SPLIT THE ICE APART! BEWARE THE FROZEN HEART. The workers pile onto the giant horse-drawn ice sled as it pulls away. Left behind, Kristoff and Sven push their ice block onto a dinky little sled then head off. We sweep up from them to the Northern Lights filling the sky...then move across the mountains...beneath the snowline...and descend upon... EXT. THE KINGDOM OF ARENDELLE -- NIGHT A humble castle, built of wood, nestled in a deep fjord. INT. CASTLE, NURSERY -- NIGHT ELSA (8) sleeps in her bed. Her little sister ANNA (5) pops up beside her. YOUNG ANNA Elsa. Psst. Elsa! Psst. Elsa doesn't stir. Anna sits on Elsa and bounces. YOUNG ANNA (CONT'D) Wake up. Wake up. Wake up. YOUNG ELSA (grumbling) Anna, go back to sleep. Anna rolls onto her back and spreads all her weight on Elsa. 3 FROZEN - J. Lee YOUNG ANNA (drama queen-ish) I just can't. The sky's awake, so I'm awake, so we have to play. YOUNG ELSA ...Go play by yourself. Elsa shoves Anna off the bed. Anna lands butt to floor, sighs, defeated. But then she gets an idea. She hops back on the bed and lifts one of Elsa's eyelids. YOUNG ANNA (mischievously) Do you want to build a snowman? Elsa's eyes both pop open. She smiles. INT. CASTLE STAIRCASE -- NIGHT Anna, now wearing snow boots, pulls Elsa by the hand. YOUNG ANNA Come on, come on, come on, come on. Elsa tries to shush her, but Anna's too excited. INT. BALLROOM -- NIGHT The girls sneak into the ballroom. Elsa shuts the door. YOUNG ANNA Do the magic! Do the magic! Elsa laughs and waves her hands together. Snowflakes suddenly burst forth and dance between her palms, forming a snowball. Elsa throws the snowball high into the air. Snow bursts out and flurries around the room. Anna dances about, catching flakes in her palms and mouth. YOUNG ANNA (CONT'D) This is amazing! YOUNG ELSA Watch this! Elsa stomps her little slippered foot and a layer of ice suddenly coats the floor, forming a giant ice rink. Anna slides off, laughing. 4 FROZEN - J. Lee PLAY MONTAGE: -Anna and Elsa roll giant snowballs and build a snowman together. Elsa moves his stick arms around. YOUNG ELSA (CONT'D) (goofy voice) Hi, I'm Olaf and I like warm hugs. Anna jumps up and hugs him. YOUNG ANNA I love you, Olaf. -Anna and Olaf appear to be dancing. REVEAL: Elsa is actually propelling them across the ice floor with her magic. -The girls slide down snowbanks together! -Anna fearlessly jumps off a snow peak into mid air. YOUNG ANNA (CONT'D) Catch me! Elsa makes another peak to catch Anna. YOUNG ELSA Gotcha! Anna keeps jumping. Elsa keeps casting magic. YOUNG ANNA (jumping faster) Again! Again! YOUNG ELSA (struggling to keep up) Slow down! Elsa suddenly slips. Her magic accidentally STRIKES Anna in the head. Anna tumbles down a snowbank and lands, unconscious. YOUNG ELSA (CONT'D) ANNA! Elsa runs to Anna and takes her in her arms. A streak of Anna's hair, where struck, turns white. YOUNG ELSA (CONT'D) MAMA! PAPA! The room around them fills with frightening ice spikes. 5 FROZEN - J. Lee The parents burst through the frozen door. GASP at the sight of the room. KING Elsa, what have you done? This is getting out of hand! QUEEN (seeing Anna) Anna! The King and Queen rush to Anna and take her in their arms. ELSA It was an accident. I'm sorry, Anna. QUEEN (about Anna) She's ice cold. KING ...I know where we have to go. SLAM CUT TO: INT. DARK ROOM -- NIGHT The King sifts through a shelf to find an ancient book inscribed with Old Norse runes. He opens the book, scrambles to a page with an ancient map. EXT. ARENDELLE -- NIGHT Carrying the girls, the King and Queen ride their horses out of the kingdom. Snow streams from Elsa's hands, leaving a trail of ice behind them. EXT. FJORD MOUNTAIN FOREST -- NIGHT A sleepy Kristoff and Sven travel alone through the dark woods. All of a sudden, the King and Queen race by with the girls, leaving the wake of ice. KRISTOFF Ice? SLAM CUT TO: 6 FROZEN - J. Lee EXT. BLACK MOUNTAINS -- NIGHT Kristoff rides Sven as they follow the trail of ice. YOUNG KRISTOFF Faster, Sven! EXT. THE VALLEY OF THE LIVING ROCK -- NIGHT Kristoff hops off Sven at the edge of a deep valley. They hide behind a rock and peek out. Down below, the King holds a frightened Elsa. The Queen holds the still unconscious Anna. KING Please, help. My daughter! Suddenly, a bunch of rocks tumble down the valley toward them. It looks as though they'll be crushed! But, luckily, the rocks stop at their feet. The rocks then unfold, revealing bright faces. YOUNG KRISTOFF Trolls...? The rock in front of Kristoff "wakes up." Meet BULDA. BULDA Shush. I'm trying to listen. She grabs Kristoff and Sven by hand and hoof and hugs them close. Sven licks her face and she eyes them both. BULDA (CONT'D) Cuties. I'm gonna keep you. Back below, the crowd parts for a troll as old as the Earth. They call him GRAND PABBIE. He approaches arthritically, but determined. He nods respectfully to the king. GRAND PABBIE Your Majesty. (referring to Elsa) Born with the powers or cursed? KING Born. And they're getting stronger. Grand Pabbie motions for the Queen to bring Anna to him. She does. He examines her. 7 FROZEN - J. Lee GRAND PABBIE (about Anna) You are lucky it wasn't her heart. The heart is not so easily changed, but the head can be persuaded. KING Do what you must. GRAND PABBIE I recommend we remove all magic, even memories of magic to be safe.... But don't worry, I'll leave the fun. Grand Pabbie pulls out a glowing blue energy from Anna's head. We see her memories floating right above her. Grand Pabbie changes all of her magical memories to ordinary memories -- snowy play indoors with the girls in their nightgowns changes to outdoors on the winter fjords with the girls in winter gear. He puts the ordinary memories back in her head. GRAND PABBIE (CONT'D) She will be okay. YOUNG ELSA But she won't remember I have powers? KING It's for the best. PABBIE Listen to me, Elsa, your power will only grow. As he speaks, he conducts the Northern Lights to show a silhouette of an adult Elsa creating magical snowflakes. PABBIE (CONT'D) There is beauty in your magic.... But also great danger. The snowflakes turn to sharp spikes. PABBIE (O.S.) (CONT'D) You must learn to control it. In the Northern Lights display, the sharp spikes cause human figures to panic and attack Elsa. PABBIE (CONT'D) Fear will be your enemy. 8 FROZEN - J. Lee Elsa gasps and buries her face in the King's chest. The King wraps his arms around Elsa, protectively. KING No. We'll protect her. She can learn to control it. I'm sure. Over the King's words we... DISSOLVE TO: -The Arendelle castle gates shutting. KING (O.S.) (CONT'D) Until then, we'll lock the gates. We'll reduce the staff. We will limit her contact with people and keep her powers hidden from everyone... including Anna. -The castle shutters close. -Anna sits on her bed as Elsa's furniture disappears. -Anna rushes to the hall to see Elsa shut the door to her new room. Anna watches, confused and sad. DISSOLVE TO: INT. CASTLE WINDOW -- DAY We look out on a gentle snowfall. Little Anna skips up to the window. She lights up at the sight of the snow and rushes down the hall. INT. HALLWAY, ELSA'S DOOR -- DAY Anna knocks on Elsa's door and SINGS. "Do You Want to Build a Snowman?" YOUNG ANNA DO YOU WANT TO BUILD A SNOWMAN? COME ON LET'S GO AND PLAY. Anna peeks under the door. YOUNG ANNA (CONT'D) I NEVER SEE YOU ANYMORE. COME OUT THE DOOR. IT'S LIKE YOU'VE GONE AWAY. 9 FROZEN - J. Lee -INT. ANNA'S ROOM -- Anna plays with two dolls, gives up, sad. YOUNG ANNA (CONT'D) WE USED TO BE BEST BUDDIES AND NOW WE'RE NOT. I WISH YOU WOULD TELL ME WHY. -ELSA'S DOOR. Anna peeks through the key hole. YOUNG ANNA (CONT'D) DO YOU WANT TO BUILD A SNOWMAN? -Anna calls through the keyhole. YOUNG ANNA (CONT'D) IT DOESN'T HAVE TO BE A SNOWMAN. YOUNG ELSA (O.S.) Go away, Anna. YOUNG ANNA (hearbroken) ...OKAY BYE. -BEHIND THE DOOR -- DAY. Elsa sits at the window looking out, longingly. Suddenly, her icy hands freeze the windowsill. -LATER. The King slips leather gloves onto Elsa's hands. KING The gloves will help. He pats her gloved hand. KING (CONT'D) See? You're good.... (starting their mantra) Conceal it. YOUNG ELSA Don't feel it. YOUNG ELSA & KING Don't let it show. -INT. HALLWAY, ELSA'S DOOR -- DAY. Anna, now 9, knocks on Elsa's door. ANNA (9) DO YOU WANT TO BUILD A SNOWMAN? -INT. HALLWAY -- DAY. Alone, Anna rides a bicycle built for two in the hall by standing on the back seat. 10 FROZEN - J. Lee ANNA (9) (CONT'D) OR RIDE OUR BIKE AROUND THE HALL? I THINK SOME COMPANY IS OVERDUE... -INT. PORTRAIT ROOM -- DAY. Anna runs around the portrait room, gaining momentum to flip over the arm of the couch. ANNA (9) (CONT'D) I'VE STARTED TALKING TO THE PICTURES ON THE WALLS. Anna lands PLOP on the cushions, then looks up at the painting above her of the courageous Joan of Arc. ANNA (9) (CONT'D) Hang in there, Joan. -INT. EMPTY LIBRARY -- DAY. Looks like no one's around. ANNA (9) (CONT'D) IT GETS A LITTLE LONELY ALL THESE EMPTY ROOMS. But then we find Anna, laying at the base of the grandfather clock, playing with her braids, bored out of her mind. ANNA (9) (CONT'D) JUST WATCHING THE HOURS TICK BY. Anna's eyes follow the grandfather clock's pendulum. ANNA (9) (CONT'D) TICK TOCK. TICK TOCK. TICK TOCK. -INT. ELSA'S ROOM -- NIGHT. Elsa (now 12) paces as she panics. The entire wall is frozen behind her. ELSA (12) I'm scared. It's getting stronger. KING Getting upset only makes it worse. The King goes to hug her. ELSA (12) No. Don't touch me. I don't want to hurt you. He and the Queen look at each other with alarmed sadness. -INT. LIBRARY -- DAY. Anna, now a teenager, slides past Elsa's room without stopping. 11 FROZEN - J. Lee -INT. KING AND QUEEN'S QUARTERS -- DAY. Anna runs into the room and throws herself into her parents' arms. TEEN ANNA See you in two weeks. -INT. ELSA'S ROOM -- DAY. Elsa curtsies in front of her parents, formally, not touching them. TEEN ELSA Do you have to go? KING You'll be fine, Elsa. -EXT. DOCKS -- DAY. The King and Queen leave on a ship. -EXT. ROUGH SEAS -- NIGHT. Lightning flashes. The sea rages in a storm. The King and Queen's ship is lost in the waves. -INT. CASTLE -- DAY. A portrait of the King and Queen is covered in mourning cloth. -EXT. CEMETERY -- DAY. Anna looks small, standing before her people, beside burial stones. -INT. HALLWAY, ELSA'S DOOR. Anna, still in her mourning clothes, approaches and knocks. ANNA (singing) Elsa? PLEASE I KNOW YOU'RE IN THERE PEOPLE ARE ASKING WHERE YOU'VE BEEN THEY SAY HAVE COURAGE AND I'M TRYING TO I'M RIGHT OUT HERE FOR YOU. PLEASE LET ME IN. Anna slides down the door and sits with her head against it. ANNA (CONT'D) WE ONLY HAVE EACH OTHER. IT'S JUST YOU AND ME. WHAT ARE WE GONNA DO? (weak, internal) DO YOU WANT TO BUILD A SNOWMAN? We move through the door... -INT. ELSA'S ROOM -- DAY. Elsa is sitting in the exact same pose as Anna. Her bedroom is frozen with ice. Snowflakes hang in the air, suspended by grief. FADE OUT. 12 FROZEN - J. Lee EXT. THE KINGDOM OF ARENDELLE -- MORNING A new dawn rises over the fjords. Ships pull up to the docks. Guests pile out. DOCK MASTER Welcome to Arendelle! A BOY tries to get away as his MOTHER tries to stuff him in his bunad jacket. BOY Why do I have to wear this? MOTHER Because the Queen has come of age. It's Coronation Day! BOY That's not my fault. They pass the May Pole being raised and a Sami ice harvester chatting with his reindeer. We recognize them as Kristoff and Sven, all grown up. Sven hops around excitedly like a dog and nuzzles Kristoff's chest. KRISTOFF What do you want, Sven? Kristoff leans in and speaks for Sven, as if he can. KRISTOFF (AS SVEN) (CONT'D) Give me a snack. KRISTOFF (CONT'D) What's the magic word? KRISTOFF (AS SVEN) (CONT'D) Please! Kristoff pulls a carrot out of his shirt pocket and hands it to Sven. Sven tries to bite the whole thing. KRISTOFF (CONT'D) Hey, hey, hey! Share! Sven takes a smaller bite. Kristoff then has a bite himself, not seeming to care that it's covered in reindeer slobber. We move on to PERSI and AGGIE, a super-excited couple who rush towards the castle. 13 FROZEN - J. Lee PERSI I can't believe they're finally opening up the gates! AGGIE And for a whole day! Faster, Persi! They pass a tiny but menacing DUKE, who wears taps on his shoes to "enhance" his presence. Two THUG guards follow close behind him. DUKE Ah, Arendelle, our most mysterious trade partner. Open those gates so I may unlock your secrets and exploit your riches. (catching himself) ...Did I just say that out loud? We leave him and head down the bridge towards the castle gates, passing an Irishman and a Spanish Dignitary. IRISHMAN Oh, me sore eyes can't wait to see the Queen and the Princess. I bet they're absolutely lovely. SPANISH DIGNITARY I bet they are beautiful. We move past them, to a particular castle window. CUT TO: INT. CASTLE, ANNA'S BEDROOM -- DAY Anna, 18, snores. Drools. KNOCK. KNOCK. KAI (O.S.) Princess Anna...? Anna sits up. She's got major bedhead. She coughs. Snorts. Pulls a hair from her mouth. ANNA ...Huh? Yeah? KAI (O.S.) Sorry to wake you, ma'am but-- ANNA No, you didn't. I've been up for hours. 14 FROZEN - J. Lee She falls back asleep while sitting. She snores. Her head drops, startling her awake. ANNA (CONT'D) Who is it? KAI (O.S.) It's still me, ma'am. Time to get ready. ANNA Ready for what? KAI (O.S.) Your sister's coronation, ma'am. ANNA My sister's cor-neration... One eye opens enough to catch sight of her coronation dress. She bolts, wide awake in excitement. ANNA (CONT'D) Coronation Day! Ha ha! SLAM CUT TO: EXT. CASTLE HALL -- DAY Anna bursts out of her room, wearing her coronation dress. She finishes pinning ribbons in her hair. Seeing the hustle and bustle of preparations, she can't help but SING. "For the First Time in Forever" ANNA THE WINDOW IS OPEN! SO'S THAT DOOR! I DIDN'T KNOW THEY DID THAT ANYMORE. WHO KNEW WE OWNED 8000 SALAD PLATES...? -Anna slides along the floor of the ballroom in her socks. ANNA (CONT'D) FOR YEARS I HAVE ROAMED THESE EMPTY HALLS WHY HAVE A BALLROOM WITH NO BALLS? FINALLY, THEY'RE OPENING UP THE GATES! -She shakes hands with a suit of armor. Breaks it. Hides the evidence. 15 FROZEN - J. Lee ANNA (CONT'D) THERE'LL BE REAL, ACTUAL PEOPLE - IT'LL BE TOTALLY STRANGE. BUT WOW AM I SO READY FOR THIS CHANGE! -Anna comes to a window and jumps out onto a window washer's pulley. She raises herself up to see the ships arriving. ANNA (CONT'D) FOR THE FIRST TIME IN FOREVER, THERE'LL BE MUSIC, THERE'LL BE LIGHT. FOR THE FIRST TIME IN FOREVER, I'LL BE DANCING THROUGH THE NIGHT. -Anna walks through the garden and follows a family of geese. ANNA (CONT'D) DON'T KNOW IF I'M ELATED OR GASSY, BUT I'M SOMEWHERE IN THAT ZONE 'CAUSE FOR THE FIRST TIME IN FOREVER, I WON'T BE ALONE. (speaking) I can't wait to meet everyone.... (GASP) What if I meet THE ONE? -Anna twists herself in a velvet drape like it's a gown. She acts like she looks gorgeous, but she looks ridiculous. ANNA (CONT'D) TONIGHT, IMAGINE ME GOWN AND ALL- FETCHINGLY DRAPED AGAINST THE WALL. THE PICTURE OF SOPHISTICATED GRACE. -She notices the bust of a man across the room. ANNA (CONT'D) (google-eyed) I SUDDENLY SEE HIM STANDING THERE, A BEAUTIFUL STRANGER TALL AND FAIR. (mouth full of chocolate) I WANNA STUFF SOME CHOCOLATE IN MY FACE! -She grabs the bust of the man and swings it around. ANNA (CONT'D) BUT THEN WE LAUGH AND TALK ALL EVENING, WHICH IS TOTALLY BIZARRE. NOTHING LIKE THE LIFE I'VE LED SO FAR. The bust goes flying and lands on the top of the cake. -Anna bursts into the portrait room, bounces on the furniture, and interacts with the paintings. 16 FROZEN - J. Lee ANNA (CONT'D) FOR THE FIRST TIME IN FOREVER, THERE'LL BE MAGIC, THERE'LL BE FUN. FOR THE FIRST TIME IN FOREVER, I COULD BE NOTICED BY SOMEONE. AND I KNOW IT IS TOTALLY CRAZY TO DREAM I'D FIND ROMANCE. BUT FOR THE FIRST TIME IN FOREVER, AT LEAST I'VE GOT A CHANCE! -INT. LIBRARY. ELSA, now a very poised 21, watches out the window as the coronation guests arrive. ELSA DON'T LET THEM IN. DON'T LET THEM SEE. BE THE GOOD GIRL YOU ALWAYS HAVE TO BE. Elsa moves to a painting of her father's coronation. She takes off her gloves and mimics the painting by holding a candlestick and ornament in place of an orb and scepter. ELSA (CONT'D) CONCEAL. DON'T FEEL. PUT ON A SHOW. MAKE ONE WRONG MOVE AND EVERYONE WILL KNOW. The candlestick and ornament ice over. Elsa gasps, slams them back down onto the table. She tries to reassure herself. ELSA (CONT'D) BUT IT'S ONLY FOR TODAY. We cut between Anna's excitement and Elsa's nerves. ANNA IT'S ONLY FOR TODAY! ELSA IT'S AGONY TO WAIT. ANNA IT'S AGONY TO WAIT!!! ELSA TELL THE GUARDS TO OPEN UP THE GATE. ANNA THE GATE!!! -Finally, the gates are open! Anna moves through the crowd, admiring the people around her. 17 FROZEN - J. Lee ANNA (CONT'D) ELSA FOR THE FIRST TIME IN DON'T LET THEM IN FOREVER. DON'T LET THEM SEE ANNA ELSA I'M GETTING WHAT I'M DREAMING BE THE GOOD GIRL OF YOU ALWAYS HAVE TO BE ANNA ELSA A CHANCE TO LEAVE MY SISTER'S CONCEAL. WORLD CONCEAL. DON'T FEEL. A CHANCE TO FIND TRUE LOVE DON'T LET THEM KNOW. -Anna hurries over the bridge and into the village square. ANNA (CONT'D) I KNOW IT ALL ENDS TOMORROW, SO IT HAS TO BE TODAY!! `CAUSE FOR THE FIRST TIME IN FOREVER. . . FOR THE FIRST TIME IN FOREVER! NOTHING'S IN MY WAY!!! -Anna SLAMS right into the breast of a HORSE! She falls back and lands in a small wooden boat. It tips off of the dock. She's heading overboard. But just then, the horse slams his hoof into the boat and steadies it. ANNA (CONT'D) (frustrated) Hey! HANS I'm so sorry. Are you hurt? The rider, HANS, sure is handsome and regal. ANNA (gentler) Hey. I-ya, no. No. I'm okay. HANS Are you sure? ANNA Yeah, I just wasn't looking where I was going. But I'm okay. He hops down from his horse and steps into the boat. ANNA (CONT'D) I'm great, actually. 18 FROZEN - J. Lee HANS Oh, thank goodness. He offers her a hand and their eyes meet. Chemistry. He helps her to her feet. HANS (CONT'D) (bowing) Prince Hans of the Southern Isles. ANNA (curtseying) Princess Anna of Arendelle. HANS Princess...? My Lady. He drops to his knees, head bowed. The horse bows too, curling his hoof up and out of the boat. The boat tips. Hans tumbles on top of Anna. Awkward. ANNA Hi...again. The horse slams his foot back into the boat to stabilize it. Anna and Hans tumble the other way. Anna lands on top of him. HANS Oh boy. ANNA Ha. This is awkward. Not you're awkward, but just because we're-- I'm awkward. You're gorgeous. (did she just say that?) Wait, what? Hans quickly gets to his feet and helps Anna up again. HANS I'd like to formally apologize for hitting the Princess of Arendelle with my horse...and for every moment after. ANNA No. No-no. It's fine. I'm not THAT Princess. I mean, if you'd hit my sister Elsa, that would be-- yeash! `Cuz, you know... (patting the horse) Hello. (MORE) 19 FROZEN - J. Lee ANNA (CONT'D) (to Hans) But, lucky you, it's-it's just me. HANS Just you? Hans smiles, amused. She smiles back. The bells RING. She doesn't notice at first; she's too busy drinking in Hans's handsomeness. ANNA ...The bells. The coronation. I-I-I better go. I have to...I better go. She hurries off, stops, turns back. Gives Hans a little wave. ANNA (CONT'D) Bye! As she rushes off again, Hans waves back. The horse waves too, once again taking his hoof out of the boat. HANS Oh no. The boat falls, with Hans in it. SPLASH! It lands upside down in the water. Hans raises it up off of him, gasping for air. CUT TO: INT. CHURCH CHAPEL -- DAY Elsa stands at the alter. Anna stands off to one side. She peeks out to the audience. Hans waves at her from the pews. He's changed his clothes. The crown is placed on Elsa's head. The scepter and orb are presented to Elsa on a pillow. She slowly reaches for them. BISHOP (a whisper) Your Majesty, the gloves. Elsa hesitates. She breathes nervously, removes her gloves, places them on the pillow. Her hands shake. She takes the orb and scepter, then turns to the people. BISHOP (CONT'D) (formal, in Old Norse) Sehm hon HELL-drr IN-um HELL-gum AYG-num ok krund ee THES-um HELL- gah STAHTH, ehk teh frahm FUR-ear U- thear... 20 FROZEN - J. Lee The scepter and orb start to freeze over. BISHOP (CONT'D) ...Queen Elsa of Arendelle. CROWD Queen Elsa of Arendelle. Just in time. Elsa manages to set the orb and scepter back down on the pillow before anyone notices the ice. She picks up her gloves and slips them on. She made it. CUT TO: INT. GREAT HALL -- NIGHT Springy music fills the Great Hall. Guests dance. Eat. Laugh. TRUMPETS SOUND. KAI (announcing) Queen Elsa of Arendelle. Elsa enters, poised and looking surprisingly content. She stands under a formal awning. KAI (CONT'D) Princess Anna of Arendelle! Anna runs into the room, waves awkwardly. Kai ushers her over to stand right next to Elsa. ANNA Here? Are you sure? She and Elsa sneak awkward peeks at each other. ELSA ...Hi. ANNA Hi me...? Oh. Um. Hi. ELSA ...You look beautiful. ANNA Thank you. You look beautifuller. I mean, not fuller. You don't look fuller, but more beautiful. 21 FROZEN - J. Lee ELSA Thank you. They look out at the celebration. ELSA (CONT'D) So, this is what a party looks like? ANNA It's warmer than I thought. ELSA And what is that amazing smell? They both close their eyes and inhale. ANNA AND ELSA (TOGETHER) ...Chocolate. Their eyes pop open. They laugh. Elsa looks back out at the party. Anna looks at Elsa. She wants to say so much, but she can't think of where to start. Just as she finds her way, Kai interrupts. KAI Your Majesty. The Duke of Weaseltown. DUKE Weselton. The Duke of Weselton. (to Elsa) Your Majesty, as your closest partner in trade, it seems only fitting that I offer you your first dance as queen. The Duke does a funny flitter of his feet, a hitch-kick, and a deep bow. DUKE (CONT'D) (whispers to himself) One, two, three. Jump. As he holds out his hand, head down, his toupee dips forward. Anna giggles. Elsa looks at Anna, stifles a giggle herself. ELSA (to the Duke) Thank you...only I don't dance. 22 FROZEN - J. Lee DUKE (offended) Oh...? ELSA But my sister does. ANNA What? DUKE Lucky you.... ANNA Oh, I don't think-- The Duke grabs Anna's arm and yanks her away before she can protest. DUKE If you swoon, let me know, I'll catch you. Anna looks back at Elsa, desperately. ELSA Sorry. OUT ON THE DANCE FLOOR: The Duke showboats, but he's just awful. Anna tries to make the best of it. DUKE Like an agile peacock... CLUCK- CLUGGLE-CLUCK! He lands on her feet. ANNA Ow. Ow. DUKE Speaking of, so great to have the gates open. Why did they shut them in the first place? Do you know the reason? Hmm? He gets in her face, suspicious. ANNA ...No. 23 FROZEN - J. Lee DUKE Oh, all right. Hang on. They don't call me the little dipper for nothing. He dips Anna back. Elsa peeks through the crowd, can barely hold in her laughter. Anna shoots Elsa funny, help-me looks. DUKE (CONT'D) (groove fully on) Like a chicken...with the face of a monkey...I fly. JUMP CUT TO: MOMENTS LATER... Anna limps back to Elsa. DUKE (O.S.) Let me know when you're ready for another round, M'Lady. ELSA Well, he was sprightly. ANNA (rubbing her sore feet) Especially for a man in heels. ELSA Are you okay? ANNA (loving Elsa's attention) I've never been better. This is so nice. I wish it could be like this all the time. ELSA (sincere) Me too.... But then Elsa catches herself. She stiffens up, looks away. ELSA (CONT'D) But it can't. ANNA Why not? If-- ELSA It just can't. 24 FROZEN - J. Lee Anna's smile drops. She tries not to get emotional. ANNA Excuse me for a minute. She walks away. Elsa watches her go, saddened. Moving through the crowd, Anna gets bumped by a bowing man's butt. She falls. Just before she hits the floor, Hans catches her. He smiles perfectly. HANS Glad I caught you. ANNA Hans. He smoothly sets his drink down on a passing tray. He lifts her up and leads her in a romantic dance. DISSOLVE TO: LATER: Anna and Hans drink and chat. ANNA (CONT'D) I often had the whole parlor to myself to slide... Oops. Sorry. She hits him in the face by mistake with her hand. He laughs. DISSOLVE TO: -THE CASTLE DOORS: Anna and Hans stroll out of the castle. ANNA (CONT'D) ...Your physique helps I'm sure. DISSOLVE TO: -THE ROSE GARDEN... Hans notices her white streak. HANS (about her white streak) What's this? ANNA I was born with it, although I dreamt I was kissed by a troll. HANS I like it. DISSOLVE TO: 25 FROZEN - J. Lee EXT. BALCONY -- NIGHT Anna teaches Hans how to eat krumkake. ANNA Yeah, the whole thing! You got it. They laugh as the krumkake crumbles in his face. ANNA(CONT'D) Okay wait, wait. So you have how many brothers? HANS Twelve older brothers. Three of them pretended I was invisible... literally...for two years. ANNA That's horrible. HANS It's what brothers do. ANNA ...And sisters. Elsa and I were really close when we were little. But then, one day she just shut me out, and I never knew why. He takes her hand. Leans in close. HANS I would never shut you out. ANNA Okay, can I just say something crazy? HANS I love crazy. "Love is an Open Door" ANNA (singing) ALL MY LIFE HAS BEEN A SERIES OF DOORS IN MY FACE. AND THEN SUDDENLY I BUMP INTO YOU. HANS I was thinking the same thing, because like. . . (MORE) 26 FROZEN - J. Lee HANS (CONT'D) I'VE BEEN SEARCHING MY WHOLE LIFE TO FIND MY OWN PLACE. AND MAYBE IT'S THE PARTY TALKING, OR THE CHOCOLATE FONDUE. ANNA BUT WITH YOU- HANS BUT WITH YOU, I FOUND MY PLACE. ANNA I SEE YOUR FACE. BOTH AND IT'S NOTHING LIKE I'VE EVER KNOWN BEFORE. They jump to the neighboring balcony and enter a door. They come out on top of one of the castle's towers. BOTH (CONT'D) LOVE IS AN OPEN DOOR! LOVE IS AN OPEN DOOR! Cut to them sliding across an empty hallway in their socks. BOTH (CONT'D) LOVE IS AN OPEN DOOR ANNA WITH YOU! HANS WITH YOU! ANNA WITH YOU! HANS WITH YOU! BOTH LOVE IS AN OPEN DOOR. They hop up on the castle roof and watch a shooting star. HANS I MEAN IT'S CRAZY. ANNA What? 27 FROZEN - J. Lee HANS WE FINISH EACH OTHER'S- ANNA SANDWICHES! HANS That's what I was gonna say! They slide down the back of the roof out of sight. We next find them strutting on a bridge ledge. ANNA I'VE NEVER MET SOMEONE- BOTH WHO THINKS SO MUCH LIKE ME. BOTH (SPOKEN) (CONT'D) Jinx.. . .jinx again. Are they doing the robot? No. They're imitating the mechanical figures on the clock tower. BOTH (CONT'D) OUR MENTAL SYNCHRONIZATION CAN HAVE BUT ONE EXPLANATION, HANS YOU- ANNA AND I- HANS WERE- ANNA JUST- BOTH MEANT TO BE. Anna and Hans dance on top of the lighthouse and cast dancing shadows across the sails of ships in the docks. ANNA SAY GOODBYE- HANS SAY GOODBYE- 28 FROZEN - J. Lee BOTH TO THE PAIN OF THE PAST. BOTH (CONT'D) WE DON'T HAVE TO FEEL IT ANYMORE! LOVE IS AN OPEN- They play hide and seek amongst the stable doors. BOTH (CONT'D) DOOR! LOVE IS AN OPEN DOOR! They climb to the waterfall looking out over the kingdom. Anna raises up her hands to frame the moon. Hans puts his hands on top of hers. Together their hands form a heart. BOTH (CONT'D) LIFE CAN BE SO MUCH MORE- ANNA WITH YOU! HANS WITH YOU! ANNA WITH YOU! HANS WITH YOU! BOTH LOVE IS AN OPEN HANS DOOR. ANNA DOOR. HANS Can I say something crazy...? Will you marry me? ANNA Can I just say something even crazier? Yes. CUT TO: 29 FROZEN - J. Lee INT. BALL -- NIGHT Anna pushes through the crowd towards Elsa, Hans in tow. ANNA Oops! Pardon. Sorry. Can we just get around you there? Thank you. Oh, there she is. Elsa! Elsa turns to Anna. Anna curtseys awkwardly. ANNA (CONT'D) I mean...Queen.... Me again. Um. May I present Prince Hans of the Southern Isles. HANS (bowing) Your Majesty. Elsa gives a polite but reserved curtsey. ANNA We would like-- HANS --your blessing-- ANNA --of-- ANNA/HANS --our marriage! ELSA Marriage...? ANNA Yes! ELSA I'm sorry, I'm confused. ANNA Well, we haven't worked out all the details ourselves. We'll need a few days to plan the ceremony. Of course we'll have soup, roast, and ice cream and then-- Wait. Would we live here? ELSA Here? 30 FROZEN - J. Lee HANS Absolutely! ELSA Anna-- ANNA Oh, we can invite all twelve of your brothers to stay with us-- ELSA What? No, no, no, no, no. ANNA Of course we have the room. I don't know. Some of them must-- ELSA Wait. Slow down. No one's brothers are staying here. No one is getting married. ANNA Wait, what? ELSA May I talk to you, please. Alone. Anna sees Hans's worried face. Hooks arms with him. ANNA No. Whatever you have to say, you- you can say to both of us. ELSA Fine. You can't marry a man you just met. ANNA You can if it's true love. ELSA Anna, what do you know about true love? ANNA More than you. All you know is how to shut people out. ELSA You asked for my blessing, but my answer is no. Now, excuse me. 31 FROZEN - J. Lee HANS Your Majesty, if I may ease your-- ELSA (flustered) No, you may not. And I-I think you should go. Elsa walks away. As she passes the Royal Handler-- ELSA (CONT'D) The party is over. Close the gates. ANNA What? Elsa, no. No, wait! Anna grabs Elsa's hand. She pulls off Elsa's glove. Elsa gasps, spins around and reaches for the glove in panic. ELSA Give me my glove! Anna holds the glove away from Elsa. ANNA (desperate) Elsa, please. Please. I can't live like this anymore. Elsa fights tears. ELSA (weak) ...Then leave. Elsa sees Anna's hurt face. It's too much. She can't hold it in. She turns and rushes away. ANNA (heartbroken) ...What did I ever do to you?! The party goes silent as everyone watches the sisters. ELSA Enough, Anna. ANNA No. Why? Why do you shut me out?! Why do you shut the world out?! What are you so afraid of?! ELSA I said, enough! 32 FROZEN - J. Lee Ice shoots from Elsa's hand, spikes across the floor! Guests cry out in shock, back away. DUKE (ducking behind his men) ...Sorcery. I knew there was something dubious going on here. ANNA Elsa...? Elsa rushes out of the room. CUT TO: EXT. COURTYARD -- NIGHT Elsa bursts out of the castle door. The CITIZENS CHEER! CROWD There she is. Your Majesty! Long live the Queen! Queen Elsa.... Come drink with us. Elsa ducks through the crowd, holding her bare hand. BOWING TOWNSMAN Queen Elsa. TOWNSWOMAN WITH BABY Your Majesty? Are you all right? Elsa backs away from the baby. She knocks into the fountain, grabs its edge. The waters freeze at her touch. GASPS of shock and fear sweep over the crowd. The Duke and thugs come out the door. DUKE There she is! Stop her! ELSA (to the Duke) Please, just stay away from me. Stay away! Magic accidentally shoots from her hand and turns the staircase into ice. The thugs and the Duke fall. DUKE Monster.... Monster! 33 FROZEN - J. Lee The crowd panics. A snowstorm begins. Elsa flees. Anna runs out of the palace doors, carrying the glove. ANNA Elsa! Hans follows closely behind her. GATES TO THE KINGDOM: Elsa runs out of the gates and down to the water's edge. The shoreline freezes under her feet. Anna calls to her from the gates. ANNA (CONT'D) Elsa! Wait, please! Elsa glances back at Anna, but turns away. She tentatively steps out onto the fjord. It freezes instantly. She breaks into a run, as the water freezes over with each step. ANNA (CONT'D) Elsa, stop! Anna rushes out onto the fjord ice, slips, falls. HANS Anna! Hans rushes to Anna's side. Elsa reaches the far shore. She doesn't look back. She just scrambles into the mountains. ANNA No. HANS (shocked) Look.... The fjord. The ice spreads out until the entire fjord is frozen, locking the ships in place. INT. CASTLE COURTYARD -- NIGHT Snow falls. Hans and Anna move through the panicking crowd. CROWD WALLAH Snow? It's...snow...in July. 34 FROZEN - J. Lee HANS ...Are you all right? ANNA (in shock) No. HANS Did you know? ANNA No. Nearby, the Duke flutters about in fright. DUKE Look! It's snowing! It's snowing! The Queen has cursed this land! She must be stopped! (to his thugs) You have to go after her. Anna rushes up to the Duke. ANNA Wait, no! The Duke hides behind his thugs and points out at Anna. DUKE You! Is there sorcery in you, too? Are you a monster, too? ANNA No. No. I'm completely ordinary. HANS That's right she is... (realizing how that sounds) ...in the best way. ANNA ...And my sister's not a monster. DUKE She nearly killed me. HANS You slipped on ice. DUKE Her ice! 35 FROZEN - J. Lee ANNA It was an accident. She was scared. She didn't mean it. She didn't mean any of this.... Tonight was my fault. I pushed her. So I'm the one that needs to go after her. DUKE Yes. Fine. Do. HANS What? ANNA (to the Royal Handler) Bring me my horse, please. HANS Anna, no. It's too dangerous. ANNA Elsa's not dangerous. I'll bring her back, and I'll make this right. The Royal Handler brings Anna her horse and a cloak. HANS I'm coming with you. ANNA No, I need you here to take care of Arendelle. He sees the desperation in her eyes. HANS ...On my honor. She throws on the cloak and hops right onto the horse, coronation dress and all. ANNA (to the crowd) I leave Prince Hans in charge! HANS (before letting her go) Are you sure you can trust her? I don't want you getting hurt. ANNA She's my sister; she would never hurt me. 36 FROZEN - J. Lee She snaps the reins and rides out. Hans watches after her. The snow picks up and overtakes our view. We push through a blizzard...lose our way...then finds ourselves... EXT. HIGH UP IN THE MOUNTAINS -- NIGHT Well above the snow-line, a small figure climbs the highest peak. It's Elsa. Finally, she stops, looks around. Catches her breath and sings... "Let It Go" ELSA THE SNOW GLOWS WHITE ON THE MOUNTAIN TONIGHT, NOT A FOOTPRINT TO BE SEEN. A KINGDOM OF ISOLATION AND IT LOOKS LIKE I'M THE QUEEN. THE WIND IS HOWLING LIKE THIS SWIRLING STORM INSIDE. COULDN'T KEEP IT IN, HEAVEN KNOWS I TRIED. . . DON'T LET THEM IN, DON'T LET THEM SEE, BE THE GOOD GIRL YOU ALWAYS HAVE TO BE. CONCEAL, DON'T FEEL, DON'T LET THEM KNOW. WELL, NOW THEY KNOW. Elsa takes off her glove and throws it into the air. ELSA (CONT'D) LET IT GO. LET IT GO. CAN'T HOLD IT BACK ANYMORE. Elsa creates a snowman, just like the one she made with Anna when they were children. ELSA (CONT'D) LET IT GO. LET IT GO. TURN AWAY AND SLAM THE DOOR. I DON'T CARE WHAT THEY'RE GOING TO SAY. LET THE STORM RAGE ON. THE COLD NEVER BOTHERED ME ANYWAY. Elsa lets her cape fly back into the wind. 37 FROZEN - J. Lee ELSA (CONT'D) IT'S FUNNY HOW SOME DISTANCE MAKES EVERYTHING SEEM SMALL. AND THE FEARS THAT ONCE CONTROLLED ME CAN'T GET TO ME AT ALL. IT'S TIME TO SEE WHAT I CAN DO, TO TEST THE LIMITS AND BREAK THROUGH. NO RIGHT, NO WRONG, NO RULES FOR ME...I'M FREE! Elsa creates ice steps and climbs them. ELSA (CONT'D) LET IT GO! LET IT GO! I AM ONE WITH THE WIND AND SKY. LET IT GO! LET IT GO! YOU'LL NEVER SEE ME CRY. HERE I STAND AND HERE I'LL STAY. Elsa slams her foot down and forms a giant snowflake. ELSA (CONT'D) LET THE STORM RAGE ON.... In a flurry of creative release, she raises the snowflake on ice beams, builds walls, archways, a glistening chandelier, and an intricate ceiling that leaves the sky visible. ELSA (CONT'D) MY POWER FLURRIES THROUGH THE AIR INTO THE GROUND. MY SOUL IS SPIRALING IN FROZEN FRACTALS ALL AROUND. AND ONE THOUGHT CRYSTALLIZES LIKE AN ICY BLAST- Standing firmly in her mighty ice palace, Elsa removes her crown and throws it. ELSA (CONT'D) I'M NEVER GOING BACK, (back to resolve) THE PAST IS IN THE PAST! She takes down her hair and creates a new dress made of ice. ELSA (CONT'D) LET IT GO! LET IT GO! AND I'LL RISE LIKE THE BREAK OF DAWN. LET IT GO! LET IT GO! The sun rises. Elsa struts onto out onto a balcony and into the light. She's free. 38 FROZEN - J. Lee ELSA (CONT'D) THAT PERFECT GIRL IS GONE. HERE I STAND IN THE LIGHT OF DAY. LET THE STORM RAGE ON!! THE COLD NEVER BOTHERED ME ANYWAY. She turns and slams her ice palace door on us. CUT TO: EXT. THE FJORD FOREST -- DAY Anna rides her horse through two feet of snow. She shivers. ANNA (shivering) Elsa! Elsa! It's me, Anna...your sister who didn't mean to make you freeze the summer. I'm sorry. It's all my f-f-f-f-f-f-fault. DISSOLVE TO: LATER: Anna and the horse struggle through a wooded area. ANNA (CONT'D) (hearing a wolf howl) Of course, none of this would have happened if she'd just told me her secret...ha...she's a stinker. A branch of a nearby tree snaps and startles the horse. Anna goes flying off, lands face down in the snow. She sits up. Spits out snow. Sees the horse running away. ANNA (CONT'D) Oh no. No. No. No. Come back. No. No. No. No.... Oooo-kay. He doesn't come back. Anna grabs onto a branch of a leaning conifer, tries to pull herself to her feet, but the tree snaps upright and releases all its snow onto her. GROAN. DISSOLVE TO: EXT. MOUNTAIN -- NIGHT The Northern Lights shine as Anna struggles, out of breath, reaching the top of a hill. 39 FROZEN - J. Lee ANNA Snow, it had to be snow, she couldn't have had tr-tr-tropical magic that covered the f-f-fjords in white sand and warm -- She sees smoke rising up in the distance. ANNA (CONT'D) Fire! WHOA! Anna goes tumbling down the hill. She lands with a crash in an icy stream at the bottom. ANNA (CONT'D) (from inside the snowball) Cold, cold, cold, cold, cold... EXT. A SMALL BUILDING AND STABLE -- NIGHT Anna shuffles up to the building, her dress frozen stiff. She shakes the snow off a sign and reads: ANNA Wandering Oaken's Trading Post. Snow drops off a smaller sign. She reads it, happily. ANNA (CONT'D) Ooh! And Sauna... INT. WANDERING OAKEN'S TRADING POST & SAUNA -- NIGHT Anna steps cautiously through the door--which hits her frozen butt and knocks her into the center of the shop. She looks around, sees only summer supplies. OAKEN (O.S.) Hoo hoo. Anna turns to see a bright-faced fellow sitting low behind the counter, fingers tapping tip to tip. OAKEN (CONT'D) Big summer blow out. Half off swimming suits, clogs, and a sun balm of my own invention, yah? ANNA Oh, great. For now, how about boots. Winter boots...and dresses? 40 FROZEN - J. Lee OAKEN (slight disappointment) That would be in our winter department. The winter department contains one outfit, a pick ax, and a lonely pair of boots. ANNA Oh. Um, I was just wondering; has another young woman, the Queen perhaps, I don't know, passed through here? She brings the clothes and boots to the counter. OAKEN Only one crazy enough to be out in this storm is you, dear? The front door suddenly blows open and in walks a mass of a man covered in ice. Underneath is KRISTOFF. OAKEN (CONT'D) You and this fellow.... Hoo hoo. Big summer blow out. Kristoff walks right up to Anna. KRISTOFF (in her face) Carrots. ANNA Huh? KRISTOFF Behind you. ANNA Oh, right. Excuse me. Anna moves out of Kristoff's way. He grabs a bunch of carrots, tosses them on the counter, then moves through the place, gathering other supplies. OAKEN (to Kristoff) A real howler in July, yah? Where ever could it be coming from? KRISTOFF The North Mountain. 41 FROZEN - J. Lee ANNA (to herself) North Mountain. Kristoff brings his supplies to the counter. Oaken counts on his fingertips. OAKEN That'll be forty. KRISTOFF Forty? No, ten. OAKEN (sweet as pie) Oh dear, that's no good. See these are from our winter stock, where supply and demand have a big problem. KRISTOFF You want to talk about a supply and demand problem? I sell ice for a living. Kristoff motions out the window, where we see the blocks of ice on his sled, covered in snow. ANNA Ooh, that's a rough business to be in right now. I mean, that is really... (he shoots her a look) Ahem. That's unfortunate. OAKEN Still forty. But I will throw in a visit to Oaken's sauna. Hoo hoo! Hi, family. Kristoff and Anna turn to see a naked family waving through the window of the steaming sauna. NAKED FAMILY Hoo hoo! KRISTOFF ...Ten's all I got. Help me out. OAKEN (isolating the carrots) Ten will get you this and no more. Kristoff seethes. Stalemate. 42 FROZEN - J. Lee ANNA Okay, just tell me one thing; what was happening on the North Mountain? Did it seem magical? Kristoff pulls down his scarf and gives Anna a firm answer. KRISTOFF Yes! Now, back up while I deal with this crook here. Oaken stands up, revealing his seven-foot stature. OAKEN What did you call me? EXT. WANDERING OAKEN'S TRADING POST AND SAUNA -- NIGHT Oaken stomps out the door, carrying Kristoff with one arm. KRISTOFF Okay. Okay, I'm- Ow! Whoa! Oaken throws Kristoff, who face-plants in the snow. OAKEN Bye bye. Oaken slams the door. Kristoff sits up. His reindeer, Sven, canters over, snorts, and nudges him, expectantly. KRISTOFF No Sven, I didn't get your carrots. Sven huffs in his face. Kristoff turns away and sees something. He points to a dilapidated barn. KRISTOFF (CONT'D) But I did find us a place to sleep. And it's free. INT. WANDERING OAKEN'S TRADING POST AND SAUNA -- NIGHT Anna stands watching Oaken and all his great height as he squeezes behind the counter and sits down low again. OAKEN (teddy bear) I'm sorry about this violence. I will add a quart of lutefisk, so we'll have good feelings. Just the outfit and boots, yah? 43 FROZEN - J. Lee Anna looks between Kristoff's supplies and the door. CUT TO: INT. OAKEN'S STABLES - NIGHT Kristoff, now unfrozen, relaxes on a bed of hay, playing his lute and singing to (and for) Sven. "Reindeer(s) are Better than People" KRISTOFF REINDEERS ARE BETTER THAN PEOPLE. SVEN, DON'T YOU THINK THAT'S TRUE? KRISTOFF (AS SVEN) (CONT'D) (throwing his voice) YEAH, PEOPLE WILL BEAT YOU & CURSE YOU & CHEAT YOU. EVERY ONE OF EM'S BAD, EXCEPT YOU. (speaking) Oh, thanks, Buddy. (singing, as Kristoff) BUT PEOPLE SMELL BETTER THAN REINDEERS. SVEN, DON'T YOU THINK I'M RIGHT? (As Sven) THAT'S ONCE AGAIN TRUE, FOR ALL EXCEPT YOU. (As Kristoff) YOU GOT ME. LET'S CALL IT A NIGHT. (As Sven) GOOD NIGHT. (As Kristoff) DON'T LET THE FROSTBITE BITE. The door opens. Anna enters. ANNA Nice duet. Kristoff sits up with a start...sees who it is. KRISTOFF Oh, it's just you. What do you want? ANNA I want you to take me up the North Mountain. 44 FROZEN - J. Lee KRISTOFF I don't take people places. He lays back down, closes his eyes. ANNA Let me rephrase that... A sack of supplies lands in Kristoff's lap. KRISTOFF Umph. He sits up. Looks in the bag. ANNA Take me up the North Mountain.... Please. He eyes her. He clearly doesn't take orders. ANNA (CONT'D) Look, I know how to stop this winter. He considers, lies back down, pulls his hat over his eyes. KRISTOFF We leave at dawn.... And you forgot the carrots for Sven. A bag of carrots hits Kristoff in the face. KRISTOFF (CONT'D) Ugh! ANNA Oops. Sorry. Sorry. I'm sorry. I didn't-- (catching herself) We leave now. Right now. She steps back outside and waits, anxiously. Annoyed, Kristoff offers Sven a carrot. Sven has a bite. Then Kristoff has a bite, contemplating. SLAM CUT TO: EXT. MOUNTAIN HIGH -- NIGHT Sven races, top speed, up a narrow cliff, pulling the sled, which skids precariously. Kristoff mans the reins. Anna sits beside him. 45 FROZEN - J. Lee KRISTOFF (trying to scare Anna) Hang on! We like to go fast! ANNA (fearless) I like fast! Anna leans back and puts her feet up on the dashboard. KRISTOFF Whoa, whoa! Get your feet down. He pushes her feet down. KRISTOFF (CONT'D) This is fresh lacquer. Seriously, were you raised in a barn? Kristoff spits on the dash to clean it. The spit flies back and hits Anna in the face. ANNA (grossed out) Ew. No, I was raised in a castle. She wipes off her face. KRISTOFF So tell me, what made the Queen go all ice-crazy? ANNA ...Oh well, it was all my fault. I got engaged but then she freaked out because I'd only just met him, you know, that day. And she said she wouldn't bless the marriage-- KRISTOFF Wait. You got engaged to someone you just met? ANNA Yeah. Anyway, I got mad and so she got mad and then she tried to walk away, and I grabbed her glove-- KRISTOFF Hang on. You mean to tell me you got engaged to someone you just met?! 46 FROZEN - J. Lee ANNA Yes. Pay attention. But the thing is she wore the gloves all the time, so I just thought, maybe she has a thing about dirt. KRISTOFF Didn't your parents ever warn you about strangers? Anna eyes Kristoff up and down, then slides away from him. ANNA Yes, they did.... But Hans is not a stranger. KRISTOFF Oh yeah? What's his last name? ANNA ...Of-the-Southern-Isles? KRISTOFF What's his favorite food? ANNA ...Sandwiches. KRISTOFF Best friend's name? ANNA Probably John. KRISTOFF Eye color. ANNA Dreamy. KRISTOFF Foot size...? ANNA Foot size doesn't matter. KRISTOFF Have you had a meal with him yet? What if you hate the way he eats? What if you hate the way he picks his nose? ANNA Picks his nose? 47 FROZEN - J. Lee KRISTOFF And eats it. ANNA Excuse me, sir. He's a prince. KRISTOFF All men do it. ANNA Ew. Look it doesn't matter; it's true love. KRISTOFF Doesn't sound like true love. ANNA Are you some sort of love expert? KRISTOFF No. But I have friends who are. ANNA You have friends who are love experts.... I'm not buying it. Sven suddenly stops, ears perked in alarm. KRISTOFF (to Anna) Stop talking. ANNA No, no, no. I'd like to meet these-- Kristoff clamps his hand over Anna's mouth. KRISTOFF I mean it. SHHH. Kristoff stands, looks into the dark woods surrounding them. Sensing something behind them, he holds up his lantern. Its light reflects off...EYES. Several. KRISTOFF(CONT'D) Sven, go. Go! Sven takes off. ANNA What are they? KRISTOFF Wolves. 48 FROZEN - J. Lee Flashes of white dart through the woods. Kristoff hops into the back of the sled, grabs a torch. Lights it. ANNA Wolves. What do we do? KRISTOFF I've got this. You just...don't fall off and don't get eaten. ANNA But I wanna help. KRISTOFF No. ANNA Why not? KRISTOFF Because I don't trust your judgement. ANNA Excuse me?! A wolf jumps at them, but Kristoff kicks it off. KRISTOFF Who marries a man she just met? Anna grabs the lute, swings it right at Kristoff's head. ANNA It's true love! He screams, as she...BAM!...swings past Kristoff and knocks a wolf away. KRISTOFF (shocked) Whoa. Just then Kristoff is yanked off the sled by another wolf. The torch goes flying. Anna catches it, shocked. ANNA Christopher! Kristoff grabs onto a loose rope hanging from the back of the sled and holds on for dear life as he's dragged behind. KRISTOFF It's Kristoff! 49 FROZEN - J. Lee A wolf jumps on Kristoff's back. KRISTOFF (CONT'D) AH! Anna thinks fast, uses the torch to light a blanket on fire. ANNA Duck! Anna throws the flaming blanket right at him. He ducks. The blanket hits the wolves. They tumble off Kristoff. KRISTOFF You almost set me on fire! Anna reaches out a hand, pulls Kristoff back onto the sled. ANNA But I didn't. Sven cries out. There is a massive gorge ahead. ANNA (CONT'D) Get ready to jump, Sven! KRISTOFF You don't tell him what to do! Kristoff shoves a satchel into her arms then scoops her up. KRISTOFF (CONT'D) I do! Kristoff tosses Anna onto Sven, then unhooks Sven's harness from the sled. KRISTOFF (CONT'D) Jump, Sven! Sven jumps the gorge with Anna on his back. Kristoff goes flying off behind them, still on the sled. Anna and Sven land safely on the other side of the gorge. Kristoff's sled loses momentum. It's not going to make it. He leaps off. He flaps his arms, claws at the air. He slams into the snowy edge of the cliff. Hanging by his hands, he looks down to see his sled hit the ground far below and burst into flames. 50 FROZEN - J. Lee KRISTOFF (CONT'D) (shocked sadness) ...But I just paid it off. Suddenly, he starts to slip. He claws at the loose snow, but it's clearly hopeless. He's going down. KRISTOFF (CONT'D) Uh-oh. No, no, no. To make matters worse, an AXE comes flying right at his face. KRISTOFF (CONT'D) AH! NO, NO, NO! The axe slams into the snow, inches from his nose. ANNA (O.S.) Grab on! Kristoff grabs on. ANNA (CONT'D) Pull, Sven! Pull! REVEAL: The axe is tied to a rope, then wrapped around Sven. Anna helps Sven pull Kristoff to safety. Kristoff rolls onto his back, exhausted. Anna peeks down at the burning sled. ANNA (CONT'D) Whoa.... I'll replace your sled and everything in it. Kristoff groans. ANNA (CONT'D) And I understand if you don't want to help me anymore. Anna walks off, sadly. Sven comes over and nuzzles Kristoff. KRISTOFF Of course I don't want to help her anymore. In fact, this whole thing has ruined me for helping anyone ever again. KRISTOFF (AS SVEN) (CONT'D) But she'll die on her own. KRISTOFF (AS SELF) (CONT'D) I can live with that. 51 FROZEN - J. Lee Through their conversation, they watch Anna go the wrong way...turn, go the other wrong way, turn, trip... KRISTOFF (AS SVEN) (CONT'D) But you won't get your new sled if she's dead. KRISTOFF (CONT'D) (knowing he's got a point) ...You know sometimes I really don't like you. Sven licks Kristoff happily. KRISTOFF (AS SELF) (CONT'D) (to Anna) Hold up. We're coming?! ANNA (excited) You are?! (catching herself) I mean, sure. I'll let you tag along. DISSOLVE TO: EXT. SHARP MOUNTAIN RIDGE -- DAWN Kristoff, Sven and Anna walk on a narrow rim of a mountain. DISSOLVE TO: EXT. MOUNTAIN FOREST CLEARING -- DAY As they step out of the thick trees, Anna catches sight of something far below. ANNA Arendelle. KRISTOFF It's completely frozen. ANNA ...But it'll be fine. Elsa will thaw it. KRISTOFF Will she? 52 FROZEN - J. Lee ANNA (uncertain) ...Yeah. Now come on. This way to the North Mountain? She points straight ahead. KRISTOFF More like this way. He points her finger up towards a perilously mighty mountain. DISSOLVE TO: INT. FROZEN WILLOW TREES -- DAY Anna, Kristoff, and Sven walk beneath frozen willows. The hanging branches glisten like Christmas lights. Sven knocks them with his antlers. They tinkle like chimes. ANNA I never knew winter could be so beautiful. Suddenly, a voice comes in from nowhere. We'll call that voice OLAF. OLAF (O.S.) YEAH...It really is beautiful, isn't it? But it's so white. You know, how about a little color? Must we bleach the joy out of it all? I'm thinking like maybe some crimson, chartreuse... While this is going on, Anna and Kristoff look around for the source of the rambling. They look at Sven - could he actually be talking? Sven looks back at them, his antlers tangled in branches, just as baffled as they are. In the meantime, a nose-less snowman, Olaf, wanders up behind them. OLAF (CONT'D) How `bout yellow--no, not yellow. Yellow and snow? Brrrr...no go. He stops between Kristoff and Anna. They look down at him. How did he get there? He suddenly looks up at Anna. OLAF (CONT'D) Am I right? 53 FROZEN - J. Lee Anna SCREAMS! Reflexes take over and she kicks Olaf's head, sending it flying off his body and into Kristoff's arms. OLAF (CONT'D) (cheery, to Kristoff) Hi! KRISTOFF You're creepy. Kristoff tosses the head back to Anna and they commence a game of hot potato. ANNA I don't want it! KRISTOFF Backatchya! OLAF Please don't drop me. ANNA Don't! KRISTOFF Come on, it's just a head. ANNA No! Olaf's body runs at Anna, arms waving. OLAF (O.S.) All right, we got off to a bad start. ANNA Ew, ew, the body! Anna slams Olaf's head back on the body, upside down. Olaf smiles happily, then looks confused. OLAF Wait, what am I looking at right now? Why are you hanging off the earth like a bat? ANNA (sympathetic) ...Okay. Wait one second. Anna kneels in front of Olaf and rights his head. 54 FROZEN - J. Lee OLAF Oooh! Thank you! ANNA You're welcome. OLAF Now I'm perfect. She looks over his innocent face, gets an idea. ANNA Well, almost. She digs into Kristoff's satchel, holds up a carrot just as Olaf turns toward her. The carrot accidentally slams all the way through his head. OLAF Woo! Head rush! ANNA Oh! Too hard. I'm sorry! I-I, I was just.... Are you okay? Olaf sees a tiny piece of carrot sticking out between his eyes. He lights up. OLAF Are you kidding me? I am wonderful! I've always wanted a nose. (going cross-eyed to look at his tiny nose) So cute. It's like a little baby unicorn. Anna reaches behind Olaf to the bulk of the carrot sticking out the back of his head, and pushes it forward. OLAF (CONT'D) What? Hey! Whoa. (seeing his now big nose) Oh, I love it even more! Hah.... All right, let's start this thing over. Hi everyone. I'm Olaf. And I like warm hugs. Olaf opens his arms wide to Anna. That triggers a memory. It takes her a moment to place it, but then she does. ANNA Olaf?...That's right, Olaf. 55 FROZEN - J. Lee OLAF ...And you are? ANNA Oh, um...I'm Anna. OLAF And who's the funky-looking donkey over there? ANNA That's Sven. OLAF Uh-huh. And who's the reindeer? ANNA ...Sven. Olaf looks from Kristoff to Sven, confused. OLAF Oh. They're--oh, okay.... (accepting it) Makes things easier for me. Sven tries to bite Olaf's nose. OLAF (CONT'D) Ha. Aw, look at him tryin' to kiss my nose. (gushes) I like you, too! ANNA Olaf, did Elsa build you? OLAF Yeah. Why? Curious, Kristoff takes one of Olaf's twig arms off, studies it. It seems to be moving in sync with his other arm. ANNA Do you know where she is? KRISTOFF (studying the arm) Fascinating... OLAF Yeah. Why? 56 FROZEN - J. Lee ANNA Do you think you could show us the way? OLAF Yeah. Why? KRISTOFF (bending the arm) How does this work? Olaf's dismembered arm slaps Kristoff across the face. OLAF Stop it, Sven. Trying to focus here. (to Anna) Yeah, why? KRISTOFF I'll tell you why. We need Elsa to bring back summer. OLAF (shocked) Summer? (sinking into wistfulness) Oh, I don't know why but I've always loved the idea of summer, and sun, and all things hot. KRISTOFF Really? I'm guessing you don't have much experience with heat. OLAF Nope. But sometimes I like to close my eyes and imagine what it'd be like when summer does come. DISSOLVE TO: OLAF'S FANTASY WORLD -- PERFECT SUMMER DAY Olaf walks through a grassy meadow with the sun shining behind him. He SINGS. "In Summer" OLAF BEES'LL BUZZ / KIDS'LL BLOW DANDELION FUZZ / AND I'LL BE DOING WHATEVER SNOW DOES IN SUMMER. 57 FROZEN - J. Lee -Olaf now lies in the sand on a beach. OLAF (CONT'D) A DRINK IN MY HAND / MY SNOW UP AGAINST THE BURNING SAND / PROB'LY GETTING GORGEOUSLY TANNED IN SUMMER. -Olaf sails in a boat. OLAF (CONT'D) I'LL FINALLY SEE A SUMMER BREEZE / BLOW AWAY A WINTER STORM / -Olaf floats in the water. All his pieces begin to separate. OLAF (CONT'D) AND FIND OUT WHAT HAPPENS TO SOLID WATER / WHEN IT GETS WARM. -Olaf tumbles on a sandy beach with sand-snowmen. OLAF (CONT'D) AND I CAN'T WAIT TO SEE / WHAT MY BUDDIES ALL THINK OF ME / JUST IMAGINE HOW MUCH COOLER I'LL BE IN SUMMER . . ! -Olaf and the seagull break out into a tap-dance. OLAF (CONT'D) DA DA . . . DA DOO / AH BAH BAH BAH BAH BAH BOO. -Olaf and another snowman drink hot chocolate in a hot tub. OLAF (CONT'D) THE HOT AND THE COLD ARE BOTH SO INTENSE / PUT `EM TOGETHER, IT JUST MAKES SENSE! -Olaf tap dances with a gaggle of seagulls. OLAF (CONT'D) RATDADAT DAD DADA DOO . . . -Olaf bounds down a grassy hill. OLAF (CONT'D) WINTER'S A GOOD TIME TO STAY IN AND CUDDLE / BUT PUT ME IN SUMMER AND I'LL BE A... He stops at a puddle, looks down at it. Smiles. Hops over it. 58 FROZEN - J. Lee OLAF (CONT'D) HAPPY SNOWMAN! -Olaf runs with a checkered blanket that he spreads out. He relaxes and stares at the blue sky. OLAF (CONT'D) WHEN LIFE GETS ROUGH I LIKE TO HOLD ON TO MY DREAM / OF RELAXING IN THE SUMMER SUN JUST LETTING OFF STEAM! Sven, Anna, Kristoff and Olaf have a picnic. OLAF (CONT'D) OH THE SKY WILL BE BLUE / AND YOU GUYS'LL BE THERE TOO / WHEN I FINALLY DO WHAT FROZEN THINGS DO IN SUMMER! KRISTOFF I'm gonna tell him. ANNA Don't you dare. OLAF IN SUMMER! Olaf sings the final note. We swing around him and return to: REALITY. He then straightens up and smiles. OLAF (CONT'D) So, come on! Elsa's this way. Let's go bring back summer! Olaf grabs Anna's hand and pulls her along up the mountain. ANNA (laughing) I'm coming! Sven hops along, happily following them. Kristoff watches all of them like they're nuts. KRISTOFF Somebody's got to tell him. DISSOLVE TO: 59 FROZEN - J. Lee EXT. ARENDELLE, VILLAGE -- DAY A layer of solid ice coats everything. People huddle around weak fires. Anxiety runs high amongst the villagers and guests. We pass two CITIZENS fighting over a woodpile. CITIZEN ONE No. No. You've got the bark facing down. The bark needs to be face-up. CITIZEN TWO Bark down is drier. CITIZEN ONE Bark up. CITIZEN TWO Bark down. CITIZEN ONE Bark up. Like a light in the dark, Hans moves through the crowd. HANS Cloak. Does anyone need a cloak? GERDA Arendelle is indebted to you, Your Highness. HANS The castle is open. There's soup and hot glogg in the Great Hall. He hands the stack of cloaks to a guard. HANS (CONT'D) Here. Pass these out. Just then the Duke approaches Hans. DUKE Prince Hans, are we just expected to sit here and freeze while you give away all of Arendelle's tradable goods? HANS (tall and confident) Princess Anna has given her orders and-- 60 FROZEN - J. Lee DUKE And that's another thing; has it dawned on you that your princess may be conspiring with a wicked sorceress to destroy us all? Hans's nice eyes turn to threatening slits. HANS Do not question the Princess. She left me in charge, and I will not hesitate to protect Arendelle from treason. DUKE (flabbergasted, offended) Treason?! Suddenly they hear the alarmed whinny of Anna's horse. It returns alone, bucking and kicking. Hans grabs its reins. HANS Whoa! Whoa! Whoa, boy. Easy. Easy. CROWD (various) Princess Anna's horse. What happened to her? Where is she? Hans steadies the horse, looks up at the mountain. He sees all the panicked faces of the kingdom looking to him. HANS ...Princess Anna is in trouble. (calling out) I need volunteers to go with me to find her! Volunteers, some from Arendelle, some from other lands, rush up to offer their services. DUKE I volunteer two men, my Lord! (quietly to his thugs) Be prepared for anything, and should you encounter the Queen, you are to put an end to this winter. Do you understand? His two thugs sneer. CUT TO: 61 FROZEN - J. Lee EXT. THE NORTH MOUNTAIN -- DAY Anna, Kristoff, Sven, and Olaf move through hostile terrain. Wind-swept icicles face horizontal. KRISTOFF So how exactly are you planning to stop this weather? ANNA (confident) Oh, I am gonna talk to my sister. KRISTOFF That's your plan? My ice business is riding on you talking to your sister. ANNA Yup. Kristoff, so stunned by her casual plan, doesn't look where he's going and ends up with an ice-spike to the nose. He stops short, GULP, moves carefully around the spike. KRISTOFF So you're not at all afraid of her? ANNA Why would I be? OLAF (oblivious) Yeah. I bet Elsa's the nicest, gentlest, warmest person ever. Olaf backs right into an icicle. It runs through his torso. OLAF (CONT'D) Oh, look at that. I've been impaled. He laughs it off. DISSOLVE TO: EXT. STEEP MOUNTAIN FACE -- DAY Anna and Kristoff hit what looks like a dead end. The face of the mountain goes straight up. ANNA What now? 62 FROZEN - J. Lee Kristoff looks around, sighs. Digs in his rucksack. KRISTOFF ...It's too steep. I've only got one rope, and you don't know how to climb mountains. ANNA (O.S.) Says who? Sven nudges Kristoff, who looks up to see Anna trying to climb the cliff's flat face. KRISTOFF (finding her ridiculous) What are you doing? ANNA (straining) ...I'm going to see my sister. KRISTOFF You're going to kill yourself. Kristoff watches her searching for footholds and hand-holds. KRISTOFF (CONT'D) I wouldn't put my foot there. ANNA (O.S.) You're distracting me. KRISTOFF Or there. How do you know Elsa even wants to see you? ANNA (O.S.) I'm just blocking you out cause I gotta concentrate here. KRISTOFF You know, most people who disappear into the mountains want to be alone. ANNA (O.S.) Nobody wants to be alone. Except maybe you-- KRISTOFF I'm not alone.... I have friends, remember? Anna kicks a foot above her head to catch a foot hold. 63 FROZEN - J. Lee ANNA You mean the love experts? KRISTOFF Yes, the love experts! Anna realizes she's stuck. ANNA ...Please tell me I'm almost there. REVEAL: she's only about six feet up. Her muscles shake. ANNA (CONT'D) ...Does the air seem a bit thin to you up here? Kristoff smiles, getting a kick out of her. KRISTOFF Hang on. He pulls the rope from his bag. Just then Olaf steps out from behind a rock and waves to Kristoff. OLAF Hey, Sven? Not sure if this is going to solve the problem, but I found a staircase that leads exactly where you want it to go. ANNA Ha ha. Thank goodness. Catch! Anna drops off the cliff. Kristoff catches her. ANNA (CONT'D) Thanks! That was like a crazy trust exercise. She hops down, brushes off her dress, and bounds off. Kristoff watches after her, digging her fearless pluck. EXT. BASE OF THE ICE PALACE -- DAY Anna, Kristoff, and Olaf approach Elsa's elegant ice palace. ANNA Whoa. KRISTOFF (in awe) Now that's ice. I might cry. 64 FROZEN - J. Lee ANNA Go ahead. I won't judge. Anna climbs the steps with Olaf. Sven tries to follow. His hooves slip out. He scrambles but can't get traction. Kristoff runs to his aide. KRISTOFF All right, take it easy. I gotcha. Kristoff settles Sven back down the stairs and pats him. KRISTOFF (CONT'D) You stay right here, buddy. Sven obediently plops his reindeer butt down and wags his tail. Kristoff climbs the stairs, admiring the ice details. KRISTOFF (CONT'D) ...Flawless. Anna arrives at the door. Hesitates. OLAF ...Knock.... (she doesn't) Just knock.... (she doesn't. To Kristoff) Why isn't she knocking...? Do you think she knows how to knock? Anna finally KNOCKS. The sound echoes inside. The ice doors slide open. ANNA Ha. It opened. That's a first. Anna goes to step in. Kristoff follows. She gets a thought, stops him. ANNA (CONT'D) You should probably wait out here. KRISTOFF What? ANNA Last time I introduced her to a guy, she froze everything. KRISTOFF But, it's a palace made of ice. Ice is my life. 65 FROZEN - J. Lee OLAF Bye, Sven. Olaf starts to head inside. Anna stops him. ANNA You too, Olaf. OLAF Me? ANNA Just give us a minute. OLAF Okay. As Anna walks inside. Olaf starts counting. OLAF (CONT'D) One...two... Kristoff joins in. OLAF AND KRISTOFF Three...four... INT. ELSA'S PALACE -- DAY Anna walks into a great foyer. The place is beautiful, but also eerie. ANNA Elsa? It's me...Anna?! Anna slips. Steadies herself. ELSA (O.S.) Anna. Elsa steps out of the shadows onto a balcony. She sees Anna, looks to her longingly. Anna can't help but be struck by Elsa's beauty. ANNA Elsa, you look different.... It's a good different.... And this place is amazing. 66 FROZEN - J. Lee ELSA (cautious, polite) Thank you, I never knew what I was capable of. Anna starts to climb the stairs. ANNA ...I'm so sorry about what happened. If I'd known-- Elsa backs up, away from Anna. ELSA (on guard) No, it's okay. You don't have to apologize.... But you should probably go, please. ANNA But I just got here. ELSA ...You belong in Arendelle. ANNA So do you. Anna takes another step up. Elsa backs up more. ELSA No, I belong here. Alone. Where I can be who I am without hurting anybody. ANNA ...Actually, about that-- OLAF (O.S.) 58...59...60. ELSA Wait. What is that? Olaf comes running in the front door. He waves. OLAF Hi, I'm Olaf and I like warm hugs. ELSA (shocked) Olaf? Olaf stops beside Anna, looks up at Elsa, intimidated. 67 FROZEN - J. Lee OLAF (bashful) You built me. You remember that? ELSA (astonished) And you're alive? OLAF Um...I think so? Anna kneels down beside Olaf. ANNA He's just like the one we built as kids.... We were so close. We can be like that again. Elsa smiles, but then a memory returns to her. FLASH CUT TO: FLASHBACK: Young Anna is struck by Elsa's powers. YOUNG ELSA Anna! Young Anna falls unconscious. Young Elsa races to her. FLASH CUT TO: THE PRESENT: Elsa's face sinks in pain. ELSA No, we can't. Elsa turns and heads up the second story steps. ELSA (CONT'D) Goodbye, Anna. ANNA Elsa, wait-- ELSA (calling back) I'm just trying to protect you. Elsa continues to flee. Anna pursues. ANNA You don't have to protect me. I'm not afraid. Please don't shut me out again. 68 FROZEN - J. Lee Anna SINGS. "First Time in Forever, Reprise" ANNA (CONT'D) PLEASE DON'T SLAM THE DOOR. YOU DON'T HAVE TO KEEP YOUR DISTANCE ANYMORE. `CAUSE FOR THE FIRST TIME IN FOREVER, I FINALLY UNDERSTAND. FOR THE FIRST TIME IN FOREVER, WE CAN FIX THIS HAND IN HAND. WE CAN HEAD DOWN THIS MOUNTAIN TOGETHER. YOU DON'T HAVE TO LIVE IN FEAR. `CAUSE FOR THE FIRST TIME IN FOREVER, I WILL BE RIGHT HERE. They arrive on the top floor, Elsa's main living space. Elsa turns back to Anna, grateful, but determined. ELSA Anna, PLEASE GO BACK HOME. YOUR LIFE AWAITS. GO ENJOY THE SUN AND OPEN UP THE GATES. ANNA Yeah, but-- ELSA I know! YOU MEAN WELL, BUT LEAVE ME BE. YES, I'M ALONE BUT I'M ALONE AND FREE. Elsa opens up the balcony doors. ELSA (CONT'D) JUST STAY AWAY AND YOU'LL BE SAFE FROM ME. ANNA ACTUALLY, WE'RE NOT. ELSA WHAT DO YOU MEAN YOU'RE NOT? 69 FROZEN - J. Lee ANNA I GET THE FEELING YOU DON'T KNOW? ELSA WHAT DO I NOT KNOW? ANNA ARENDELLE'S IN DEEP DEEP DEEP DEEP SNOW. ELSA What? Elsa looks past Anna's shoulder out white-peaked mountains. ANNA You kind of set off an eternal winter...everywhere. ELSA Everywhere? ANNA It's okay, you can just unfreeze it. ELSA No, I can't. I don't know how. ANNA Sure you can. I know you can. Snow starts to swirl around the room. ANNA (CONT'D) CUZ FOR THE FIRST TIME IN FOREVER, ELSA (panicking) I'M SUCH A FOOL! I CAN'T BE FREE! ANNA YOU DON'T HAVE TO BE AFRAID. ELSA NO ESCAPE FROM THE STORM INSIDE OF ME! The snow picks up. Anna tries to fight through it. ANNA WE CAN WORK THIS OUT TOGETHER. 70 FROZEN - J. Lee ELSA I CAN'T CONTROL THE CURSE! ANNA WE'LL REVERSE THE STORM YOU'VE MADE. ELSA ANNA, PLEASE, YOU'LL ONLY MAKE IT WORSE! ANNA DON'T PANIC. ELSA THERE'S SO MUCH FEAR! ANNA WE'LL MAKE THE SUN SHINE BRIGHT. ELSA YOU'RE NOT SAFE HERE! ANNA WE CAN FACE THIS THING TOGETHER... But as Anna sings, we lose sight of her in the thickening blizzard taking over the room. ELSA NO! ANNA (O.S.) WE CAN CHANGE THIS WINTER WEATHER, AND EVERYTHING WILL BE... Anna's voice disappears in the storm as Elsa cries out. ELSA I CAN'T! Elsa's fear, so strong, sucks the blizzard back into her and then it bursts out, unwittingly, like a sharp snowflake. Anna is STRUCK right in the heart. She grasps her chest in pain and stumbles back. She falls to her knees. Elsa gasps when she sees Anna. Just then, Olaf and Kristoff rush into the room to Anna's side. KRISTOFF Anna. Are you okay? 71 FROZEN - J. Lee ANNA I'm okay.... I'm fine. Anna gets to her feet, determined to hide the pain. ELSA (scared) Who's this? Wait, it doesn't matter. You have to go. ANNA No, I know we can figure this out together-- ELSA (desperate) How? What power do you have to stop this winter? To stop me? Anna doesn't have the answer. Kristoff sees spiky ice shadows creeping down the walls. Puts a protective arm around Anna. KRISTOFF Anna, I think we should go. ANNA (close to tears) No. I'm not leaving without you, Elsa. ELSA (heartbroken but decisive) Yes, you are. Elsa waves her arms and builds a giant, menacing snowman. We'll call him MARSHMALLOW. SLAM CUT TO: EXT. ICE PALACE -- DAY Marshmallow holds Anna and Kristoff by the scruff of their necks in one hand and Olaf in the other. ANNA Stop. Put us down! OLAF (to Marshmallow) You are a lot stronger than I think you realize. Marshmallow tosses Kristoff and Anna down the steps. 72 FROZEN - J. Lee MARSHMALLOW (like a bouncer) Go away! Anna and Kistoff slide past Sven, who's got his tongue stuck to the ice railing. OLAF (O.S.) Heads up! Olaf's head smashes into a snowbank nearby. ANNA Olaf! OLAF Watch out for my butt! Anna and Kristoff duck as the rest of Olaf slams into the snowbank. Marshmallow turns to go back into the castle. Incensed, Anna tries to march back up the stairs. ANNA It is not nice to throw people! Kristoff grabs her, pulls her back. KRISTOFF ANNA All right feisty pants. Calm Let me at him. I want to get down. Woaw. Just let the snow him. I.... Okay. I'm Calm. man be. Anna backs down...for a moment. Then she grabs a snowball and throws it at Marshmallow. The tiny little ball hits Marshmallow's back, not making even the slightest dent. But it's enough to infuriate him. He ROARS. Spikes shoot out of his joints. KRISTOFF Uh-oh. Now you made him mad! OLAF ...I'll distract him. You guys go. Kristoff pushes Anna along. Sven runs off in the opposite direction. Olaf's belly and butt fall and follow Sven. OLAF (CONT'D) No, no, not you guys. 73 FROZEN - J. Lee Marshmallow goes charging after Anna and Kristoff as Olaf's head falls and lands face down in snow. OLAF (CONT'D) (muffled) This just got a whole lot harder. Anna and Kristoff leap and slide down a steep slope. They tumble to a stop at the bottom just as Marshmallow lands hard right behind them. They're off again...through a maze of conifers that sag under the weight of the snow, Marshmallow hot on their trail. KRISTOFF This way! Anna grabs a branch of a sagging trees and releases all of the snow. The tree snaps upright, knocking Marshmallow back. KRISTOFF (CONT'D) (impressed) Ho-ho-ho! ANNA I got him! Anna and Kristoff burst out of the conifer forest and almost run right off a cliff. They stop short, toes on the edge. KRISTOFF Whoa, stop! ANNA It's a hundred foot drop. KRISTOFF It's two hundred. Kristoff ties the rope around Anna and pulls tight. ANNA Ow. He drops to his knees and starts digging a u-shape in the snow with a pick axe. ANNA (CONT'D) What's that for? KRISTOFF I'm digging a snow anchor. 74 FROZEN - J. Lee ANNA (not trusting) Okay. What if we fall? KRISTOFF There's twenty feet of fresh powder down there; it'll be like landing on a pillow.... Hopefully. They hear an angry ROAR coming closer. KRISTOFF (CONT'D) Okay, Anna. On three. Anna preps for the jump like a boxer getting ready to fight. ANNA Okay. You tell me when... KRISTOFF One... ANNA ...I'm ready to go.... KRISTOFF Two... ANNA (pumped up) ...I was BORN ready! Yes! KRISTOFF Calm down. A huge tree flies through the air toward them. ANNA (O.S.) TREE! Anna jumps and pulls Kristoff over the edge with her. They hang upside down over the cliff by the rope. The rope catches their fall. KRISTOFF Whoa! That happened. Back up top, Olaf emerges from the woods. He's a complete mess, all his body parts are in the wrong places. He huffs and puffs, struggling to run. OLAF Ah. Ah. Man, am I out of shape. 75 FROZEN - J. Lee He stops. Puts his body back together in the right order. OLAF (CONT'D) There we go. Hey, Anna! Sven! Where'd ya guys go? We totally lost Marshmallow back there! Marshmallow steps up behind Olaf. Olaf turns to face him. OLAF (CONT'D) (happily) Hey. We were just talking about you. All good things, all good things. Marshmallow roars and approaches Kristoff's snow anchor. OLAF (CONT'D) NO! Olaf jumps onto Marshmallow's leg trying to stop him, but not making much of a difference. OLAF (CONT'D) This is not making much of a difference! Marshmallow flicks Olaf off his leg and right over the cliff. OLAF (CONT'D) WHOA! Olaf passes Anna and Kristoff. ANNA Olaf! OLAF Hang in there, guys! Marshmallow starts yanking Kristoff and Anna's rope up. ANNA Wait, what? Kristoff's head hits the cliff. KRISTOFF Aargghh! Kristoff passes out and hangs like a rag doll. ANNA Kristoff! 76 FROZEN - J. Lee Marshmallow pulls them up. He roars and breathes snow all over them. MARSHMALLOW Don't come back! ANNA (grossed out by his snow breath) Ugh. We won't. Anna whips out a knife and cuts the rope. Kristoff comes to just as they fall. They both SCREAM! SLAM! REVEAL: Anna opens her eyes to find herself buried up to her shoulders in the soft thick snow. She laughs. ANNA (CONT'D) Hey, you were right. Just like a pillow. She looks up to see Olaf's upper half hanging onto Kristoff's boots, which are sticking out of the snow. OLAF (shaking the boots) I can't feel my legs! I can't feel my legs! Suddenly, Kristoff's head pops up. He spits out snow. KRISTOFF Those are my legs. Olaf's bottom goes running by. OLAF (to Kristoff) Ooh. Hey, do me a favor, grab my butt. Kristoff grabs Olaf's head and puts it on his body. OLAF (CONT'D) Oh, that feels better. Sven walks up and sniffs Olaf's nose. OLAF (CONT'D) Hey, Sven! 77 FROZEN - J. Lee Olaf turns to Anna and Kristoff just as Sven goes to bite off his nose -- and misses. OLAF (CONT'D) He found us. (to Sven, funny voice) Who's my cute little reindeer? KRISTOFF Don't talk to him like that. Kristoff goes over to help Anna, who is stuck in the snow. KRISTOFF (CONT'D) Here. He lifts her out easily. ANNA (impressed) Whoa! KRISTOFF You okay? ANNA Thank you. They meet eyes. Wait. Is that chemistry? ANNA (CONT'D) ...Um.... How's your head? She touches the spot where he banged his head. KRISTOFF (in pain) Ah! Ooh! He catches himself. Waves off the pain with a giggle. KRISTOFF (CONT'D) I mean, It's fine. Ah...I'm good. Ha. I've got a thick skull. OLAF I don't have a skull.... Or bones. KRISTOFF ...So.... The awkwardness is killing him. 78 FROZEN - J. Lee KRISTOFF (CONT'D) (shy) Now what? ANNA (shy) Now what? (then...panicking) Now what?! Oh! What am I gonna do? She threw me out. I can't go back to Arendelle with the weather like this. And then there's your ice business-- KRISTOFF Hey, hey, don't worry about my ice business... (noticing something) Worry about your hair?! She thinks he means it looks bad. She smooths it down. ANNA What? I just fell off a cliff. You should see your hair. KRISTOFF No, yours is turning white. She grabs her braid as a tendril turns white. ANNA White? It's what? KRISTOFF It's because she struck you; isn't it? ANNA Does it look bad? KRISTOFF (thinking) ...No. Olaf's head pops up. He's holding his head up off his body to join the conversation. OLAF You hesitated. KRISTOFF No, I didn't. Anna, you need help. Now, come on. 79 FROZEN - J. Lee He heads towards the sunset. Sven and Olaf follow. OLAF Okay! Where are we going? KRISTOFF To see my friends. ANNA (catching up) The love experts? OLAF Love experts?! KRISTOFF Yes. And don't worry; they'll be able to fix this. ANNA How do you know? He looks her over, remembering the moment he saw the trolls heal her as a child. KRISTOFF ...Because I've seen them do it before. As they round the bend, the sun sets and Olaf turns to Sven. OLAF I like to consider myself a love expert. CUT TO: INT. ELSA'S PALACE -- DAY Elsa paces, distraught. She talks to herself. ELSA (mantra-style) Get it together. Control it. Don't feel. Don't feel. Don't FEEL! She hears ice cracking. Stops. Looks around. She's left a sharp wake of ice spikes behind her on the floor. They grow up the wall, taking over the castle. DISSOLVE TO: 80 FROZEN - J. Lee EXT. BLACK MOUNTAINS -- NIGHT The Northern Lights are bright. Olaf stares at them in awe as he rides on Sven's back. OLAF Look, Sven. The sky's awake. Behind Olaf and Sven, Anna walks with Kristoff. She shivers. KRISTOFF Are you cold? ANNA ...A little. He reaches like he might put an arm around her, but decides against it. He looks around as if he doesn't know what to do, then gets a thought. KRISTOFF Wait. Come here. He takes her hand and pulls her around a bend into a rock- lined pass. Steam vents, powered by the volcanic activity, dot the path. He holds her hands over one of them. ANNA Oooh.... That's nice. They continue on the path, walking from vent to vent. KRISTOFF (taking a deep breath) So, about my friends...well, I say friends, they're more like family.... Anyway, when I was a kid, it was just me and Sven...until they took me in. ANNA (moved) They did? KRISTOFF (nervous ramble) Yeah. I don't want to scare you, they can be a little bit inappropriate...and loud...very loud...they're also stubborn at times, and a little overbearing. And heavy. Really, really heavy. (MORE) 81 FROZEN - J. Lee KRISTOFF (CONT'D) But they're fine.. You'll get it. They mean well. Anna touches Kristoff's arm, reassuringly. ANNA Kristoff, they sound wonderful. Kristoff smiles, appreciating her sincerity. KRISTOFF Okay then.... Mustering the courage, Kristoff steps forward and with a wave of the arms announces-- KRISTOFF (CONT'D) Meet my family. REVEAL: he's surrounded by rocks. KRISTOFF (CONT'D) (to the rocks) Hey, guys! As Kristoff and Sven move through the rocks, waving and greeting, Olaf and Anna stand frozen, dumbfounded. ANNA (to herself) ...They're rocks. OLAF (realizing) He's crazy. (covertly, to Anna) I'll distract them while you run. (Loud and slow to a rock) Hi, Sven's family! It's nice to meet you! (quietly to Anna) Anna, because I love you, I insist you run. (to the rock) I understand you're love experts! (to Anna) Why aren't you running? Anna snaps out of her shock and starts backing away. ANNA Okay. Um...I'm gonna go-- Just then the rocks around her start rolling. 82 FROZEN - J. Lee ANNA (CONT'D) (panicking) Kristoff! Olaf lights up and chases the rocks, who surround Kristoff and unfold as trolls. BULDA KRISTOFF'S HOME! TROLLS (VARIOUS) Kristoff! Kristoff's home! It's been too long! Kristoff's home! Olaf jumps around all excitedly. OLAF (excitedly) Kristoff's home. He then stops, confused, and looks to one of the trolls. OLAF (CONT'D) Wait? Kristoff? Anna watches, shocked and confused. The trolls all want Kristoff's attention. One troll yanks him down with a boulder's strength. TROLL ONE Oh, lemme look at you! Another troll tries to pull off his clothes. TROLL TWO Oh, take off your clothes, Kristoff; I wash them. KRISTOFF (holding up his pants) Ah! No. I'm gonna keep my clothes on, thank you. KRISTOFF (CONT'D) Great to see you all. Where's grandpa? MUSHROOM KID TROLL He's napping. But look, I grew a mushroom. TROLL SCOUT KID And I earned my fire crystal. 83 FROZEN - J. Lee KIDNEY STONE TROLL I passed a kidney stone. PICK ME UP TROLL Pick me up. The kid troll jumps up on Kristoff's arm. Kristoff sinks under the weight of him. Anna still stares, confused, then realizes... ANNA Trolls? They're trolls. Silence. All troll eyes turn to Anna. Blink. Blink. BULDA ...He's brought a girl! TROLLS (TOGETHER) He's brought a girl! Suddenly Anna is surrounded by trolls. They body-surf/roll Anna over to Kristoff. She falls into his arms. ANNA What's going on? KRISTOFF I've learned to just roll with it. Bulda climbs on top of her husband, Cliff, to get a good look at Anna. She studies her like she's a piece of cattle. BULDA Let me see. Bright eyes. Working nose. Strong teeth. Yes, yes, yes. She'll do nicely for our Kristoff. ANNA Wait. Oh. Um. No. KRISTOFF You've got the wrong idea. That's not why I brought her here. ANNA Right. We're not. I'm not-- Anna laughs, uncomfortable, not knowing what to say. 84 FROZEN - J. Lee BULDA (to Anna) What's the issue, dear? Why are you holding back from such a man? Bulda SINGS. "Fixer-Upper" TROLLS (VARIOUS) IS IT THE CLUMPY WAY HE WALKS? OR THE GRUMPY WAY HE TALKS? OR THE PEAR-SHAPED, SQUARE-SHAPED WEIRDNESS OF HIS FEET? AND THOUGH WE KNOW HE WASHES WELL HE ALWAYS ENDS UP SORTA SMELLY. BUT YOU'LL NEVER MEET A FELLA WHO'S AS SENSITIVE AND SWEET. TROLLS (CHORUS) (CONT'D) SO HE'S A BIT OF A FIXER UPPER, SO HE'S GOT A FEW FLAWS- HIS PECULIAR BRAIN, DEAR. HIS THING FOR THE REINDEER THAT OUTSIDE A FEW OF NATURE'S LAWS. SO HE'S A BIT OF A FIXER UPPER, BUT THIS WE'RE CERTAIN OF- YOU CAN FIX THIS FIXER UPPER UP WITH A LITTLE BIT OF LOVE. KRISTOFF Can we just stop talking about this?! We've got a real, actual problem here. BULDA I'll say-- (To Anna) IS IT THE WAY THAT HE RUNS SCARED? TROLLS (VARIOUS) OR THAT HE'S SOCIALLY IMPAIRED? KID TROLL OR THAT HE ONLY LIKES TO TINKLE IN THE WOODS? TROLLS (VARIOUS) ARE YOU HOLDING BACK YOUR FONDNESS DUE TO HIS UNMANLY BLONDENESS? OR THE WAY HE COVERS UP THAT HE'S THE HONEST GOODS? 85 FROZEN - J. Lee TROLLS (CHORUS) (CONT'D) HE'S JUST A BIT OF A FIXER UPPER- HE'S GOT A COUPLE A' BUGS. KRISTOFF No, I don't. TROLLS HIS ISOLATION IS CONFIRMATION OF HIS DESPERATION FOR HEALING HUGS. SO HE'S A BIT OF A FIXER UPPER, BUT WE KNOW WHAT TO DO. THE WAY TO FIX UP THIS FIXER UPPER IS TO FIX HIM UP WITH YOU. The girl trolls sweep Anna away. The boys take Kristoff. KRISTOFF (to the male trolls) Enough! She's engaged to someone else. Okay?! TROLLS beat. Blink. Blink. The boy trolls turn, huddle... TROLLS (VARIOUS) SO SHE'S A BIT OF A FIXER UPPER, THAT'S A MINOR THING. THIS QUOTE "ENGAGEMENT" IS A FLEX ARRANGEMENT. KID TROLL AND BY THE WAY, I DON'T SEE NO RING. TROLLS (VARIOUS) SO SHE'S A BIT OF A FIXER UPPER, HER BRAIN'S A BIT BETWIXT. GET THE FIANCE OUT OF THE WAY AND THE WHOLE THING WILL BE FIXED! GIRL TROLLS WE AREN'T SAYING YOU CAN CHANGE HIM TROLLS (VARIOUS) 'CAUSE PEOPLE DON'T REALLY CHANGE. WE'RE ONLY SAYING THAT LOVE'S A FORCE THAT'S POWERFUL AND STRANGE. PEOPLE MAKE BAD CHOICES IF THEY'RE MAD OR SCARED OR STRESSED. (MORE) 86 FROZEN - J. Lee TROLLS (VARIOUS) (CONT'D) BUT THROW A LITTLE LOVE THEIR WAY (THROW A LITTLE LOVE THEIR WAY) AND YOU'LL BRING OUT THEIR BEST! TRUE LOVE BRINGS OUT THE BEST! Kristoff looks over at Anna. She actually looks shockingly beautiful dressed in moss, lit by shimmering crystals. ALL TROLLS EVERYONE'S A BIT OF A FIXER UPPER, THAT'S WHAT IT'S ALL ABOUT FATHER, SISTER, BROTHER WE NEED EACH OTHER TO RAISE US UP AND ROUND US OUT By this time Kristoff and Anna are being ushered into a pit by the sheer force of numbers. TROLLS EVERYONE'S A BIT OF A FIXER UPPER, BUT WHEN PUSH COMES TO SHOVE- THE ONLY FIXER UPPER FIXER THAT CAN FIX A FIXER UPPER IS TRUE TRUE TRUE TRUE LOVE During this last bit Anna and Kristoff are looking at each other differently. Hmmm. Maybe those trolls are right? Sparks! Chemistry! TROLL PRIEST Do you, Anna, take Kristoff to be your trollfully wedded-- ANNA Wait, what?! TROLL PRIEST You're getting married. TROLLS LOVE! Just then, Anna collapses. Kristoff catches her. She's shivering something fierce. KRISTOFF Anna? He pulls off her cape and hat. 87 FROZEN - J. Lee KRISTOFF (CONT'D) She's as cold as ice. Just then Grand Pabbie pushes his way through the crowd. Trolls clear the way for Pabbie. He stops at the edge of the pit. GRAND PABBIE There's strange magic here! KRISTOFF Grand Pabbie! GRAND PABBIE Bring her to me, Kristoff. Kristoff helps Anna over. Pabbie looks into her weak eyes. GRAND PABBIE (CONT'D) Anna, your life is in danger. There is ice in your heart, put there by your sister. If not removed, to solid ice will you freeze, forever. ANNA What...? No. KRISTOFF So remove it, Grand Pabbie. GRAND PABBIE I can't. If it was her head, that would be easy. But only an act of true love can thaw a frozen heart. ANNA An act of true love? BULDA (googley, to her hubby) A true love's kiss, perhaps? A bunch of trolls give each other kisses. Anna shivers again, collapsing into Kristoff's arms. More of her hair turns white. KRISTOFF Anna, we've got to get you back to Hans. ANNA (still weak) ...Hans. 88 FROZEN - J. Lee KRISTOFF Help us out, Sven. Kristoff grabs Sven's antlers. Sven pulls them out. Kristoff helps Anna onto Sven and hops up behind her. KRISTOFF (CONT'D) Come on, Olaf! Sven takes off. Olaf grabs Sven's tail, rides with them. OLAF I'm coming! Let's go kiss Hans! Who is this Hans?! CUT TO: EXT. ELSA'S PALACE - DAWN Hans and the men tread cautiously towards the castle. HANS We are here to find Princess Anna. Be on guard, but no harm is to come to the Queen. Do you understand? The Duke's thugs exchange a look. Suddenly, a mass of snow rises from the ground behind Hans. It's Marshmallow, Elsa's snow guard. MARSHMALLOW Go away! He slams a fist inches from Hans. Hans deftly dodges out of the way. All of the guards take up arms against Marshmallow, who quickly knocks them over. Marshmallow throws down a guard and his horse, who topple over Hans. Marshmallow raises his foot to stomp on Hans, but Hans barrel-rolls himself to safety. He sees his sword, leaps, and grabs it. Just then, Elsa peeks out the front doors. The Duke's two thugs see her. DUKE'S THUG The Queen. The thugs charge up the stairs. 89 FROZEN - J. Lee INT. ELSA'S PALACE -- DAY They guards burst through the ice doors. Elsa flees to the top floor of her palace. The guards pursue. They trap her on the top floor, raise their crossbows. ELSA (scared) No. Please. One of the thugs shoots an arrow right at Elsa. At the last moment she creates an ice wall. It stops the arrow, inches from her face. The thugs reposition to take another shot. ELSA (CONT'D) Stay away! Elsa shoots ice at the thugs. They duck out of the way and continue the attack. THUG Get her! Get her! Elsa fights for her life. BACK OUTSIDE: Hans is nearly crushed by Marshmallow. He rolls away. Jumps to his feet. And with agile might, he slices Marshmallow's leg off with his sword. Marshmallow stumbles back, off balance. And falls off over the cliff, but not before striking Hans. Hans goes over the edge. REVEAL: Hans clings to the ice steps. His men help him up and they rush into the ice palace. INT. ICE PALACE -- DAY Elsa is surrounded. It's do or die. In two swift moves, Elsa traps one thug in a cage of spikes that threaten his neck. The other she pushes back with a wall of ice....up against the balcony doors...which BURST and CRACK. OUT ONTO THE BALCONY.... The balcony doors shatter. The thug is pushed to the edge. He's inches away from falling to his death. BACK INSIDE: Hans and his men run in. See the destruction and the thugs near death. 90 FROZEN - J. Lee HANS Queen Elsa! Don't be the monster they fear you are. Elsa snaps out of her rage. She sees the men, frightened, moments from death. She stops. Elsa looks to Hans, overwhelmed, frightened. The wall retreats from the thug on the balcony. The ice spikes lower from the second thug's neck. He takes advantage and aims his crossbow at Elsa's back. Seeing it. Hans runs and pushes the crossbow up just as the arrow releases. The arrow hits the ice chandelier, hanging directly above Elsa. The chandelier comes CRASHING DOWN. Elsa dives out of the way but she falls in the blast. All we see is ice smashing like glass, and all we hear is the sound of it shattering as it rings out. CUT TO BLACK. FADE IN ON: Elsa's face as her eyes flutter open. She sits up. She's surrounded by stone. INT. ARENDELLE, DUNGEON -- DAY Elsa looks to the nearby window. Tries to rush to it. She's pulled taut by giant shackles that fit like iron gloves. She's chained to the wall. Elsa strains to looks out a window... INSET WINDOW: Arendelle is outside, frozen solid and getting further buried under the ice and snow that is falling. ELSA No....What have I done? Hans enters. He hangs a torch by the door. ELSA (CONT'D) Why did you bring me here? HANS I couldn't just let them kill you. 91 FROZEN - J. Lee ELSA But I'm a danger to Arendelle. Get Anna. HANS Anna has not returned.... Elsa looks to the storm with worry. HANS (CONT'D) If you would just stop the winter, bring back summer...please. Elsa meets his eyes, desperate. ELSA Don't you see...I can't. Hans sees the sincerity in her eyes. ELSA (CONT'D) You have to tell them to let me go. Hans walks to the door. He takes the torch. HANS I will do what I can. He opens the door and leaves. Elsa, distraught, hears cracking. She looks down as her shackles begin to freeze over. The storm outside picks up. CUT TO: EXT. THE FJORDS -- DAY Sven charges down the mountain with Kristoff and Anna on his back. Olaf slides along beside them, penguin-style. Anna shivers in Kristoff's arms. She's weakening. Kristoff takes off his hat and puts it on her head. KRISTOFF Just hang in there. (to Sven) Come on, buddy, faster! They arrive at the walls of Arendelle. Olaf slides past them, out of control. OLAF I'll meet you guys at the castle! 92 FROZEN - J. Lee KRISTOFF Stay out of sight, Olaf! OLAF I will! He disappears into the village streets. OLAF (O.S.) (CONT'D) Hello! TOWNSWOMAN (O.S.) Ah! It's alive! CUT TO: EXT. CASTLE COURTYARD -- DAY Guards see Kristoff and Anna approaching. GUARD It's Princess Anna! Sven skids to a stop outside the gates. Kristoff slides off, holding Anna, and carries her to the gate. KRISTOFF I've got you. Anna looks up at him, gratefully. ANNA ...Are you g-gonna be okay? KRISTOFF (touched, reassuring) Don't worry about me. Just then the castle gates open. Gerda, Kai, and a handmaid rush to help Anna. GERDA Anna! Oh, you had us worried sick. KAI My Lady. You are freezing. GERDA You poor girl, you're freezing. Let's get you inside. 93 FROZEN - J. Lee KRISTOFF Get her warm and find Prince Hans, immediately. KAI We will. Thank you. Anna is swept away from Kristoff and into the palace grounds. KRISTOFF Make sure she's safe! Kristoff is shut out as the castle gates close on him. Kristoff stands there with Sven for a beat, staring with worry at the closed gates. Finally, he sighs, turns and walks off. Sven reluctantly follows. CUT TO: INT. LIBRARY -- DAY Hans stands with the dignitaries and guards. HANS I'm going back out to look for Princess Anna. FRENCH DIGNITARY You cannot risk going out there again. HANS If anything happens to her-- SPANISH DIGNITARY If anything happens to the Princess, you are all Arendelle has left. Hans hesitates, realizing how much this kingdom has come to depend on him. Is he really all they have left? Just then the door opens and Gerda and Kai bring in Anna. KAI He's in here. Prince Hans. HANS Anna. 94 FROZEN - J. Lee Hans rushes to Anna. She falls into his arms. HANS (CONT'D) You're so cold. ANNA (weak, but desperate) Hans, you have to kiss me. HANS What? ANNA Now. Here we go. She tries to kiss him, but is too weak to pull herself up in his arms. GERDA We'll give you two some privacy. Everyone shuffles out, leaving Hans and Anna alone. HANS What happened out there? ANNA Elsa struck me with her powers. HANS You said she'd never hurt you. ANNA I was wrong. Anna crumbles, weak. HANS Anna. Hans carries her to a couch, sets her down. ANNA (shivering more) She froze my heart and only an act of true love can save me. HANS (understanding) A true love's kiss. He takes her chin in his hand and gives her a tender smile. He leans in slowly...gently... 95 FROZEN - J. Lee Then he stops. HANS (CONT'D) Oh, Anna. If only there was someone out there who loved you. ANNA What? Hans gets up, leaving her there. ANNA (CONT'D) ...You said you did. He goes to the window and shuts the curtains. HANS As thirteenth in line in my own kingdom, I didn't stand a chance. I knew I'd have to marry into the throne somewhere-- ANNA What are you talking about? HANS (putting out the candles) As heir, Elsa was preferable, of course. But no one was getting anywhere with her. But you- ANNA Hans? HANS You were so desperate for love you were willing to marry me, just like that. Hans crosses the room, grabs a pitcher of water from a table and goes to the fireplace. HANS (CONT'D) I figured, after we married, I'd have to stage a little accident for Elsa. Hans pours the water on the fireplace, putting out the fire. Anna tries to stop him. She falls to the floor, weak. ANNA Hans. No, stop. 96 FROZEN - J. Lee HANS But then she doomed herself, and you were dumb enough to go after her. ANNA Please. HANS (chuckles) All that's left now is to kill Elsa and bring back summer. Hans approaches Anna. ANNA ...You're no match for Elsa. He bends down, takes her chin in his hand again, this time not so gently. HANS No, you're no match for Elsa. I, on the other hand, am the hero who is going to save Arendelle from destruction. She wrenches her face out of his hands. ANNA (anger) You won't get away with this. Hans rises and crosses to the door. HANS Oh, I already have. Hans leaves and shuts her in, locking the door. Anna struggles to the door, yanks on the locked handle. ANNA (hoarse and weak) Please, somebody help. The rest of her hair turns white and she crumbles to the floor. CUT TO: 97 FROZEN - J. Lee INT. COUNCIL CHAMBER -- NIGHT The Duke looks out the window at the growing snowstorm. He rubs his arms and shivers. DUKE It's getting colder by the minute. If we don't do something soon, we'll all freeze to death. Hans comes in, putting on his most distraught face. SPANISH DIGNITARY Prince Hans. HANS Princess Anna is...dead. VARIOUS DIGNITARIES What...? No.... Mon dieu. Hans stumbles, weak with grief. The men help him to a chair. DUKE What happened to her? HANS She was killed by Queen Elsa. DUKE Her own sister. HANS (really putting it on) At least we got to say our marriage vows...before she died in my arms. He bows his head in a brilliant display of teary grief. DUKE There can be no doubt now; Queen Elsa is a monster and we are all in grave danger. SPANISH DIGNITARY Prince Hans, Arendelle looks to you. Hans nods; he knows what he's being asked to do, and he'll do it with the perfect amount of authority and gravitas. 98 FROZEN - J. Lee HANS With a heavy heart, I charge Queen Elsa of Arendelle with treason and sentence her to death. INT. ELSA'S DUNGEON -- DAY The cell ices over. Elsa looks out at the storm that is devastating Arendelle, then hears the guards approaching. GUARD (O.S.) She's dangerous. Move quickly and with resolve. Elsa pulls at her shackles. They crack. Just as the door busts open, the weight of the ice crumbles the walls. The men duck out of the way. Hans pushes his way into the room...sees... The back wall is blown open. Broken shackles rest on the floor. Elsa is gone. CUT TO: EXT. MOUNTAIN SLOPE -- DAY Kristoff heads into the mountains. Sven lags behind, not wanting to follow. He looks back at the kingdom, then shakes his head. Enough. He runs past Kristoff. Stops and turns to face him. He snorts and grunts. KRISTOFF What is it, buddy? Sven nudges Kristoff with his antlers. KRISTOFF (CONT'D) Hey, watch it. What's wrong with you? Sven snorts with more conviction, moos, brays. KRISTOFF (CONT'D) (avoiding) ...I don't understand you when you talk like that. 99 FROZEN - J. Lee Kristoff tries to walk on ahead, but Sven uses his antlers to lift Kristoff off the ground. KRISTOFF (CONT'D) Ah! Stop it! Put me down! Sven drops him hard then "yells" at him once more. KRISTOFF (CONT'D) No, Sven! We're not going back! Sven shakes his head, angrily. KRISTOFF (CONT'D) She's with her true love. Sven makes an "of-course-she-isn't" face. Kristoff gets it; he's made his point. Just then the wind picks up. Kristoff looks back at the kingdom. Sees a violent winter storm swirling over the castle. Sharp ice claws its way up the castle, encasing it. KRISTOFF (CONT'D) Anna. Without hesitating, he dashes back down the mountain. Sven runs after him, catches up. Kristoff grabs Sven's harness and jumps onto his back. CUT TO: INT. LIBRARY -- NIGHT Anna shivers by the door. She looks up to see ice overtaking the ceiling. The door handle suddenly jiggles. Stops. Jiggles again. ANNA (barely a whisper) Help. CLICK. The door swings open. We see a carrot in the lock and hear a giggle of victory. Olaf takes the carrot, puts it back on his face. Then he sees Anna lying there. OLAF Anna. Oh no. He runs to the fireplace. Throws in some fresh wood, including one of his own arms, which he quickly rescues, before striking a match and relighting the fire. 100 FROZEN - J. Lee ANNA Olaf? Olaf. Get away from there. OLAF Whoa! So this is heat.... (considering) I love it. He reaches a twig finger toward the flames. It catches on fire. OLAF (CONT'D) Ooh! But don't touch it! He shakes the flame out, as he rushes over to help Anna to the fire. OLAF (CONT'D) So, where's Hans? What happened to your kiss? ANNA I was wrong about him. It wasn't true love. OLAF (confused innocence) Huh. But we ran all the way here? ANNA Please Olaf, you can't stay here; you'll melt. OLAF I am not leaving here until we find some other act of true love to save you. He sits down behind her, stubbornly. Leans his back against hers and thinks. OLAF (CONT'D) ...Do you happen to have any ideas? ANNA I don't even know what love is. OLAF (confident) That's okay, I do.... Olaf hops back up and puts a soothing hand on her shoulder. 101 FROZEN - J. Lee OLAF (CONT'D) Love is...putting someone else's needs before yours, like, you know, how Kristoff brought you back here to Hans and left you forever. ANNA ...Kristoff loves me? OLAF Wow, you really don't know anything about love, do you? His face starts to melt. ANNA Olaf, you're melting. OLAF (sweet and reassuring) Some people are worth melting for. But then...his face REALLY melts. He panics, pushes the snow back in place. OLAF (CONT'D) Just maybe not right this second. Suddenly, the window blows open, cold wind sweeps in. OLAF (CONT'D) Don't worry, I've got it! Olaf flitters to the window. He pulls one panel of it shut but struggles with the second panel. OLAF (CONT'D) (determined) We're going to get through-- (distracted) Oh, wait. Hang on. I'm getting something. He breaks an icicle off the window, uses it as a telescope and sees... Kristoff and Sven running back down the mountain. OLAF (CONT'D) It's Kristoff and Sven! They're coming back this way. ANNA ...They-they are? 102 FROZEN - J. Lee OLAF Wow, he's really moving fast. Huh.... I guess I was wrong. I guess Kristoff doesn't love you enough to leave you behind. Anna tries to get to her feet. ANNA Help me up, Olaf. Please. He hurries over, tumbling over the couch, knocking over the chess set and water jugs. OLAF No, no, no, no, no. You need to stay by the fire and keep warm. ANNA I need to get to Kristoff. OLAF (clueless) Why...? (realizing) Oh, oh, oh, I know why. He hops around in an excited display of hope. OLAF (CONT'D) There's your act of true love, right there, riding across the fjords like a valiant, pungent reindeer king! Come on! The walls crack under the ice pressure. OLAF (CONT'D) Look out! They rush out the room just as the ceiling collapses. INT. CASTLE HALLWAY -- DAY Anna and Olaf struggle down the hall. Ice spikes grow and block their path. OLAF We're trapped. Anna looks around desperately for a way out. 103 FROZEN - J. Lee EXT. FJORD -- DAY Elsa runs, but is nearly blinded by the snow and wind. EXT. CASTLE -- DAY Anna and Olaf bust open a window. The storm is so strong it sweeps the window panes away. OLAF Slide, Anna. It's a long, snowy way down. But what choice do they have? They slide down the iced-covered building. Anna arrives at the bottom, weak but uninjured. Olaf gathers snow along the way. He arrives at the bottom as a giant snowball. OLAF (CONT'D) We made it! He shakes off the extra snow as Anna struggles to her feet. EXT. FJORD -- DAY Kristoff and Sven bound off the mountain and sprint across the frozen fjord waters and right into the heart of the storm. Its white-out wind pushes them back. But they fight through. KRISTOFF Come on, buddy, faster. CUT TO: Anna and Olaf reach the shore of the fjords. ANNA Kristoff! The wind lifts Olaf up and pulls him apart. He goes swirling off into the storm. OLAF Keep going, Anna! Anna struggles on. 104 FROZEN - J. Lee ANNA Kristoff! PAN TO: Kristoff rides Sven past cracking, frozen ships. Sven struggles over the uneven surface. KRISTOFF Come on! Come on! Suddenly, a mangled ship, risen by ice, capsizes over them. They give it all they've got as debris falls all around them and the mast shatters. They make it past just as the entire ship slams down and cracks the thick ice beneath their feet. The ice opens up. Sven bravely jumps over a gap. But it's too wide. He bucks Kristoff to safety, but lands in the freezing water and disappears below. KRISTOFF (CONT'D) Sven? Sven! At first there's nothing but the wind and the tumbling icy water. But suddenly, Sven surfaces and claws his way to a floating ice chunk. He calls out, signalling for Kristoff to go on. KRISTOFF (CONT'D) Good boy. CUT TO: Anna moves blindly across the fjord. Anna's hands frost over an icy blue. She stumbles on, determined. But she's running out of time. She clutches her chest. The color in her eyes fades, the inevitable is coming. CUT TO: Kristoff, lost in the white-out, doesn't know which way to turn. But then he hears a faint-- ANNA (O.S.) Kristoff. KRISTOFF Anna...? Anna! WHITE OUT TO: 105 FROZEN - J. Lee Elsa struggles through her own storm, but the fear is consuming her. A dark shadow approaches. It's Hans. HANS Elsa. You can't run from this! Elsa backs away from him. ELSA ...Just take care of my sister. HANS Your sister? She returned from the mountain weak and cold. She said you froze her heart. ELSA What? No. HANS I tried to save her, but it was too late. Her skin was ice. Her hair turned white... Elsa's face sinks as she realizes what she has done. HANS (CONT'D) Your sister is dead... because of you. Elsa drops to her knees, emotionally broken. And with that, the swirling storm suddenly stops. The snow freezes mid-air, hangs suspended, trapped in grief. Citizens and dignitaries rush to the wall's edge and look out to see... Anna, barely able to move but now able to see across the fjords to... ANNA (a whisper) Kristoff. KRISTOFF Anna. Anna pushes on towards Kristoff. He runs top speed towards her. There's still a lot of fjord to cross, but Kristoff is giving it all he's got. He's going to make it. But then, Anna hears the sound of a sword being drawn from its scabbard. She turns and sees Hans, behind Elsa, as he raises his sword over his head. 106 FROZEN - J. Lee ANNA Elsa. Anna looks back at Kristoff as he runs for her. She gives him a longing look, but then turns away from him and then... Using all of her remaining strength, as Hans brings his sword down, Anna throws herself in front of Elsa. ANNA (CONT'D) No! In that instant, Anna freezes to solid ice. The sword hits her instead of Elsa. The sword shatters completely. The force of it sends Hans flying back and knocks him out. ELSA Anna! Elsa rushes to Anna and touches her sister's frozen face. ELSA (CONT'D) Oh, Anna...no...no, please no. Olaf walks up and sees Anna, frozen. OLAF (confused, sad) Anna? Elsa hugs Anna and cries. Kristoff watches in shocked despair. Sven steps up to his side. Citizens and dignitaries on the castle walls bow their heads. All of Arendelle is joined in somber silence. But then, Anna warms. She begins to thaw. Olaf looks up and gasps. Kristoff and Sven notice, light up. Anna bends her arm and embraces Elsa. ELSA Wha-? Anna? Anna opens her eyes. She smiles at Elsa, relieved. ANNA Oh, Elsa. They embrace. 107 FROZEN - J. Lee ELSA ...You sacrificed yourself for me? ANNA (weak) ...I love you. Olaf realizes what's happened. He's so excited about it, he lifts his head right off his body and exclaims-- OLAF An act of true love will thaw a frozen heart. ELSA (processing) Love...will thaw... (realizing) Love.... Of course. Elsa looks at Anna with confidence. ANNA Elsa? ELSA Love. Elsa lifts her arms, and the ground shakes and cracks. The ice and snow breaks away and rises high into the air. Beneath their feet the bow of a ship thaws. The entire fjord melts and other boats right themselves. The villagers come out to see the warmth returning. In one final wave, Elsa draws all of the snow into a giant snowflake in the sky, then waves it away, leaving only a warm summer day. ANNA I knew you could do it. OLAF (melting, good-naturedly) Hands down, this is the best day of my life...and quite possibly the last. ELSA Oh, Olaf. Hang on, little guy. 108 FROZEN - J. Lee Elsa waves her hand and surrounds Olaf with a swirl of cold air. He refreezes. Above his head she leaves a little, perpetually-snowing storm cloud. Olaf loves it. OLAF Hey, my own personal flurry. Kristoff sees Hans trying to get to his feet. He marches toward him, prepared for a fight. But Anna puts up a hand and stops him. ANNA Uh. Uh. Uh. She'll handle this. She goes over to Hans. HANS (confused) Anna? But she froze your heart. ANNA The only frozen heart around here is yours. She turns away from him, proud of her words. But not yet satisfied, she turns back and punches him right in the face. HANS Ah! Whoa, whoa, whoa! He falls overboard. Elsa comes over to Anna and hugs her. Over her shoulder, Kristoff meets Anna's eyes. She smiles brighter, happy. DISSOLVE TO: EXT. ARENDELLE -- DAY It's a beautiful summer day. The mighty ships have been repaired and are sailing away. On one of the ships, HANS is thrown into a brig. FRENCH DIGNITARY (to Kai) I will return this scoundrel to his country. We shall see what his twelve big brothers think of his behavior. KAI Arendelle thanks you, my Lord. 109 FROZEN - J. Lee Down on the dock, Arendelle guards lead the Duke and his two thugs to their ship. DUKE This is unacceptable. I am innocent. I'm a victim of fear. I've been traumatized. (bad acting) Ow! My neck hurts. Is there a doctor I could...No? And I demand to see the Queen! Kai steps down from the gangplank to the dock. KAI I have a message from the Queen. (reading a scroll) Arendelle will henceforth and forever no longer do business of any sort with Weaseltown. DUKE Weselton. It's Weselton! The guards usher him and his thugs onto their ship. EXT. VILLAGE SQUARE -- DAY Anna runs through the crowd, pulling a blindfolded Kristoff along behind her. She's so excited she can't stand it. ANNA Come on. Come on. Come on. Come on! She runs him right into a pole. KRISTOFF Pole. ANNA Oops. Sorry. EXT. ARENDELLE DOCKS -- DAY Anna skips to the perfect spot and stops. ANNA (stopping) Okay. Okay. Here we are. 110 FROZEN - J. Lee She takes off the blindfold. Kristoff opens his eyes. Before him sits the most beautiful, suped-up sled. Sven poses in front of it -- Vanna White-style. ANNA (CONT'D) I owe you a sled. KRISTOFF (blown away) Are you serious? ANNA Yes. And it's the latest model. KRISTOFF No. I can't accept this... ANNA You have to. No returns. No exchanges. Queen's orders. She's named you the official Arendelle Ice Master and Deliverer. Sven shows off the Ice-Master-and-Deliverer medal like he's king of the bucks. KRISTOFF What? That's not a thing. But he can't help but admire her enthusiasm. ANNA Sure it is. And it even has a cup holder.... Do you like it? KRISTOFF Like it? He sweeps her up high overhead and spins her around. KRISTOFF (CONT'D) I love it.... I could kiss you! He drops her, suddenly embarrassed. KRISTOFF (CONT'D) ...I could. I mean I'd like to. I'd... may I? We me....I mean, may we? Wait, what? She gives him a quick kiss on the cheek. ANNA We may. 111 FROZEN - J. Lee He smiles and goes for it. It's a true love's kiss, alright. We move past them to find Olaf enjoying the summer. With his snow cloud safely overhead, he's free to smell the flowers, which he does. Then sneezes his carrot nose off. Sven catches it between his teeth. Olaf gasps as Sven sucks the whole carrot into his mouth. It's gone. Olaf's face sinks in sadness. But not to fear, Sven spits the carrot back out and jams it into Olaf's face where it belongs. It's completely covered in reindeer spit, but Olaf doesn't seem to mind. He hugs Sven happily. CUT TO: EXT. CASTLE COURTYARD -- DAY The gates to the castle are wide open. In the courtyard, stands Elsa. ELSA Are you ready? Villagers cheer. Elsa stops and creates an ice rink. The people, skates at the ready, hope onto it and twirl about. Elsa then freezes the fountain in a beautiful design and adds some snow flurries for atmosphere. Anna comes slipping in. Elsa catches her. ANNA I like the open gates. ELSA We are never closing them again. Elsa then waves her hand and magical ice skates (literally made of ice) form on Anna's boots. ANNA What? Oh, Elsa, they're beautiful, but you know I don't ska-- Elsa grabs Anna's hands and pulls her along on the ice. Anna slips and slides, but laughs in delight. Sven goes slipping past. Kristoff runs after him. KRISTOFF Look out. Reindeer coming through! 112 FROZEN - J. Lee Olaf skates and helps Elsa coach Anna. OLAF That's it. Glide and pivot and glide and pivot. We pull away slowly, into the sky. We arrive at a bird's-eye view to see that where the castle had crumbled has been repaired with a ice. All is right in Arendelle. FINAL FADE OUT. THE END | 1 |
| Other values (14) |
Length
| Max length | 219850 |
|---|---|
| Median length | 54157 |
| Mean length | 72255.368 |
| Min length | 1516 |
Characters and Unicode
| Total characters | 1372852 |
|---|---|
| Distinct characters | 101 |
| Distinct categories | 15 ? |
| Distinct scripts | 2 ? |
| Distinct blocks | 3 ? |
The Unicode Standard assigns character properties to each code point, which can be used to analyse textual variables.
Unique
| Unique | 19 ? |
|---|---|
| Unique (%) | 100.0% |
Sample
| 1st row | Welcome everyone! This is natural Language processing for Law and Social Science. Thanks for joining remotely today. It still is a bit up in the air how we will do the hybrid verses in person versus Zum format. This term, you hear, I'm a little stuffy today. This is true. Avoid nineteen case I Caught it from my daughter who caught it in daycare. It's very mild so I hope if any of you catch it, it's not worse than this. It's definitely manageable. You can see I'm here to teach even though I have it so it's not that bad. I'm a little congested and we might end a bit early today. I Hope that's all right, but going forward. I Would like to do the course hybrid where we have some impression meetings at least, but before text money you'll hear from me about that. Thank you for all of you who filled in the students survey. There was a broad agreement that we should have an online aspect or at least should be recorded so well. we will work with that. So I have a few slides to introduce the course and then we'll have a chance to answer any sex questions about the format. So this course is a applied Natural Language Processing. It's not a course where we will just start with different texts, data tools, or different help models and learn how to come them up in there. We care just as much about the applications of those methods in the law and in social science and this is in the news all the time. Here is an example from a recent legal product called Clarity which uses in all tools to to analyse contracts and for example terms of use to highlight different clauses that are unusual. You also hear these really exciting ideas such as the World's First Robot Lawyer I'm excited about this, I think everybody should you. I think that this technology is is improving rapidly and dramatically and there is scope for many interesting things happening in the law and inoup world. But there also is a lot of hype and we will take a skeptical view of of these strong statements such as the World's First Robot Lawyer And in turn I Think that while there is a lot of interesting about tools coming about for law and social science and other applications, I Do not think we're close to having a judge to be replaced by a contact. some other reasons to be skeptical or to be concerned about the arrival of these legal inopetuls is that they can be biased So northwest. One of the classic examples is different languages have different degrees of being rendered having notions of gender, mail and female in the language and if you translate from English such as she is a doctor he is a nurse to Turkish which does not have notions of gender pronouwns and then you translate back the gender switches so basically they they have since fixed to this in google translate but it used to be where if you to see as a doctor translated to Turkish and then translated it back it will change to him as a doctor just because similarly he as a nurse would be transformed she as a nurse and this is just because. Theories this basis this statistical correlation in language where doctors tend to be male and nurses tend to be female and statistical language models and translation systems will capture that bias. These issues are based. the language models are as the technology comes more powerful, these issues become more intense to get more intense benefits but also more more intense risks and good. It is now a few years a few years old but this is language whose hole it came out in the Tousadand nineteen that could. It was basically among many other things a fake news production engine and it could produce a lot of context appropriate prose. So you can imagine to know Twitter and email. can the news being filled with all this machine produced speech that would drown out all other speech and I think that those those concerns are still relevant. but now that got two has been out for it for three years and there's an even better version called just There that's been out for a year and we have not seen the internet employee. That means that maybe it was not as bad as we thought and so in this course we want to know. Can we take legal Gptto illegal? Get there to help judges in their work? So this is the course. It's natural Language processing for law and Social science and our engineering goals are doing these kind of two pieces. This course we're going to develop skills in applied natural language processing which will include machine analysis, interpretation, generation of documents and those could be on news articles or contracts or judicial opinions or political speeches. And we want to also take a social science approach where we do not just care about sequencing language in different ways, we care about relating it to attend data, and to understand the social forces that underlie these documents. What are their determinations and what are their outcomes and so you knowsome. Of the examples that we will frequently come back to are: what are the motivations for judges? Can we analyze judicial opinions and see what's driving their decisions? Can we look at the writings and the speeches of politicians and to the end where they're coming from And I Think this is kind of the broader agenda or the goal in this research area is Knpowders language matter in society in human relationships s and what can help do to help understand that? So what we will do. We're going to read text documents as data so there's you know many ways to do this and will go over many of them. We will use supervise learning techniques for dimension reduction, topic modeling, groups interpretation, supervise learning for text regression and text classification can be predict from a speech. Is this form a left wing politician or a right wing politician will get at World embeddings, document embeddings, a lot of exciting technologies there for producing these learned representations of language of words and concepts in geometric spaces. and towards the end will get into disclosure analytics. So this is where the linguistic side of natural language processing, cynicism, and a summarization question answering I'm checking. These are all really relevant to legal applications for example. So some course logistics or beating times that will be two even fourteen to it in an sixteen so we'll have a ten minute break in the middle going back to what I started mentioning at the beginning. These are the survey results for the the course format and there were only a handful of you who would be register if there only online and everybody else wanted some online aspect or the indifferent. The based on these surveys we will certainly have a online component. like everything in the class will be durable online. but I'm hoping that we can have some impression component as well. so there's a question in the chat I have the the chat year so I'm sorry but it in general how keep track of the chat so you can always ask questions three were asked to. We need to have knowledge about law I said are to be a good in the class. The answers no no not at all so you do not to have any knowledge of it, you need to open to learning about it. So if you have some interest in social science applications or legal applications of help it will make the class much more enjoyable. but there will be no substantive aspects of health or social science that will be tested. and so given that we will have this important online component to the class I Want to make the most of the hybrid learning. The lectures will be recorded by but in some view that contacted me about different special cases which is fine but if you can it's going to make the class more fun and and more better for everyone if everybody comes in if you're going to be absent let me or the tea now and if you have questions or comments from online you can just type them in the chat as you as the doing or you can use the the raise and front in which a no monitor so help asks are people who either know Pytha nor beeligible for this course so aim going to get to that in a little bit. But the short answer is if you've never seen Python I do not recommend taking this course, it will be too difficult. I mean you can try to stay for the first two weeks and see if it's manageable for you. but in previous versions of the class, people who were new to Python and people who had never done any text analysis it was frustrating for them. and so I do not recommend the course for anyone who's sure asked and well tell you that as some emails if you'regoing to be absent for a lecture, a email email after the tea to let her know and if you have to miss a significant number of courses the email of cause you might have to do an additional assignment. So ya so relax. If you're anyone who is kind of the new to Python or has not done any ex data and turnout sure let me know you can talk about it so avoid asks and can homework only be permitted in Python or can we draw also try to this in or sure yeah you're welcome to try it in our for me I should. I wouldbe great if if anyone is interested in converting the course materials to war that would actually be a great course project so we can arrange to get extra credit for that report. asks what's the do registration deadline? There is not an official one I'm not sure exactly. I think it's varies a bit by department but I do not have an official de registration that line. If you're going to be just for for our grading purposes, it's better to do it before the first response essay which is like six weeks in five or six weeks in because others take a long time for grading and so I would rather you deregister before that. So I would say I think by five or six weeks you should know whether you will stay or not. So smart. Asks if we attend the lecturers and only watch the videos. there will be additional projects yes, so mandatory. The live attendance is mandatory and so if you're going to just watch the videos then you have to do another assignment. but I have not decided what that is yet. Okay so yes, so this is related to newly keep track of course participation through in class activities. So young asks, do you recommend someone who also general machine her knowledge but just to experience with help. If you're actually pretty comfortable machine learning with Python then this course actually wopolity Fine. So if I think that if you're doing the first two assignments, the first two home work assignments and they're not taking you a long time to do if you can finish them within of hours then then your on track. but it mainly do not recommend it. I mean it, if you're quite new to Python then I do not recommend it if you have some machine learning experience than that's good. But as I said some text analysis or snap experiences is recommended. So we have course syllabus I've already sent by email and I be in oppose to league to it again so also asks why this course worked for people who intend buillier of it judge course at so if you took my course in the spring and you l in off if you've done if you finish the assignments of the course in the spring then this course will find freedom so there's there's a little bit of overall. So I would say that he saw in the fall course it was say in the fall course ably report judge it would be fine as a prerequisite for this course. If you've done that then this should be fine for you. So those links are a bit is going to change to it screenshare to the syllabus so I just pose I did a link to this in the home Here's my contact details. Here's area's contact details: the lecture schedule that sessions which I'll cook you a bit in a second but those are at ten a man on Fridays they're not mandatory but these will. They also be recorded and Afro will go over the example coding notebook for the week and then also go over the previous week's homework. This is our daughter's the structure, All the slides including Iardaploi today slides here there in the slides thotfolder notebooks. These are the simple notebooks for learning the material and so before before doing the assignment. You should read the notebooks so you can see. You can kind of skim though you can read through these, ensure you understand them and everything that's in the homework which is here under homework. The homeowners will follow will have similar content to what's in the notebook so you can see we fill in part of the notebook and then you have to add something in a new column text which contains the lower case, title and lead. So here's lead. here's title and so No Nights is an example and here you can just like to nowtype lower so thiswill be how to get lower case. So this is just to show you that these the notebooks and the homework are designed off for self motivated learning of all the coding as aspects of the course. so find asked how do the homers omissions work? So there's there's a homework every week and so it's like this: homework Here you download this Jupiter notebook and fill it out and finish it and then you upload it to add you that it's not working yet but it's going to be on the course model. There's a submission system or using coal ufous up load it on the website and going to be due. The homework are done on thousands but the first homework is done next. Thursday So I can not actually show you if you scroll down so everything you need is going to be highlighted here. So for example, do this week. next week the homework one is done on Thursday fin is that what you were asking? I'm going to come back to that as let me know if you have some other questions about it. So here's me. I'm going to put in the correct system still working on this but camera acts are all homework mandatory if you want it mean you lose point if you do not do them but they are. The homework are a completion grade so you know we're not grading them. We're not grading all the answers but if will check that you did like you tried to do every piece. and if you say you get full credit and basically so in terms of the grading it's thirty percent for the programming homeowners and I do not go eleven homework so mistake. three points per homework or that's thirty percent and so for every lecture will have the slides. I'm going to post the links to the recordings here so like after it today, you'll be able to get the recording for for everyone here. there's going to be a tea session on free, there will be a recording link heiress about. So unique asks what can we think, what the response essays are. Can you be a bit more specific? like do you want to see an example? Okay, well get to that. We'll get to that next week. It may be the We attributes. You do not have to do one of those for a time until a month from now. time is talking about the response essays. Whether's some information here I'll provide some example response essays from previous previous years, but it was not going to get into that into detail today because it's a more complex topic but you can read about them here. But basically what it is is reading a paper one of these papers in writing a response as I about it. Like a review here, I have a link to the bibliography of references. So these are all of like the materials that the slides are based on so you do not. Someone of these are required readings but it's worth skimming through this just to see where things come from. and if you are interested and you want to to go back to add to fill in your knowledge from the slides then you can read these the other. the other required assignment is there is going to be there required readings for example in week for you have to read one of these papers and then we'll do an inner class activity about them but it's going to be. We will form groups and do short presentations to each other about these press but I'm going to provide more information to that in week there before we do that. So the the three pieces of the assessment are the homework on the coding homework which I showed you the response essays which I mentioned or reading a paper and writing a report about it and in third there's a end of there's an end of the course assignment and its in the you So we would call them an exam but I think here you would just say it's an end of course assignment where you have a few days to do an assign. For those of you who were in my class in the fall you know this is like it's a questionbasicly a question about every like sure in some questions about the required readings and so that the end of the course assignment is one the things that we will cover in the lecture are covered their of sotthat's how the course will be assessed I'm going to cover some of that information again now just in the slides so it mentioned awards the is the first that session will be on fairly area's here as well. After do will introduce yourself sure here one man after I'm a packed student at an centre and I hope to see you in the first session. So in these is sessions it's what would expect far will go over the notebooks they code note books from the bathtub and then the last week's homework after you've submitted it and then usually there will be some time left over an area can turn the recorder off and you can ask some office hours time questions. I'm going to pose course announcements on Model and if you were registered for the course this morning you should have gotten an email about that if you did not send me a note after class and so we can try to figure out your muddle registration. it's not ready yet but I'm going to work with airfare to post it but we will have this to in a forum on model and so you can post questions there before next week's class and I'll go over them at the beginning of class or I'll just answer on the model. So I wanted to make a note about the course work load because this is not like other science and Perspectives classes like it's not much work to the extent that I've gotten negative. I mean I just want to say expectations I have got a negative feedback on the class because people thought it was too much work and so the thing is, it's actually a lot of work for me too because I have degraded the response essays. So it's actually easier if there's fewer students. So if if you're worried about the course load, then there's other classes you can take that do not take as much time, but according to it, would increase. The number of credit points at I is not the credit system is the maximum for a Science and Perspectives course, but the number of hours for most people If you satisfied the course prerequisites such as having some Phantom background and a little bit of blip background. the amount of work in this course is less than ninety hours. And so it's twelve lectures, eleven programming assignments. There required readings to response essays, and then the final assignment. So that's about sixty hours of time just actually doing these things. And so that includes three more hours. So that includes the tea sessions and then study time if you are new to pythem especially, but if you're new to help then it will take longer. So I just wanted I Want to say expectations about that beforehand? Also, if you were interested in this topic of applied Up for Social science then I would highly recommend you also sign up for the two additional credits for the course project so we'll talk about that more after class next week. So if your interested in it, just stay after it you. This is simply recommended for people who might want to do graduate research after because the previous course projects have turned into conference and journal publications. two of the projects were part of into Swiss startups as well. So if your interested in legal tracker or other entrepreneurial projects based on applied help then the course project could be interesting for you so then asked one where doing for the submission of the project. there's there's a link from the syllabus on course projects that has the rough deadlines. Basically you need you haveyouhave to pick a topic within the next month and then have an outline within the next month and then the full draft of the paper is then day until remember September first so you can work on it over are so a related system of what we've talked about already. Thank you to everybody who filled out the course survey. if you registered since I said this out, send me a note, email me a no because it send you a link to the course survey. Oab'll just send out another link so you can fill it out as well if be curious who else has joined. It's about half master students and few old students and then the rest bachelor students and mostly computer science some data science. He's actually zero people from law which is somebody asked do we need substantive law background so if we did not, we would lose all these students. So we do not require that so that two you guys are So I Already went through this a bit in the syllabus, but the required readings are indicated in the syllabus schedule. In addition, there's the bibliography of references that has additional readings if you want to complement the slides in the link related to the response essays. there's the list of applications papers for response essays, which will talk about more next to be. So I wanted to just give a quick outline of some of the books that were mentioned in the references list. Again, none of these are required readings, but for those who want a deeper understanding, I do coming these books. So Natural Language Processing with Perception is the book that accompanies the Natural Language Tooloquate, which is just this classic blip trouble with kind of more standard classical like old school machine, old school natural language wing tools. If you want to learn machine learning, this is my favorite book for earmachine learning with Physicist learn and wood courses and Monster Flow. It's more generic and's not about Inop specifically, but there are a lot of top applications and for those of you in my course in the Fall you would have already seen this and this is available on oil through the Earth Library. you should be able to get it. This is a free book if you're interested in more of the theory and guess for natural language processing more product than mathematical formalization. If you want to do graduate work research in Blip, then I really recommend this book the Your Goldberg book. I think this is available for download as well on the Earth Libraries. If you can not find it, let me know I can get you a Pdf even though it came out in to this and seventeen. It's actually a little bit out of date unfortunately, so it basically has everything up until Transformers, which as we'll talk about have kind of remade inilp. but a lot of the issues and approaches here are still quite good. Another kind of classic textbook is necessary in Martin. Its kind of more than the does not really focus on neural nets inalp and is kind of more than the older generation of help research, but it's very good for some of the more linguistics oriented in semantics oriented part of the course, so this came up a bit already. Python is a course prerequisite see here for the example notebooks and you know I'm sure many of you as I said, Python is a country register, so you should already have it set up on fairy affairs can provide some help in setting things up, but we would trust everybody. Everybody should be able to get their own another environment running. As a prerequisite to this course. these are the main piping packages that we will be using. As I mentioned in all to is this broad collection of older Inalp tools. Finish is great for topic models and award embedding. Spicy is another kind of generic tool. It's great for named in any recognition, parsing in reference resolution, things like this as well as a library of pre trade world factors. and then as it mentioned this new inilp this new neural net architecture called Transformers in particular large pre train transformer models that are trained on large corporate these have really remade how help is done and hugging base transformers as the hugging base system is the standard for that. To provide an overview on the course, here are your objectives: seventy percent if you want to learn how to use help tools and fifty their parents if you want to learn how to apply opinion tools for law and social science so this is great. We're going to do both in this course which are the followings best Matches your goals for learning in top: sixty percent want to learn it for Engineering in Software development Thirty seven percent for social science research and fifty three percent for computer science research. This is good. We're going to be doing all three of these goals are going to be covered in this course so avoid asks if we need to into processor to no no and maybe you're asking if you know like at you for the assignments. The short answer is no and you do not need any special computer equipment the yeah so we you should be able to the examples on the books and assignments. We use kind of small corporate things so you do not need you do not need any specialized competition for that. If you have problems you can use a Google collaps right and so Afro will cover that in the tea. sure you can just those Google could for everything. So why this course right now we live in this really amazing time I Think for language processing where with our lifetimes there's been these new social structures that have remade the data landscape. the Internet Social Media digit join efforts by governments and Google Books for example just as amazing amazing initiatives for digitizing tons of text at the same time as having these huge crops are. We also have had this amazing increase in computational resources as from cheap disease to efficient databases, solutions for quarrying all of those corporate and then having cups give rise to go Pus and then also tips for training these gigantic volunteers and in particular for natural language analysis. We have these really interesting tools in machine learning, a blip and casual inference for the legal and the social science applications of these tools. And for me I Think it's fair to say that at least from a research perspective a lot of as these trends are especially amplified in the law and Legal language. Political Science and Political Language Here many doors that are being opened in these old fields by these new technologies and so we care about legal and political institutions such as what judges write in their opinions, what politicians say speeches, what's written in patents or in newspaper articles about the law or in legislation and regulations. Those are all millions of lines or millions of documents of unstructured texts. and there's no way that humans could read them even if they wanted to add. So this is why bring in these tools for computers to read them is so exciting. So manual asks, could you share the response to the questionable students background acknowledging presence is up. I do not have that in the slides but if all talk about that next week. I don' think there's anything not notable from that. or do have a specific question manual right? All talk about that a bit next to be but but you do not need to worry about that. So here's an outline of the course and actually I would say let's will all just go through this and they will take them break. So I know this is like an eyefool and I made this last year, but we're actually going to follow basically the same format and justice. but you can visualize everything that we're going to learn in this chorus from this gap. And you what? what we're starting with as raw data today. and next week we go from raw data to segmented documents that are pre processed in different ways. And once you have these documents, you can only use these in some social science analysis just to say oh well, how long are the documents you know? How many bonus do they use? What's the word link to the sentence Link This public a measure of reliability or sophistication in language. The second would be dictionary accounts. and I Think if you're a example researcher, a computer scientist, the fact that you should just go and count different words and count the number of the times the word good shows up as a measure of sentiment that just seems so primitive. it's like the stoneage it. But it I think that we should consider those models cape seriously and I'll give you a good reason at the end of today why dictionary methods are are not to be ignored. And so next week we'll get into tocanization. So the different ways that documents are split up in to sendances and words and looking at part of speech things like this. Once we have documents as these end up being teprimitives for all the models that we will analyse including in gram. So that's converting tokens to phrases. Topic models that's converting documents to distributions over topics so you can imagine in illegal groups there's a crime in contracts on tutors and patterns and things. Each stations are left wing politician I Think my internet might be unstable but I'll give it a second. Can you go hear me now? Can you put it in the cash? I back market thank you So think asks do do we take a look at how he's methods roughly work or do we may learn to use them or what were weregoing to do both rateboth so we will in the notebooks in homework. In the tax sessions we're going to be learning how to do these things in Python we will implement them, but in the lectures were going to be focusing on whether's the purpose, how whatever, going to help us understand the judges or the lawyers things and's so after machine learning will get into neural nets and a particular if we'll learn about word embendings which is a way to represent language in a low dimensional space we'll get into passing, getting at syntax, the relationship between subjects and objects, agents and patients. This is getting into linguistic sides things. We will then get into Transformers and that part of the course which will lead to are ample language modeling knowledge graph's entitlement. So this is getting into asking a computer does a empty public, does sentence A empty B or unknown We do information and extra going to extract relations from a corpus and learn about what the corpus is about and towards the end will be getting into these of more global semantic aspects of summarization. Question answer, automatic claim checking, casual inference from documents, identifying casual relations in documents and a lot of this gets way past that social scientists are using right now. But I think these technologies are improving rapidly in the case of the legal domain at least the there going to be clear applications that really have not been done yet but or will be running right to the frontier in this course to take a look at this if you like will be following us as long as we go in the course. Okay so I know we've already gotten abunchab logistical questions but I wanted to take break now and give everybody a chance to ask a question and then we'll take a quick coffee break. So are there any questions at the moment that have not been covered? You can put them in a chat or the race hand function so manual ask how relevant is reinforcement learning to snap. That's a very interesting question You next's not universal, but it is certainly relevant. There are many very interesting reinforcement learning applications help for example the recent paper that cannot used as reinforcement learning to to improve the quality of summaries and I have a paper with a old student which actually came out of this course using reinforcement learning for attractive summarization. So if you understand reinforcement learning, there's a lot of interesting applications and I can provide some more resources on that. So so fun. Also, memory can note of that area. can you make it note of that? They set up environment script that must have been something from last year so we'll fix that thank you find so report access. Is it possible to do the course this semester near the party next year? Sure it the mean things change right but I'm planning to teach the course again next year so should be fine. Thank you for asking that. So think asks on right, think is asking. theatre's not a simple yes or no. answered to that question Sometimes yes he mean it depends on it mean we're not. We're not going to be like trying to implement them and see Plus Pals or something so but you knew will be talking about you know will have some review of Nstarcastic Radiant Descent. You know how volunteers learn. You know it is not as it said it and this is not a substitute for machine learning in our Pop course so we will do some. but if you need that then you had take a different course or take both of right? So we're going to take a breaknow and will return in nine minutes at there fifteen. I'm also going to turn the recorder off now So if you want to ask a question the recorder please do of really's We really started the content of the course for the remainder of today so nfuzum in on this on the course outline We get to the Corpera. so text data is a sequence of characters called documents and the set of documents is the corpus which we will call them in this course. And what makes text data challenging but also interesting is that it's structured. It's just this stream of characters and the information that we want is induced in the symbolic information in the those characters and it's mixed up with a lot of information that we do not need for any and task and so a lot of what we're going to do in this course is its information extraction or if you prefer information disposal. we're trying to get rid of the information we do not need and will be the main focus of what will do next week and a week after. And to recap something that was mentioned earlier, this course is appealing in place and we do not care about that much of the documents by themselves. What we care about is related into better data and that could even just be like the time Theatre document is published. So we might say well, how syntimate towards a impose social group, How a sintimate towards immigrants changed over time And we can. we can. make a systematic measure toward immigrants and end show that How that's evolved over the last ten years less one hundred years. And the important point there is that we have met a data on the time and so just to say that it be more specifically new might start off with some positive negative syntimate capital by and judicial opinions. And that by itself is not that interesting for a lawyer or for such a scientist. But what if we had the dynamite in opinion is by Judge J at time It and so we will often use this type of notation with these subscripts for the indexing corresponding to the meta data that the document corresponds to. And we can say you know how to sintimate very over time. Or let's say we have the information on the political party of the judge are they do. They have more negative sentiment towards defendants from groups Go. So let's say that go is a dummy variable enjoying one for cases where the defendant is an immigrant and so then we can is information in our data set to say. Well, the right wing judges. They have more negative sentiment towards immigrants for example and you can see that one you relate the text features to meet data. There's many interesting questions that one can answer and a precursor question to you. This type of analysis is what's the document. So now often have this big data set. If we might have a bunch of books on what do we do with them, we need to zoom input to some degree to find out which document, which observation is useful for our meditative variation. And so you know if if we do not want to just arbitrarily make the documents really small because they do not, they do not help us answer a research question such as you know our republican judge is more negative towards of immigrants. If we made a data seat at the sentence level for that, the sentence would both abstinence. data would be super high dimensional, but we would not have any additional variation comparing the right wing judges. did the left win judges. And so this is going to be one of our first Islands activities. what should we use as a document in these contexts? So I Want to take about five minutes six minute to answer these questions? We're going to set up them, going to set up breakout groups to put you guys into pairs and this is just a way too please pull up the slides on your own computer because you're going to lose it when we stop sharing and I'm going to put everybody into pairs and this is just a way for us to start to know each other. So you going to be. You're in a breakout room of two or three people. introduce yourself and said that your major and what you are interested in the chorus and answer these three questions together. What counters the document and were is going to open the breakout rooms for six minutes and will be back at Fifteen Twenty seven so only so handle in corporate. So we have some examples in the notebook about some breed processing to especially if our working on a project not just in this course but you knlater on in your career in life. It's good to think about from given questions or given to ask the data sets out there that have been compiled and so far for court cases like in the U is for example court listeners good but in social media there is really excellent data from Twitter and Facebook for example. for Red it is also for Wikipedia All these data that's are really useful for such a science analysis. This is not in part of the course necessarily, but you know it will be useful for you later on to learn about queueing ships running web scalpers doing processing to remove home markup and there is also the hugging face hub and hugging face system. They have a lot of amazing data sets so it's good to just kind of produce through that ski that a bit can fu access should learn or should have learnt so it would say it mean that should learn because it do not be tested in this course but it will help you to know it for you for other things. So I recommend learning it but you can do it after that. We do it in the summer so all everything that we talk about this course is kind of language agnostic I'm a native English speaker so everything is going to be focused on English. but for your course projects and things you're welcome to do things in other languages. After one of area's special teas is multilingual implants so she can help you get started with it. And there's also just really great machine translation So Elevthos asks why should we use Cyllinium is not cycle evercovtalate that it depends on what what you trying to do then I think that doing a webscraper with that you are a Lib is like a nice start. but if you keep going on up for social science for data science then we will come to a point where you will need a crime driver or something on those lines. So how can we use the quantity of text as data So you know most of this course is going to be about the semantic or conceptual or symbolic content of language. but there is also interesting things to be learned just from the service features just the statistics in the characters in documents and one thing that me and old students in my group had looked at is how the style of writing changes over the life cycle for judges. and one of odd or curious thing we found is that older judges that use shorter words but longer sentences and so whether this is like better or worse writing thing is kind of subjective but it shows that there is. It seems like there is this of biological component of writing style for judges. Another relevant debate in the case of legal quantity of text, the law is on legal complexity where this depends on what country you're from but like in the U's and Finance for example they're always complaining about the law being too complex on but using help we can actually try turn this into an empirical question and ask how complex is the law and certain bedroom which is one of the applications. Papers linked to the syllabus is about measuring the complexity across different areas of the legal code and they produce the number of words in a code title which is a section of the code and they also produce a word invite measure which is basically it's the tiny of the probability distribution across words and what they show is that Public health and the Tax code for example is very complex. It has a a lot of words in it but if you look at the codes if you look at the code titles that have high word intern so a more diverse vocabulary it's quite different and so you. The Commerce in Trade or Public Health and Welfare scores highly on both measures of complexity and so from a tax lawyer perspective. This is kind of interesting because it shows that even though the Internal Revenue Code is complex in terms the number of words, it's very repetitive so it might not be as complex as if sounds. So the next set of methods that will talk about today our dictionary based method which is counting words or patterns and so in the notebook we show some examples of using regular expressions which is this really powerful string matching system and this is going to be depending on what your research, question or engineering task is how you would want to to do this. So one theme or question in the law or judicial behaviour is do judges? Are they making decisions based on justice or morality or are they doing it based on cost and benefit in analysis or efficiency? And so you can imagine using a dictionary method to go through to what different judges say and say, do they say justice demands or do they say efficiency demands And so a dictionary based method could help you get it that. There are also general dictionaries such as Wardnet and Luck which will talk to you about in a second. One example from Economics and this is also one of the applications readings is Baker Bloom and Divas where they wanted to know what is the amount of uncertainty in the macro economy and this is like if you're a Mcroeconomister, you're going to finance a big beam of when there's more uncertainty in the markets, they're more volatile. That could be costly. And so they built this data set of newspapers and they use this kind of simple dictionary method. Does the article that the word uncertain does it have a word related to the economy and then does it have some words related to the government and then they ploddedin that and this is actually just the main result in the paper is just showing that this policy of uncertainty index based on the newspapers, it spikes during these kind of stock market shocks. So like Black Mundane or the Wars elections, nine even. This is the financial crisis to two thousands and nine the Euro Death Sealing crisis. And so you can see that these kind of intuitive moments of market uncertainty are captured by this newspaper based measure. There are are some problems with that to be sure if you're curious about using this in like financial applications is recommend to keep that all paper. another set more fundamental dictionary methods are are available in World That which is this nice python package with a database of it's a big dictionary with all the words in English and how it related to each other and so you can see for example, all all the different senses of the word bass or bass are located here. so it's like music or voices or the fish. There's many meanings of the word base or bass and the the word that it captures these synonym sets of groups of beer anonymous and has information on the anthiponyms and also parts and holes and then also all towns are organized in this categorical hierarchy and so you could have employers which are the higher word in a category and then symptoms are the members of a category and so you. There's many ways this could be used, so if you want to do definition reduction, you could replace words by their hypernem for example. And if you keep on going up Twilight of Categories word that has twenty six lexacer, gray graphic suppresses and so like action, animal partifact person relations shape things like this. I think it's pretty cool that these linguists have gone together and really tried to organize all of language in this and it's all automated now and a Python and so they did the same thing for verbs as well. So this is for bonus and this is for verbs and so you could take a corpus in, say, well, which books are mostly about perception versus possession for example. and you can imagine there's a kind of digital humanities or cultural analytical types applications for these methods. Some other general dictionaries which are relevant for us include lists of function words so thinking of words like four rather than also these get at the style features of text so most of them have you going to drop them because they're not going be very useful for topics, but they can be used to get it nonpolitical dimensions. So if you want to identify authors of texts without that being correlated with the topics, then using software is a good way to do like. or luck is it's kind of the standard licenses for linguistics and Social Psychology and the like team. They have seventy lists of word's of category relevant words including the commotion, cognition, family, positive, negative. We will see some applications using Luck later on in the course and in more recently there is these big lists of words that are initiated by people on crowd scouring platforms. So Mohammed and turns have joy and sadness, anger, fear, trust, disgust, anticipation, surprise and can warmer at all. They could fourteen thousand words along violence, arousal to dominance dimension. So these last two on kind of emotional content. Those are part of this broader set of tools in blip on sentiment analysis and this. You'll see this in many papers, but also in an industry like in advertising digital marketing, they really care about determine right for obvious reasons and we want to know for a given like a review of a movie. Is it positive, negative or neutral And the standard approach could be licensed. base research for the word good, but it's easy to break that right. Like what if somebody said though the movie was not good or the movie was not very good and so just like amends other words totally transforms the meaning. and so it means that just counting words often is not going to work very well. And the more moderate approach is to use pre trained transformer based syntimate models in area I Think is going to add an example of this into this week's notebook to show how you can download a pre trained sentiment analysis from the Hugging Faced Model hub and apply that to documents and you should be careful about these off the shelf scores through these are trained models because they might be trained on a corpus that's quite different from yours. So if you take a corpus that's like initiated tweets and then you apply it key contracts, that problem will work right and so you have to be careful about that. Check that it works in your new domain and there is also some methods which will get too in the world embeddings for making some domain specific licences so you can start off with some positive negative words and then you might find that in read it. the positive words are different than the ones in Instagram for example. So I wanted to just point out a specific issue with syntimate awareness that are based on initiated data. So if you take a syntimate analysis like from hugging bass or something and this is from from some slides by crime car I saw on his Twitter feed. so I do not know exactly which model this is but he made this where if you just take let's go get initial food versus let's go get medical food this has a very positive sentiment and this has a low sentiment. This is bit so just changing the word from relation to Mexican but just that soon as by itself the sentiments exactly the same right? but they changed it from relation to medical and the sentiment went from positive to almost negative and then this is an even kind of more striking example. If you say my name is email, you get this positive sentiment. but if you say my name is unique while you get negative sentiment and so this is like really kind of striking and important and concerning issue in all Pi systems and you want to ask the mean, Is this sentimental model racist? What's what's going on here? How did that happen? And this shows how you have to be careful about using symptomatic scores that are trained on initiated data sets because they're also going to be learning this correlated information in the data set. so that's not unsintimate, but then it will learn it and apply it when you use it in a new domain. And so this is just part of this broader problem or challenge or issue in an Apple for social science. But also this is going to be relevant to many things for products as well that we care about. some true sentiment in language. It but what we get is a systematic scorer. and the model that we trained to learn a sentimate score from language has all these confounding factors. So you nonwhite examples for medical food versus Italian food. You can kind of see how that might have happened right where initial restaurants. maybe they tend to be more of scale or like thecritics like them more. and so because the language model or the competent classifies trained on these biased data sets. that food this, let's say, food critics like Italian food, then that gets baked into the intimate model even though it has nothing to do with even though it's using these words that are syntimate neutral. So this is not easy to fix me. You know, because there's not going to be, you can not just make data set that's neutral, like every data set going to have some biases in it. And so I Think that trying to fix it is this really important and interesting area of upon research that's happening and and this is not unique to determine either. This is a universal problem that I want you guys to be thinking about Throughout this whole chorus is that we are trying to measure this true quantity in language, trying to extract it to learn about science or to learn about to solve a legal task to make a prediction and whizzl we care about. But we get this. We can only make a measurement of its indirect measurement and it's going to be confounded by other language features and sometimes non language features like where it comes from or the large age or style and supervised models are just by construction how they're built. they learn features that are correlated with the label being initiated and unsupervised models you in a topic model or world embodying there also going to learn those correlations and so you like. A classic example is like the similarity between the word banana and the word yellow is actually pretty low. but the similarity between the word banana and the word green is actually really high. And it's not because bananas are green, but it's because if a banana is yellow you we just never mention it right and so you have to be very. This is just some examples of these these problems or limitations with these language models you have to be careful about when you're interpreting their their outputs. But and this is what I mentioned at the beginning dictionary methods. They do have these other limitations But They very significantly mitigate this problem. And because the researcher is familiar with the context, they know what would the words mean and so they can always regularize out any serious surroundings. And so if I'm like trying to make a sentiment analysis and my model tells me that captain means high positive, I can easily fix that with a dictionary method, right? And so this is like a defense for dictionary methods. potentially. And I think it's why economists in particular and just other empirical social scientists. they might still use dictionary methods because of this reason. And I mean they have these other limitations which you you can not. Those are difficult to deal with. but we have to be careful with these blackbox models. Okay, so let's wrap up so the first assignment is already put on the gatehub as well as the coding example that Afro will go over on. Friday. So this came up. We explained it a little bit earlier actually, so I'm just going I'm going to stop for a second and answer elithereosthis question sorry is missed that earlier Those elyptherias asked are those general dictionaries useful any more since we can not easily measure the similarly between words. and also such dictionaries require a lot of human labor to be kept up to date. So I think that's an important point. So I mean the general dictionaries. They are built up over time by these research groups. but I think they have limitations and I mean I think they should be updated over time but meant I think of now all the methods that we'll talk about in this class in terms of dictionaries and classified things. They're all going to have kind of prose and cons and we's want to identify those problems and think of ways to address them. So the way that time the timeline will be in this course about the coding practice and the homework is that for last week to so like week one, the notebook, the coding notebook for week it is got to be reviewed in the tea session on fairy week it and the home work for a week it is going to be done on Thursday week to plus one. So for example, week one notebook is going to be reviewed this fairly homework one is going to be done next Thursday arch third and then reviewed in the week to to session. the notebook two will be reviewed in the week to that session and so on. All of this is in the syllabus in terms of the days, so you now something is confusing. We'll cut you guy some slack in the first few weeks of course, but I think it'll be self explanatory as we move forward. That's the first week we have five minutes left for any questions. |
|---|---|
| 2nd row | I Just want to recap quickly what is already announced to the class because we now have this beginning of room for everybody to joining persons. We will prioritise the impression teaching, but there are a number of students for various reasons who will be joining remotely. Of course, if health or other issues are relevant for you, please do not feel an obligation to coming frozen if you need to from home. That's totally fine. So to start, there is a number of questions on the module queue on a page that I'd like to answer. Imjus can I take these in the order of top of posting. So one of the questions is is it okay to copy paste from the note looks for the homework in general yes so you know you're doing it. It's good to understand. I Do not recommend copying and pasting and just blindly just copying the notebooks. It's better if you understand what you're doing, but so formal there's no formal constraint on doing that. The second question is homework emissions, so there's a number of them about what to do with the field. out, no books. The edge flow page has now been set up and you should be able to reach it from the course model if there's any questions with that. let me your after to now but there should already be submission pages for all of the reducing homewares and I think in ask the same thing. The edge flow had not been installed yet, but it should be up there now. so another question is about or these this? thank you for those questions but I think those are okay now. Okay, so could you raise your hand if you in the Computational Statistics recession only I Think with three students, we can not change it. Unfortunately the the attaches are recorded and so please just watch the recordings. And if you join at eleven at the end of the computationastatistics season, you could still be able to ask questions andsomebody asked if computer science bachelor students can count the credits as gas or minor courses and actually do not know about that So it would say ask or someone in registration for your department at experience be two of that because it was one of the guys who actually ask you only not looking that based people on Du Tesigns datchelers this force is ramble even as an agantion step as well as a go as a guest from basically depend on how you begin when you chose your when you chose to basically woman in store, future of future enron, an ex sex vacations of them that be used as a agitation scope or or compulercize nature if you show that there's a guest hop analysis over as a guest. Score Great Thank you. Relax Are there some other questions or clarifications about that? I Probably is not answer but check with your registration officials at your department and let me know if something if you need more information about that right and somebody just asked to clarify if the course projects and the district are counted differently and mean they are counted actively in terms of course credits. but I'm not sure if you can count them under different programs. from my perspective if you're department will let you then it will let you. I do not have any problem with that or right? I Thank you Dived for posting that you can coordinate this with the drink study administration and that's on the model if you want to check out. Answer: Okay I think those are all the you and a page questions. Are there any others that anybody want to bring up live to Asked: Can I do the course project in the summer or or is it does during the semester You're supposed to do it during the semester but it's not due till the end of the summer. But if you have any question about locals just set me a note role you can figure that about. Okay so this is related to the question about the homework what's due So on Thursday by midnight area will go over the homework homework one on Friday morning on the to session. it's best to submit the just the ipi and believe the Jupiter notebook few directly onto edgeof for their completion grades so we do not go through every homework and check them. But so you'll get full credit for completing the substantial completion so this is going to be targeted to combination you will just spotcheck radio, low spot check up and also a programmatic component if you have any questions about whether you are whether a specific programming assignment got for fighters or not a just a check with autism we do not like you know it it becomes available for late later you its past fails so it's just zero or one whether you is if you in our judgment and I think the homewerks are not that sophisticated so I think you can give in every section and more can try a good try even if like if there's an error that you just can not figure out just limited anyway and if there's evidence that you tried it then you get clean if there are there any questions or feed back about the first to session. does that a formate work for everyone at you mean should every person work on their own notebook? Yes yes So I mean for the assignments you can work as groups to solve it but prepare your own notebook and limit that on your does I answer your question. So for this data session on fairy app will go over as the homeortist to be ready and then also the week to notebook on tocanizatioand you can use the Qna page from on on model for questions in the tax session as well. So I mentioned this a little bit last week about the final assignment this is going to be first. It's just a kind of an exam you might say where it's just covering all the material of the course and it will be distributed a week or two after the class ends. It'll provide the exact dates with plenty of time in advance and it will just be based on the ladies and the required readings. It will be designed to be pretty quickly two hours to complete, but you'll have a few days there. Four days may be a week to complete it so you can schedule this around your other obligations at that time and we will do some review and will show you some sample questions during the class or so. Last week we started with collecting and cleaning corporate and doing a quantitative measure from corporate and also a dictionary methods. It wanted to introduce an example of a dictionary method paper from my own research to see how these dictionaries are designed to solve social science problems. So this is joint work with David So who is a student with me here as well as Art at Being At Work and warm rural at all. And this paper is motivated by the recent discourse about race in ethnicity, issues of discrimination and prejudice in equity as a policy problem and does that policy problems motivate that there of above more research about this issue And there is this anecdotal or stereotypical understanding that Economics as a social science compared to political science or Sociology has done less work on these issues and previously this was just in people's heads. But in this paper, we actually look at this eventuality by applying nil up methods to the text of scientific articles. So we built a corpus of five hundred thousand academic publications from these three disciplines, and we used a dictionary approach to identify against the relevant articles. and so please read the paper if you want to see all the details. But I Wanted to note a few highlights just as an example of the dictionary methods approaches that we started discussing last week. So first we considered all publications in the journals that J Store characterizes as comprising the disciplines of Economics, Sociology in Political Science. We also topped up this corrupt with additional articles from the Scoops data set and also the Web of Science data set. And so the goal was to get close to universal coverage of published articles from and Beaten and Sixty up until to Thunyad Twenty In that end up with a half a million in publications. So this is exemplary of this first step in social science research, which normally in an opinion class. You do not worry about this right you just take the courses given, but in order to answer questions such as how much have the different social science disciplines been spending on these issues of discrimination and so much time have they been spending on it, you might not be able to answer that immediately because the data and does not exist. This shows that building data and cleaning it's this important part of social science applications of us even before you get to the modeling part. So read this live can you like? But I just wanted to point out that our dictionary method is actually quite subtle in the sense that we we be matched a number of patterns. so we have a number of synonyms which we which we refine by looking at examples of different race inathnicity groups as well as related topics. So we did not just look for the word black because that light refers to the work to the colour right and so in addition to requiring the mention of a research group is also required a mention of of a topic of discrimination, inequality, prejudiced bias, historical issues like slavery or him cargo so that in order to yoincrease precision of the resulting dictionary based classified identify articles that are had do not just mention these relevant words but also have a related a substantive topic and shown we did invitations of articles. This turned out to help a lot because often times you'll have an article that's about you now minimum wages and you'll have like in a system abstract describing what the paper does and then it's only the last sentence that says and then we looked for heterogene eighty by white, black in hospital that's actually quite pharmaceutical and so we wanted to focus on articles that were very so typically about the issue of race and acidity and these different subtle revolutions and refinements of our dictionary base method allowed us to do that. These are the main results from the paper showing that even though Economics in the top left, economics is in blue, political scientists and green and Sociology is in red. Even though Economics has the most papers, it has the least papers about race in ethnicity issues and Sociology has the least papers but the most papers about race in activity issues and so this anecdotal since that sociology especially but also political science are paying attention to these prejudice discrimination in equality issues that been anecdotal since that kind of conventional wisdom turns out to be right at yes that's just because of coverage. actually they're not in the database so if you so probably to get higher quality data is that we should have finished around twenty and fourteen or so. But it's because it takes some time for all the journals all to their articles to get into the day to base the easing of. But but if you included everything for real it's going on Still is even speaker destroy the way the number of range related obligations who then do through thawing wanting worse or it's journal quality weighted the question. We should put populist that into the title, but it basically multiplies the number of articles by the impact factor of the journal. So basically journals, they get sides often, so they're kind of more influenced on the discipline. they count more on the graph and so this is really just. this is mainly a robustness check more than anything but we. He wanted to check that it was not that Economics might have fewer publications, but they're much more impractical. We can rule that out with the bottom left panel only that's dictionary methods. Now we're going to get tobasically the whole. Most of this lecture today is about taking documents and transforming them into restrictions, and you'll see that there's many ways to do that. Almost everything that we do in this chorus is about document representation learning, a numerical representation of plain text documents, and the process of taxation, which is nsegmenting the document like breaking it up into pieces like paragraphs or sentences or words, or where pieces or letters. That's a pivotal step in that process. So just to summarize what we're doing, we start it off with some plain text documents and what we're doing today is converting notes into tokens, and then one of the workhorse document representations in blip historically and also still today is immigrants which is basically it phrases. So we a document goes from plain text to account a frequency distribution over two or three word praises and is be going to try to keep this notation consistent throughout the course. of course some papers do it differently so it do not be hundred print. but will you use capital detail referred to the documents in capital way to refer to tensions like lists of words and capital x will be some market representation or frequency representation constructed from the tokens. So to summarize and you when you are reading newspapers and somebody asks how did they recognize the documents and why these are the country factors that would be relevant for us so they need to be informative or predictive for some task be text classification, doing a topic model, training, word embeddings, or building a language model they should be. This is one of the goals that is often not satisfied, but ideally the tokens are interpreted so the holding everything equal would be useful to be able to count the number of times each word it's mentioned rather than each letter. If you just had a representation of a document that was counting a number of times each letter like G X C L was was mentioned, you would not be able to look at that data and understanding anything about the document was about. But if you could look at the talk words or phrases in a document then that will tell you we' much more interpretable and then finally treatable. So this is becoming less and less of a problem as computational resources increase. But it's still the case that you know you have a million sentences in pose and you need to compute the paradise similarity between the sentences. Let's say I want to have a million sentences I want to know how similar are they to each other? That's a million times a million comparisons, right? So how those sentences are and presented will be really important compucationally. So there are two broad approaches to this one. if you might call in the standard or classical approach that is from pre neural nets or at least pre recurrent natural nets blip where documents were represented by counts and by the longer sequence longer and sequence information in the document was removed. And then there's a more recent approach which is to maintain the sequential representation documents so you can imagine in Tea. In the first approach, you take a playtext document and you get some counts over different words. In the second approach, you get a list just the origin, all words you take in the whole document as data rather than accounts or clear over vocabulary. This is a no kind of Opstrak, but we will see examples. Here's a askematic for the first kind of of recognizing pipeline, so this would be a way to do this in Python in ineltik. The first line reads in raw hotel text from your website. You can clean out the hotel using the number of approaches such as a beautiful soup. take some snip it of the data tokens would be Taking the raw text and splitting it on the space is what it normally means. You get a list of words rather than strings and then for example you could say you do lower that will give you the lower case that will start doing so. preprocessing, putting everything in the lower case for example. and then the vocabulary at the end is just the set of its heat of unique words. In the purpose in this process of tocanization, you can think of it as building a degree. arguably most of the time it's kind, the space of possible representations that a document can be mapped into. The second kind of tocanization as used in transformer models. This new generation of help. Most of the time they used what you would called subway tocanization and the form A practical standpoint for this type of recognizing this standard type, you probably want to use spicy or Ginsim to do that. Those are currently standard tools that have like these standard recipes for doing that or Psychic Learned Tide Factorizer Back to for this type of toconization you want to use the Hugging Face toconizer. So I think after you are ready introduced that hugging base system in the tax session. So we will be. For those of you who are doing transformer based help models using context sensitive help, context sensitive embeddings, then the hugging Face Stadium is something that you will want to invest in in Learn Houdworks. So for example for Bit which is that is guess to' kind of one I just be the work horse for short document in all using Transformers rather than a vocabulary that includes all words, they have a vocabulary that includes subwords. and so for example you a word in the best vocabulary. it could be three words or four words. So here you it says playing could be reduced to play an wing. And so that's why it's called forward Tocanization because it would take the two world pieces and treat them as two separate words in the vocabulary. We'll come back to this. So before we're getting to this token representation or either list or or counts over tokens, the first set of choices that you'd want to make in developing good data is preprocessing the text. So there are many steps to this process potentially. And as it mentioned, there's a number of recidities such as a physicist learns stiff facterizer or Chinsim preprocessing or or some of the space functions that will make a lot of these decisions for you. But it's important to think about these because before example, whether you move, capitalisation or punctuation could make a big difference for your downstream outputs depending on your task. So the first usually is taking full documents and splitting them into smaller pieces so you might want to take the one that often uses letting into sentences and as a mention of a you have this task that's doing pairwise comparisons between a million sentences you have had to get the sentences in the document first right and so spicy I think would be a standard sentence solution. so when you input a document into space it will immediately do the splitting and it will adjust for nperiods at the end of more or messes or abbreviations on us a porch things like this and you'll get informative set of sentences till work. A lot better than just splitting on periods or full spots. In terms of splitting paragraphs and documents, there actually is no standard solution to that because that's going to be a quite domain specific how paragraphs are split. If you are using hotel there will usually be in be the pop tag or the bar tag and if you're like in digital documents then you'll have to have a costumed solution. I will tell you that for most digitized documents like over a line that ends with the period or it full stop orquestion market explanation point that's almost always the end of a paragraph so it do not be perfect but you can use that as a short cut to split paragraphs using our data. Just align it into the period and you know part of this breaking up process. Preprocessing is the idea. Threats is something that will repeatedly come back to is that unstructured text date has lots of noises, is a lot of extra information that is not useful and so we need to develop our pre processing and featurization steps to extract important information and exclude the irrelament. So of course theatres many papers about this, but the dining supporting paper, for example. They undertake a number of systematic investigations of different pre processing choices and especially for unsupervised learning. So remaining a topic model or clustering things like this, the preprocessing makes a big difference. Fortunately, this paper also shows that for supervise learning, classic machine learning, classified fiction, and regression, the pre processing choices do not matter as much as long as you have a sufficiently rich representation of documents. Text class birds will work well, so are choice is whether to remove capitalization and so usually the capitalized and non capitalized version. This is everything that going talk about is mostly about English, but you can imagine that there is going to be similar or parallel issues in garden or other languages. The capitalization issue is think even more nsomekind. It's like more interesting in garden, but in English of course knows are only capitalized at the beginning of a sentence. Risk in titles and things like that and so often times you increase the size of the feature space by treating capital letters differently and it's usually better to remove the there are so many exceptions to this rule. So in the legal context you can think of the First Amendment having a capital effort. Capital A Three is deteriorating about American law. So the First Amendment refers to Freedom of Speech and religion and things like that. And if you read the phrase of the First Amendment without capitalization, you know that they're talking by about you no specific law or specific contract or something but have the capital to capital as they're talking about the Bill of Rights to the U's Constitution is and so that will be an example. For legal documents, including capitalization could be pretty important. Also you of course if you're doing linguistic initiation of documents like part of speech tagging statistics, passing semetric role labeling, capitalization is really important for that. What causes what you might often have is the source documents are not capitalized correctly. So in the case of contracts for example, you'll often have sections of contracts that are all caps that all capital letters just to lie, highlight them. And and that's a problem because like things are a part of speech tagging and a synthetic parsimple break on on text like that. so you might need to do some custom checking. This is a nice example where punctuation is quite important. So I got these from Chairs Bail's slides. In general the rules about punctuation whether you should keep them or not. It's kind of the same as capitalization where usually they're not very informative if you have immigrant representation of your documents. but if you going to use your data for some linguistic invitations such as sentence splitting or part of speech tagging things like this then you need to keep the punctuation and information in your doctonants similar for numbers. Most of the time you should remove numerical information from the document just because if you are breaking up your document into matrix accounts is just counting that. How many times the number one mention you not going to be very useful or especially if you're counting like the number of times the word the number nine hundred, nine, two thousand and one hundred is mentioned that do not be very informative and no would say replaced numbers with a special character like a hashtag would you be a decent approach for language models like get two gptthree bright numbers are just treated the same as any other sea are and so they'll be part of the subdued tocanizer and will get to language models in weak nine. but this these big language models like God there actually can not solve mat problem. So if you give us some tiny pleas to plus two equals for two plus seven equals in nine things like this and give it a new problem by will actually often times provide the the correct inverse. This is particularly why we're doing this kids right now because this amazing technology of language models is transforming how lawyers and social scientists can use language as data and on for practice. But the this really exciting and intriguing finding that language bottles can solve math problems. It means that they're starting to have this kind of conceptual representation of language under the hood, but it still does not work very well. So it's easy to find math albums that get there does not can not solve. And so this is an active area of research and there are many projects to do for having language models understand numbers. And as a side note to that, having language models that can do things like fact checking for example, scientific claim verification and having them understand numbers is going to be a really important piece of that. You can see how this is practically important. Dropping software is a similar type question as punctuation and capitalization. There's a standard list of words that show up all the time in English, but do not really mean that much by themselves. On the other hand, it's easy again easy to think of counter examples in the word not is often troops a stop word. But if we're reading legal documents and we're trying to say you know what is the judge deciding having the phrase not guilty is inappropriate to include right and then just more generally in the law and in other technical all domains. Oftentimes specific phrases or means are going to have an independent and pivotal meaning beyond the words that are composing them. So like the classic example in America in law is beyond a reasonable doubt. This is like a very specific evidentiary standard in criminal cases. and the words beyond reasonable doubt by themselves. If you counter those, those would not be that informative and with all deliberate speed. That's another procedural phrase that's very important in the U's law. And so you can imagine that even though those those phrases has contained stockworks, right? So if you drop stockwords before doing in gardens, these would not come out in your future representation to one option here you're practically speaking would be to drop software when they appear by themselves so you do not care how many times the word a shows up. But if they show as part of these phrases beyond a reasonable doubt, you would keep them in the future representation. Another way to you refine the corpus to make it more informative is limiting or limiting. and so rather than include every word that shows up on its own, it could help to drop surfaces. So in English this would be to consigned consigned, consigning consigned with all four of those things are talking about the specific word route consigned and so applying a swimmer will remove all all those suffixes following a rule based algorithm porter sometimes I think we pretty well in most cases and there is something called limiting as well which spicy will do for you which rather than just split take off into the sentence, it will actually look it up in a dictionary and give you the world route only. So this was going to be designed for the hybrid zoom chat but lets lets just do this in person instead. So just to start practicing with these issues of of non word or style features in language, consider the following four dimensions of language based on the first letter of or last name. So read either if get disturbed based on your last name, think of a social science analysis or important legal or social dimension for example or judicial opinions for newspaper years for social media for political speeches, think of something that's interesting to you, perhaps related to your work or other things that can be either can be measured by capitalization punctuation would change depending on the use of software or we change depending on the use of simulator limbertizing and so we're actually at the the point of the break so is Let's think about this now and you can take a few minutes but also take a break and we're going to resume and a fifteen minutes a ten actor the hour and will give some examples of will have people to volunteer some examples for these categories so will see you in fifteen minutes stmepo degree to at it on the alright we're going to resume now so can I get a couple of designs from those where the last name we direct for something that can be measured with capitalization. Yes annual air last name starts with off someone who thought about capitalization being a relevant legal or social science point to deal animal which for you and do you want to start us So I'm asking for some volunteers for these four groups. So what is your last thing to start with? So great. So you are in the first category then did you think about this you come up to an idea for something social science on social media, law, political reached where capitalization would be this important dimension. Let's okay we havepouned the back here yes at and that's interesting right? So you can imagine you people who have the more years of schooling. maybe they have a longer's longer a more complex language it has or fewer periods but more cocombas for example interesting And the abundance of pertulatory events using doctors bacors be cause more along plots heat this shorter grade umation interesting right? So you can imagine like in liked transcription of patient's speeches. if there's more like dots that they're breaking up their language more because thneither'r taking more extra both is like that. That's interesting yes and at often the J owl are done although it are a simple passport and analysis version by the biological and that so just an etaspoto to the court other is so and what was the role of the porter semester in your approach analysis there can be built that would plate time analysing, forge in on and so porter sometimes I guess it depending on what you're in your context. the swimming was worth while because the world endings did not have as much information is needed. Is that fair to say He I see thank you any other exact yes constitution age weights that like the First Amendment ex ample or like you's constitution article there and do not be among number is for example woman numerals are often capitalized and so like luck it is going to be different than just looking for lower case and so interesting at at to can topecifoe report on a rise to iyaitionsa to get and of new truly vote changers sometimes right But I think that ye turkish and maybe Japanese like this too much to but through examples sometimes like a whole synthesis in one word right like you could have like addition or her like tin surface saves is one thing. So why you won you very careful what's standing in Turkish so I think that's a very good very good. totally once here and or son the second cussificationak shout interesting Totally right yet so like tweet the yealternating fur lower or K R I d' I did not put it here but I guess you could include a mode jesus punctuation as well about thinkffor a syntomate classieria all cap it's going to make a big difference. you have a well at of routine on tree was root seeing airport announced special rises of interesting totally right. So like all naninnity recognition for example. so if you want to who you let them second but like if you want to in English especially when things are capitalized they have some kind of special properties to get them of personal about. for example procedure for it is important another point there yet he and interesting right? So that's a interesting thought right? So you could imagine like there could be some systematic methodical way to see know for what words are, timing or limiting capital that's a good project idea potentially to right. Let's move on. Thanks everyone for those on the zoom if you want to add examples please type of Imachap. Also now we're going to move forward with recognizing the most basic unit of representation in a text is a token. So thats what were going to refer to as when we break documents down into pieces those pieces usually words are tokens and so you could imagine. One simple way to do this would be you represent documents as a list of characters or letters. The standard approach would be splitting into words so you go to like a split on the space and then the third would be immigrants or phrases so we capture local world ordered as it for example with direct becomes with sir So when this is like the kind of classic workhorse in help would be a backwards representations. Basically just a corpus is broken down to a vector and for each word in the vocabulary you also get a number corresponding to how many times that word showed up in a document. And so you know just to introduce some languages here which we'll use repeatedly. For a given word in a vocabulary we can say we can have the document count which is the number of documents that had that particular word tocentype appears in a in the curbs and then term count would be the total number of appearances in a corpus. In then we can define a term frequency within a document as the number of times a token type appears in a document divided by the documentligso there's the number of words in the document and so sometimes that we might be a bit imprecise, but going forward if you tried to use the word counts to refer to integer, the number of times it competent occurs and frequency will be the share are the count divided by the links. So as an application of this in the political realm Monro and all or this is bighton words paper which is linked on the bibliography they recognize congressional speeches and ten identify words that are distinctive of Republicans and democrats in the U's Congress. So to do that first they run a topic model late to reached directly allocation which we will talk about next week to identify speeches about abortion, reproductive rights. So that's what is. They use a top model for purposes sick so this is it. Think pretty useful in the congressional record because it's very diverse what they're talking about. Sometimes the speeches are just like procedural there is saying we should not on air, not on the and those are not going to be very interesting politically and in the paper that they provide a number of ways to identify language that's distinct five of the different parties. so you know the Democrats at the left wing Republicans are the right wing party. This is from around to thousand and five I think they hundred six Congress and so you can see that when you just looking at the difference in the proportions of how many words are used, the most democrat word is to do it's a stop word and the most republican word is the T. And so this shows that this very simple metric of just the difference between the frequencies is not. There's not to a very good job of extracting distinctive tycoons. The second approach that they do is compete the law of auzry. Go for the words. This you would think this would help a lot because as racioit adjusts for the proportions of other works that are used and within this actually or living works. even worse because the most democrat phrase is bankruptcy which has in snow which have nothing to do with abortion right. So this is not extracting an interesting dimension of partisanship. Then you can look at the paper for the statistical the mathematical specifics of this. but they provide this interesting multinational lesbian model underlying language. If they specify that model and estimate the associated parameters, you get a much more interesting ranking of the words. So for within the topic of reproductive rights, democrats talk about women's rights for example their decision up to the family whereas republicans are talking about abortion procedure and killing babies. So you can really see how this kind of like difference in the difference in the framing of this topic really comes out once you try to these other token drinking methods and then finally this one they had a regularisation parameters to really to shrink most of the word premieres to zero while still maintaining this distinctive language. Yet that's great. Yes so not now. and I think that's a very good question of this specific ranking of the terminology and their sporting well in this context, but it seems quite specific what they tried so we actually do not know. they might havetried a hundred other things. So in terms of exciting we do not want to draw from this we sold always apply this method in every context, but in in this course you you know when we're when we're doing the required readings, presentations who were running the response essays. That's exactly the type of question do you want to ask the morons and their co authors they wanted. They're also probably selectively presenting this right and so some other questions would be used so I didnt even include. that's a good one right. Will this work in other domains? We actually do not even know if it works on other topics, even within the congressional record. They only try their reproductive rights at the abortion topic and so get these nice ranking of the words that we kind of know intuitively that if we want to extract this of slain dimension of reproductive rights discourse that it worked. but if it was another topic where we do not know such about, how would we know to trust the results So so waiver this is kind of. It's lucky that they chose this topic because the results are indicative. So in another domain where we would not have a lot of previous knowledge about the topic in mind to work and we might not even barely equivalent in another question or comment about ewhyheard some words so out of that supporter batter. So when you do a porter simmer on words it will replace them with these kind of placeholder suffixes. So words that end in why given ending in eye. So this is because like baby or bad they both become paid and but that's a good question. Do I about the presentation of this right? Maybe they should have used a limitizer instead because it makes it were to and of have to read the paper to even know it. but that's a fact. So I think and that's kind of related to you known they were saying that the difference in the frequencies do not work. The mean if they had just dropped stock words beforehand this would have looked okay right because they just dropped software. Then you can say women's rights their the Democrats abortioned baby procedures and publican that would actually already looked pretty good. And so this is the type of kind. Like what else could they have tribe alternative methods to get it the same place? That's way we want to start speaking about the you mean on horizontal axis it says frequency of word within topic. So if I think it's just how its the log of the of the frequency of the words. So but I Do not think it played a big role in our analysis. I Think it was just a way to add more space on the graph. So let's not road these types of questions. Okay, well, you know, is this figure effective? What else could they have done? What information is being concealed or left out? That's what we want to ask for. All these applications appears okay. So building a vocabulary. As already mentioned, this is basically the goal of frustration or tocanization. We need to think about. Well, how big is the space where our documents are being represented And so no. there is number of ways to do this and depending on your context it might not be do not afford it. but there's just kind of simple historical that usually work pretty well. So like any word that shows up in less than ten documents, for example, you know it's not going to be very informative for any analysis that you undertake, so you should drop it from the clear. probably. You could also impose a more complex threshold. sigh you not needs to appear twice in over twenty documents. I Think there's kind of limited returns to this, but I Like the ten document minimum or take the top twenty thousand most frequent words is usually a decent approach. A more subtle way to rank words is in the frequency is called to term frequency inverse document frequency waiting. So what? this does is in addition to avoiding information about how often as a word or a phrase shows up, that's weighted by how often it shows up across document in the whole countries. And it's inverse document frequency waiting in the sense that words that show up in all documents should be downweighted because they're not very informative about topics. So if this will kind of depend on your task. but let's say you you're trying to classify documents on breeding topics or you're trying to learn topics accidently at a topic model. The words that show up in every single document like the or a those are going can be very informative and inverse Document frequency waiting is a way to address that. And there's some example text at the farm on this slide to get a littlebit of intuition, but basic by this will upright distinctive words that show up often, but not in all documents yet. I Think that they have like Enpsycli Burn for example you would put one plus they smooth. there's a good stomping farmer btyso this is like one example formula but as you said it would be undefinend for some works this way. but they they add might made partner in this danger of implementation. So this is the the psychic learned implementation that is talking about. So there's a few different options but I think in the default one it's one plus the log of the document count. To address this issue. The intevenominer. But in terms of life for most tasks especially for doing it classification task this is your friend of if your corpus will fit into memory run stiff vactorizer on it to transform the words the plain text to earnings. When do you're ready to go you have a data set already that can be used for anything and so in terms of the preprocessing options it will actually already remove accents. Things like this remove capitalization and drops up. Words is the question at so the the idea Let's say let's see how this a regression model and am packed at up. Let's say you this will make more sense when we're doing were like comparing documents so let's say it will do this next week. But you want you want to compute the distance between two documents and if you when you vectorize documents as frequencies over words you might want words that show up and all documents to cut less and so if if you use this transformation you're basically re scaling the dimensions of the base so that words that show up in all documents means there not very distinctive you downright that dimension and so he think it's an important question because the while this will matter all for concise coinkedocument distance for other downstream tasks that actually do not make a difference and lets but I want to come back to your question next week. Buyout On think of it as it's releasing the corpus so that releasing the corpus representation so that dimensions are words that show up and all documents count less in those distance calculations. So stiff Vactorizer the is waiting. Its its optional so you do not actually do not have to do it and there is options for something. The if waiting on things like this time I don.' There has not really been enough kind of systematic comparison of whether if waiting helps or not but this will like only be task specific. so there are this. representing documents is like counts over words or counts over in brands is really the standard thing to do. Counts or frequencies and you can pick up an infinite number of other things to do right. You could say well in actual I Want take the log of the frequencies or just an indicator like for whether the phrase appears or not. Quadratic paradise interactions these winners often not done because of any given like text classification or top of motor polemhis just adding dimensionality without helping and there are not any kind of rules of them for beyond getting the ignorance frequencies. what else? What other featureization we should travel. I think immigrants is the way to go. and so just to talk about immigrants a bit more, These refer to converting lists of words to lists of phrases. so you know for this is a sentence, you have a number of bargains like this is is a sentence and then atrtrigrams are the three word fishermen. That sentence it is a sentence and this is useful for oclassifiers and things because it captures the local word order and I mean you can. This is a sentence is not a very good example, but like we talked about, you know the word tax versus tax cut or tax credit. You can emit like these two word phrases really in of a lot of important information in legal and other documents. and so this is why you normally use infants. and there's this link on the cillabus is that this is that code may be a little bit out dated now, but I think most of it is probably still accurate with this Google developer's text classification guys. They tried perturbing some of these different choices or use of standard text classification tasks and when you have long documented or relatively few documents then stiff weighted bargains is the base lie. They recommit which you know because there are so many different things to choose. This is a very satisfying to hear. I Think there really simplifies things that if you and the even they even give you like this. these told that if rose the number of rose divided by the document length is less than fifteen hundred. used tide water bridges and so they tried like words for his characters immigrants versus sequence which we're going to talk about later just bygraundi usually enough and then if I did and so this simplifies our or task. How many managers do you need in your vocabulary if you imagine that there's a hundred thousand words in the English language, the set of possible toward phase hundred thousand times on hundred thousand times to and so that is not going to feed in memory prison. The same Google Developers guide they recommend picking were thousand bygrands based on their frequency which I think is actually a really decent rule of them And so in the physicist learned stiff facterizer you can just say I want bargains and I want twenty thousands of them and it will do it automatically based on frequencies. but even twenty thousand any thing is in marriage cases is more than you need I've just like tired this in different applied classification tastes and even like to thousand features is often enough in adding more as diminishing returns. But I think that and well talk about feature selection and a bit for the if you're doing classification old let's sixty thousand immigrants based the frequency and a du feature selection to select predictive features down to ten thousand. A totally alternative way to get at this issue of high dimensionality is bashing bacterizer at E all year. Good question, right? So that probably seems a bit inconsistent with what I said. is it can to go about of painting rat? So what I would recommend is in general, you should include them frequent including those frequent words and what it would normally do is any words that show up in more than like forty percent of documents those get dropped and then after that take the option. Thousand Bees formulate this. Another historic would be drop the two hundred most frequent words that is usually actually a r by document frequency. So that's like kind of a representative of these southward usual words that are not very important, the most corporate, drop the top two hundred in the ten thousand flat. So but I think that the idea is that are the most informative words are the most frequent order, not that are formative and all the very internet words are also not that informative. So in the middle and a year we also look out at the dipran Yes, so there will be very few of those. so to be a bit more precise I would say dropstopwards and words that show more than forty percent of documents then make bargains from the remaining vocabulary and include the top to on tousan on that probably has. Not only are when we get to partings ansyntact and sometimes is to date because the right I show you an example earlier. Beyond a reasonable doubt that's like a forward phase that contains a sword but you need it to get the meaning, let of that. So these are kind of rules at Dumb and and it's if important to think the ways where they find not work for some cases is her on what performance did when need breeds or so the best. The most systematic one that I know of is this one Here is at least for text classification right they found they did is own array of stated text classification datasets and for the biogramds work as well as strike so we'll see an example later. This gains crown should appear at two thousand and ten paper at least in terms of interoperability like kind of the quality of the features. The programs are pretty nice in terms of getting phrases that are saved like in important narrative or political information in them. The programs were nice, but I think that for you just the quality of classified bygrands are normausually as good or there's diminishing returns as you had not at ignoring pills but another way to kind of dodge this issue is the washing vacterizer because with a washing vacterizer you can add even an arbitrary length of immigrants and they all get just mapped to an arbitrary idea. So has anybody heard of this before today So you know washing is this you known way function technology thats often used in cart criptography Right where you say basically you can take the string output and output a lower dimensional string and there's no way to get back to the original string. But that function is deterministic. So basically you can imagine making string and mapping it to a number. That's what a hash function does. And so I can build a hash function that has ten thousand buckets in it and any string will be randomly matched to one of those ten thousand ideas. But once I have the washing function, it's deterministic. So once if I have a washing function, I can then vectorize the whole corpus to a vector of ten thousand items. But what those individual strings include could be anything. so it could take words, Bigrams, programs, and quadrans. All of them get mapped to into this ten thousand dimensional hash space. Kethis is the illustrating it here under traditional vocabulary construction and specify the number of everywhere. so bell cats and dogs. but with the washing trick, they just get a random number. but that number is always the same so it's comparable across documents because there's collisions. So just by chance, the same number at the end of the chasing function represents multiple words. so that's what this is trying to to represent here. This is called a collision in the sense that if you only have ten thousand items in your hash function at in the output space, but there's an infinite number of strings in the input space so You can imagine that if you have ten thousand items in your hash function, two words have basically one out of ten thousand chance of being up to the same number. and so every when you're going backwards, every number at the other hash output space represents a lot of different strings. right? And love. Tune in your purpose. And so that's why the fact that this would never happen in practice, right were two or superrare that two important words would be mapped to the same idea. The way that you address this is you can actually use two washing functions And so if you use to washing functions is in the rare case where two important features are confounded in in the hash output, it will only be for one of the washing functions. enough in output. Yeah, there's two reasons that's that's by the main one is that it works for any vocabulary. You do not have to build it a crucially so new documents, it will vactorize them correctly and the other is so you do not have to have the purpose a head of time the whole purpose. So that's Ikapty. Very useful and it helps with dimensionality. So like you, if you want to use triggers for example in your feature or quatrans in your feature space, there's meusilians in them, but that will not increase the size of your questioning. A passion bactrizer. It's also very fast, like the computationally so you guess might have heard of it. The text classifier that Facebook uses it's called Fast Texts the Washing Bacterizer I be the celebrity compound technological in the fast Text of Text classified algorithm. So basically what it does is it takes documents and represents them not as the list of word ideas, but a list of hatched in grant ideas. So you get a bunch of these hash numbers for every document and those can be very quickly computed and can input into a classified so ththere'there.'s few practical issues, but also complicate ya, the can is alright. like to about isinin outfits about grids, the w he ears there or do to say ye let me plazamia That's actually nice analogy, right? So basically even in the standard model language for the same in type you you will have multiple words on it anyway, right? And this is just adding to that. Making it slightly is making it more more crook, right? The court's representation: Multiple meanings or words are loaded on to the same string anyway. but this actually just takes that a little wooden worth and allows for even phrases to be loaded onto a specific string. Yes, so this fast text model is. were going to get into this like week as or so. but like the the Fast Text Word Embedding model or Document embedding model uses passion processing and so you can embed the washing ideas the same way that ipludin bad words. So I would say it. I Would say it's under explore though so there are a couple of papers doing it. but I think that this is a very interesting technology that has not been not that to be I think in principle yes but I mean if all you if all you had was the hash ideas it would be difficult to check this. but during the corrupt verification process you can keep track of swapping alongside do things in is and then one is. I think it's a little bit unpredictable but would happen but at a minimum it might add anyone to your embedding tribe because there is going to be some other words that are included that are mapped him to men and woman's the she site you like. it's interesting. I mean you could probably do some of this theoretically like the distribution of world frequencies like this on the world frietsay who get sagazip's distribution and I think in the fast text model they allow for like a million stews or something. so Whether using embeddings they're using this hash fertilizer as well as allowing to for a million different spots and so in that case like collisions are very rare but I'm sure you could work something and based on the distribution of frequencies it would right? So yes so so I mean especially why no if you're talking about the original fast text vactorizer on social media data so you know they are like hundreds and millions of documents and sminicalliio and even for strangers or quadrans. So the that's an interesting question. I Do not think that they get it that directly. Was there another? Yet to be sure that as the stand directors of Poet, traditional economic construction and artists basic met one producing yes, Andritan a permit evident yes Exactly as so. This is great. We're get in to that like week six about getting into to embeddings. But but the standard toconizer will give you and time with a list of words, a sequence of words or account over words like that. One of the atectet is on one hop in coding or for account in coding and that nose could be limited. so be the. But you can imagine that a document is just to list of adhesives as well. so it would be the same thing you could apply an embedding look up to like a list of words or a list of ash ideas. Okay, him going to speed up a bit and then we so we can finish the next week. So reaching vactorizer is one way to deal with hydimontionality, another is to filter the vocabulary based on your downstream task. So for example, if you just take all of the world frequencies and get the correlati with your outcome variable. so let's say' we're like in the morning at All the fighting were its paper we're trying tobasicaly that was kind of feature selection right? They were trying to check which features were correlated with the different political parties and so there is a number ways to do this. but Key Squared for example would be like a classic feature selection approach. Let me think about what I should cover now. So one issue which were going to come back to the lot in this course is relating statistics to meta data. So let's say in the fighting words paper A you the number of Republicans is increasing over time, but also the frequency of the word kill is increasing over time like over like a twenty year time periodand you would measure this correlation between Republic and a kill. But it's actually just a spurious correlation that if the word kill is increasing in both parties and the number of Republican is increasing. So this's like a kind of a classic correlation virus causation issue. You're actually not estimating a partisan dimension in text yours getting this time confounding trend and this is something that you nwits not recognized very much an nip, but when you're started applying to kind of social science issues, then this is going to be everywhere. And so rather than just estimate these rather as do feature selection based on the rock correlations, you might actually want to try to deconfound some of the frequency voters or at least the outcome vector. And for example, if you deny all the word frequencies by year and then estimate the relationship to democrats and republican, then you're getting just a within year variation and excluding it this time for point. And so this is something that will will continually come in back when we're trying to get up more casual or more informative representations and correlations. This is just like some side note is that you might have. you might want to get rid of both the time component of this correlation but also the geographical component. So link to Congressmen from taxes. for example, they might. you might have the word congresor from taxes might use the word to kill up often times, right? And so in order to remove both the year and the state variation, you can use a literary reason to do that and for why it would be the outcome and x would be the frequency of each word. You redress that against less categorical variables and take the residuals and then use those and your feature selection or machine learning to bask. So that's done. Found did feature selection. We're right at at four o'clock now. so I want to wrap up. And also for those of you who were doing a project I'd like you to stay for five more minutes to talk about it. We're going to write But now and I'm going to finish these slides at the beginning of of next time. Are there any questions or logistical or otherwise before we wrap up right? Thanks and will see you next week at the same time here in the same room. If you're interested in doing a project please stick around for a few minutes. Sigiisi A but cut in be on the tide process a holeayat at a of of more is is said to always if you have to wind up where the not object is too yes please he So if you have not sign for a peak yet and your he man to hear about it feel free to is to cart but you just go to hear about it. So if it's okay I'm just going to start talking and is freed to go yesterday. There is an additional to optional course credits for a course project. It can be done individually or in groups of up to force students and it's really just doing an application or a act based on what we're the continent of the course. so it's quite broad in terms of the topic. but usually its an app project on some social or political data or legal data but we're quite go and you can choose some thing and you're interested in. So just to give you some examples of the previous year's projects, so one of the top a better startup feature Startups: the Deep Judge team. Their project or their system started as the course project in this class and so they actually built this context sensitive legal search engine which's pretty amazing and they went on to get some by funding their headquarters at the I center. Now another group did some environmental regulation analytics and they want it in a swift grant. So just to show you how successful some of these projects have been, a number of them have been published and one on doing legal language modeling, another on inspiring legal language, another on medical documents summarization of medical documents, one before biting with a old student. Here he built a basically an automated question answering system for a coding class and then no Langu who's also a old student here publishes an attractive summarization system using reinforcement learning and so those last to those are individual projects. So even if you do a project by yourself, those have been in just being successful. There's a number other projects that that I think they have good chance of being published at some point either an top conference or a social science journal. We made a partisan treat generator when standard. Stump Bombelli who's another students here. She did an analysis of immigrant attitudes and historical newspapers. One group did and deep it like normal nets and instrumental liberals. The kind of caused a little of paper. One group did parties and question answering. it does not have to be text. One of the students did an audio an analysis which was fine. if we want to do audio or other images we're not in cover that in the course material. but you're welcome to do that for your projects and then some of them have just been kind of classification projects in terms of picking a topic, You're welcoming to pick one on your own and I can provide back if I think it's a good topic or not. for how to modify it, some of you are ready. Asked about this. we have a list of suggested topics like we have a short list of topics that I think are kind of interesting to do right now and then a longer list if you want to see that. just send opera, email wearing people, less economist everybody list of the same time. With some doubts it would also be good to think of which maybe one or two or three topics you're interested in because then I can provide some thoughts about advice about what's doable for course project One you take a look at these slides from the gethub, but once you formed a group, send us a list of their team members and it be useful. You do not have to be right a Cv for this or something, but would be useful. to know what courses you' we take in and so that we can set expectations and things and then some of the projects have either new or after we advise them but there's a number of other project advisors that will help out this. So for those if you need help picking a topic we can need to disc test. maybe it could be Onzumor in person and if you already know your topic we will also need so that I help you get start. So those are the points that I wanted to to bring up for Now you do need to say you do not need to pick a top topic for a month a few weeks from now so this is none of us urgent. but I'll send the list of project ideas A we go from there, are there any questions or concerns at the moment A right to keep the such with any questions and will see you next week. Teach changes are sejetd. |
| 3rd row | I think we do reading as started. Thanks everybody for your patients with the hybrid learning experts. Just we face a bit of a tradeoff here. where in order for me to see the slides here we have to have those kind of zoom overlays and rather than me having to kind of a switchback, keep looking back and forth thing mass iksher lwilld but better so we'll try that this time. I Looked at the queue and a page this morning I Did not see any questions, were there any that have been added since then anybody wanted to to bring out loud off for those on the home I Can not see the chat so if you have questions type it in there and African can let me know. Do see any questions on there yet? Yes sure yeah each. so it did not see any there this time but yet we have Consimantos in the mood. swell do that going forward. So we ended a bit early. we did not finish the slides last time so I'm going to go through the rest of the week two slides today at the beginning today and then we'll go to the week three slides in a few minutes. We are in the middle of the discussion about immigrants which are perhaps so phrase representations of documents and why are used phrases is because in English as in many other languages the meaning of a phrase is more than some of its parts. A There's and honest clear examples of this are what you linguists would call collections such as kick the Bucket. So these are these sequences of words that you can not just say adding kick and then and bucket together here in any arrangum order is not the same as kick the buck at this particular sequence and this is what you would call on compositional non substitute So you can not just like swap and no synonym and maintain the meaning and non modifiable so you can not. These collections do not respond to the grammatical rules the same way as as normal language. So to get at capture collections in language a nice way to do that's a positive mutual information. So basically two words in a sequence A and be. If they are independent, then the probability they occur next to each other is just the the product of their probabilities that they happen on their own. So that's the denominator here. the probability that each word shows up in languages like me, the corpus frequency. If you take those two words independently, then this the numerator here. which is the probability they occur right right before and after each other. It will be the same as the denominator. And so if you compute this metric or point living mutual in motion between two words in this specific bathroom sequence, you can really rank the words in the vocabulary by their collocatiodness. They're how often they can relative to who occur a part. So I think thisactually makes for a nice feature selection approach. We're tough ways to reduce vocabulary. One way to do it would be on frequency for bankers and trainers. The pointlized mutual information is astonished to it because you capture these distinctive phrases. So back to to a social science application of of distinctive immigrants. So whether we are talking about feature selection approaches supervise feature selection is a way to identify words that are related to some metadata and in the morning at all Fighting words Paper We looked at words related to reproductive rights that were distinctive over republicans and democrats knocked and Chipiro to thousands and ten is a paper from media economics using that same type of machinery. but for a social science question of what drives media silent newspaper media and so what they do is they build this parallel corpus of next from you's daily newspapers as well as the text of political speeches from the U's congress and they have doing some pre processing So this is why the this after unharmed so you know more was looking at single words Ginscoach appear are looking at bargains and trademarks or two or three word phrases and they then use a related metric to mnmonro all do to identify polarizing phrases. So what they do is for each phrase you you have some frequency that a word shows up for democrats, a frequency use of the Republicans, and it a frequency for other phrases. and they compute this Kysquared metric for each word, which is the same as the key squared metric that you have Gipychit Learn for example for features, action. it's just the latest statistic for the equality between the political parties of using that case. Is it better? Much better. So this will be easier to read now, but still probably too small. Check out the slides if you can not read it from where you're sitting. But in terms of using these nip techniques for social science, something that readers will always ask for is some examples or some way to do some qualitative valuation of what the system is outfitting. And so here they show the lists of phrases, bargains and programs that are most associated with democrats here on the left and Republicans here on the right. And if if your European or your switch you I do not know that much about the you's politics so this might cope unomoush to you. But for those who are familiar with the Us' political setting these phrases are kind of intuitive and for four republicans for example used well site cell and around this time there of her thousands they were talking about ban and to his self research because it was time or in minority time cells that tourist reality reproductive rites and they're talking about kind of legal issues and government spending like this is in part of the Republicans being worried about about expenditures and Democrats on the other hand are more worried about the work war. So so this provides some kind of just qualitative evaluation that the distinctive phrases for democrats and Republicans they make sense this like way to add confidence or will trust them at bit more And so this is the main result from the paper. We can not see this, but this is just scattered plot of the newspaper slant so computed using this those distinctive phrases. So how often does a newspaper use republican democrat phrases and then the horizontal access is how many republicans there are in that in that congressional district. So you can see that this is consistent with the demand side media slant. So rather than you you get a republican, owner of a newspaper takes the newspaper republican. The consumers for a newspaper market are more likely to be republican and the newspapers respond to that by using more republican language vice versa At yes so this is it could be that the newspapers introduced Republican language and then over a time people became more republican for example Than there's the short answer to that is they looked at when the owner of a newspaper agent and it did not change the language. So that's that's somewhat imperfect response to that question because there's other research showing that when the newspapers do change their me as people do are persuaded by that allowing talk about an example of that mix to be in to actually know so one other just kind of like parenthetical feature to detail that I want to mention it is named in the recognition so these collections that is mentioned of low kind of high phrases with high prime phrases that tend to do locate you will get these nice phrases such as this order or meriquery which those words that that phrase is quite distinctive of those individual words. But if you want to do if you want to try to capture proper bounds or entities did a name indie recognized is the right way to do that is so space for example. it will have some examples in the notebooks where if you put a document into space it will identify different categories of entities such as people, organizations locations and so this could be useful in a range of things. So like for example if we look at news articles and you want to identify articles about a specific story then getting it overlapping entities would be the right way to do that. So you know another kind of shortcut to getting named inities would be it actually just like search for proper bonus which is a part of speech part of speech are kind of the basic functional categories for works in attendance. Solution is an object of verb is an action and adjectives and attribute and so on. And the the standard part of speech tag set in English has thirty six tag once even there and but otherwise kind of universal tag sets that work across languages as well, but kind of. These basic functional units in language are pretty universal across languages where a noun will refer to an object. So if you getting a topics in text and looking at bonus who make sense, adjectives correspond to attributes and so if you're concerned about the symptoms in the text to maybe you want to focus on adjectives and that's parts of speech. So okay, so one question you might have on to practically is why on what to do with out of vocal words So were not. We talk about the washing concrete last time. This is a method that we out. Vocabulary words are no problem, but if you want to limit your features that to ten those fifty thousand items, there's going to be a lot of words that that are not going to show up while you take it to new data. So what to do? It does. Think the standard is used to drop them. That's what psychiclearn Tfidvacterizer will do. By default, you could replace with a special unknown token. This is what is done in modern toconizers such as the housing face town years. You could replace them with their part of speech tag. That's kind of, I think a nice compromise or you could actually just have again auxiliary chasing vactorize for those out of vocabulary works. something else is thought. Bone was right in these slides. is that it you could replace with the hypernym from word not so like why if you see the word you no trout you could replace it with fish something like this. So another social science application for the specific case of parts of Arts Muse tags that is found is this Netsmer, Lamair and Person seems two dozen into paper windward sweat. They find that and the words that people use in loan applications are predictive of whether or not they'll they will pay back to loan and so maybe you guess withered this light already. But imagine that by the website exists where they have people like protein peer lending so you can actually lend to strangers on the internet if you want and one of them says I am a hard working person marry over twenty five years I have two wonderful boys please let me explain in why I need help? It would use the three thousand dollars alone to fix our roof. Thank you God bless out and is promised to pay you back borrow number two days. While the past year in our new place has been more than great, the roof is now leaking and I need to borrow the thousand dollars to cover cost of repair. I pay all bills eg. Carloan's cable utilities on time. So which of these borrowers is more likely to default Razor hadything its borrow number one of number two. it's about equal. may be a little more borroughor number one, but it's the case as borough number one actually is more like a default. And so in this paper, they have this really amazing data set of these types of snip bits and then date on where the people paid back the loan. And they find that loan requests written by Dealt and Bears might like to include words related to board's family, financial, general hardship, mentions of God I Think that's superinteresting the near future in pleading for help and also using verbs. In present, vendors feature tents so you can see that like they really are. These kind of tells in language for whether you pay back to eleven or no. And so this is one interesting craft that they show in the paper. Words like Reinvest Wedding student learn summer side. These are words that if you see this person is going to pay back to lean, they're pretty reliable. but then words that predict default actually he go. God and family is like a category if you see that in a lone application, you should stay away from this purists. which is maybe it makes sense in retrospect, but you can see like you know God bless, pay day loans, medical death. I need help. Let me explain the situation and these are pretty striking. And like this figure, they I think they used a topic model for this, but they had a way to classify words or documents into topics and these are the words that are associated with default but also typically related so they tend to show up in the same sentences. But they did that based some automated method so we do not have time to do this today. but we might come back if we have time for these four Social Science Applications papers that we talked about last week as practice. for this course and for the response essays. It's good to be able to look at these papers and think about them and ask these types of questions. So what is the research question? What data set is being used in? why? what is the paper trying to measure using the text data, and why? What's the purpose of it For the entering of research question? What no method is being used? Why was this method chosen? How was it validated? What else could they have done? And so you knowthese papers or from two thousand and nine, Two thousand and Ten inilp has changed a lot since then, right? and so often times the method they chose ten years ago it would be different now. Andasomething that that's what is that we want to think about it. time. Main results from a substantive social science standpoint: Why are they important You? The results seemed incomplete or non robust. So for you guys as mostly exams students on social science I would not expect you to know this just from run the abstract but when you're reading the paper like in the introduction they will normally have a paragraph explaining what are the other papers and so you do not need to be able to verify this objectively. but you need to be able to point in the paper where they say it and then what are the limitations in open questions? So again lie Now these one of these questions are and stars than others. but the in terms of what is the research question and what is determined in how they relate to each other. This and if I would be best to start practicing and thinking about so we talked about each of those questions a little bit in the race related research paper. the more fight and words paper what drives media silent then wind words slept so want to start practicing for the to say is tape I answered your question true love on your poll if it was saying a tenor's first montage she has on to train a time to to skin or if your zoom can you go you Asylum of right where go on the next slides. Now nobody knows how to get rid of this right? This everything only. So the week there lecture is about unsupervised learning and so this is really an example off what will be frequently doing in this course which is representation Learning Feminism reduction to extract relevant information from high dimensional data. Otherwise, in this text, next week based on the next two, We week four and week five are more on the supervise learning side where we for example have labels of documents and we want to teach a machine to reproduce those labels. If we do not have that then we we still want to extract relevant information and unsupervised learning is a way to do that. It's where an algorithm discovers themes and patterns in text and then we look at those topics or patients to interpret them so well we'll see is that you know they really is not like some kind of fine distinction here that is always clear cut. Often times will cause supervise learning to find the themes and patterns and unsupervised learning will often used to pursue a known goal or for making features for predictions and this is related to what we just saw at the end of the last weeks slides. which is these methods. If you were taking a regular plant course just learning the methods would be enough right? That would be tiny of it. But for us this is talent for law and social science and so we want to know what is the research question or the problem to be solved and then in service of that question, Why was this corpus or or data set chosen? How did they preprocess it? Do they produce enough descriptive statistics or visuals to really understand the data For unsupervised learning like a topic model, What are we trying to accomplish? What are we trying to measure by getting those topics and so other steps out of undertaking an undiscovered learning project known, pick a model based on your goal to probe sensitivity to hype parameters and try to provide indication that the model is delivering what we want and then number step forward. Empirical Analysis is doing a scope of science analysis. Usually this is testing some hypothesis like what drives media slant. For example or so the first set of methods that we look at is document distance. The classic way to do this is a set of algorithm is called text release algorithms like Smooth Watermen which really this is like Pleger is in detection software. So you know when you pick to documents and you find a sentence in one document is incidentally there and other documents. If the sentence is more than like five words there is something that's actually very unusual to five six words. Once you get to that length there's very few of them that are at tend to be shared across documents and so if you look for these shared sequences then you can really detect pretty well whether whether word text is being rescued or copy in tasted And so that's actually a shortcut that is much fat, so smooth water Inwhici looks for these shared sequences across documents That's computationally expensive. so for large corporate you can not live to it. And looking for shared hatched five grams like these quiet grams like so five ward sequences actually worked pretty well as a shortcut. If you have two documents that share a lot of five gardens, then they they're likely from the same source. So a more standard way of of doing unmet comparison, which is used for example in information information retrieval is basically just take factors of the word counts or the phrase counts in farm counts and then take the cozine similarity though those factors. And so that's what we did the last time is we take a document a list of words and convert it to a frequency distribution over words. And generally we will use the Tfida or inverse document frequent red word frequencies, and then you have each document as a non negative factor in an index space where index is the vocabulary size and so documents, Rays and similar documents have similar concerts. So you can imagine if you take a document to multiyou duplicate it so you have two documents that are like one after another, those rays are still pointing the same direction, right, but one is looted the other, but in order to make events of varying lengths more comparable. rather than use a Uclidian distance, we use this choice of the vector trees angels and so perfectly collinear documents have a concise similarity of one of documents are orthogonal means they have no words or phases in common, then they concise similarity of zero yes to and that depends so it will often provide better performance if the voters are normalized for example, so you just divide them by the norm poor even if you standardize them. let use as the standard secular in Psychic Learn which will center everything in the divide every dimension by the standard deviation. that will often improve performance and so if at those are the two kind of standard normalization approaches that you would use besides stuff and this is art of broader issue is that you often do not know which one to use when you're starting off and unless you have as if you have a metric for deciding which documents stimulate which documents you think should be together then you can evaluate it. But this is part of the broader issue that we'll talk about at the end. but often times for the document comparison document distance the method you use could to make a big difference ad so and if you do not have any document labels then you might have a problem and so in that case using the standard stupid of weighted in grams. the fact that so so many people use it is like a reason to start with it because people should have an intuitive understanding of it and it's become popular. And so the fact that we be Twenty Five indexing so like a lot of the standard indexes, comfort search and go in for databases queries use this way is a good reason for using it because at least just in practice people tend to agree with the results but will come back to that. So this is the formula for cuisine. Similarity is just the dot pro dust normalized by the norm of the victors. And you know in scholar you can get the concise similarity between all rows of a matrix or in just one line for example. So if you want to try to rank documents by their similarity, that's the way to do it. So if you have not done this before it might surprise you the first time that you run close and some party and you get Azilian metrics. So in times in this one so if you have a thousand simultaneously get a million scores back to be careful about that to if it will downright terms that appear in many documents. And so as it mentioned this stiff similarity is used in be the Twenty Five and elastic search. So that's like this standard robust matrix that's used across many domains. but this is a custodian one but there are other distance metrics that can be used. My sense is that there is not enough research kind of system out of comparing these, but this is what we have. So a research paper that we have linked on the syllabus that uses both be Twenty Five, the tariff similarity and the text Rescue Smooth Watermen. It is this burgus at a legislative influence detector. So what they do just to summarize ndthewill talk about it a bit. They take built texts legal texts across states and they compare both laws between states and across states as well as similarity to take to the model legislation provided by lobbying agencies. So here is an example they show in the paper of basically two bills between two states that are almost identical. So you would not be able to tell this but if the title of the bill is or of whosance sponsoring it or things like that. But it turns out that the same legal text is proposed in different states as it similar times. and so they use this text similarity matrix to look at the number of bills that are introduced from Black All which is this conservative lobbying group and they also look at the number of bills that are introduced by amid a liberal lobbying group. and I Think it's interesting that you would not be old to know this without this text similarity metric because they do not advertise it so likethey're kind of secretive about it. That's these bills that they propose on their own website. When a legislator is actually proposing them, they do not advertise that. and so it shows that they are trying to kind of secretly influence the legislative process. and this is a good example of a paper where it kind of makes this simple point of by taking a corpus of legislative documents and providing a measurement of similarity to these lobbying group documents. and you get this type of out of graphics and so just pink on your own. What is the research question here and why is it important and separately, what is the problem solved And I'll come back to that the second. What is being measured and how does the measurement help answer the research question. So in this class in the response essays in the exam, these are the types of questions that I want you guys to be able to answer and so noticed that it said what is the research question and what is their problem solved. Does anybody has not to guess why it put those separately, How would you distinguish those in this case or or generally setters. That's an interesting right. Let me see if let's say you could solve the research. Let's say you can answer the research question of nowhat is influencing legislatures. There could be a separate policy problem that just actually run this paper does not solve. Is that what you have had an ad in mind at Good At For Sure right? So actulyye to think it's even better. So it could be that this paper has a research question as they'd like to answer, but there are such statistical or computational other problems that have not been solved and so they actually could not answer the research question yet. And actually this paper is a good example of that. Where the problem is these lobbying organizations are secretive about further influence on legislatures and so this paper was to solve that problem of matching up the legislation with the model bills to measure. This shot is the problem that they try to solve, But the research question is what is influencing legislatures or what is the influence off of these model bills on legislative. This is important because this is not very democratic right? Like these are kind of the lobbing organizations and they are adding policies in this secretive manner. Other any questions and or reflections on that so ye comment or analyst at treatment in words and a person is brat more about brother as hasatist plan. Yes that's great right? So let's actually's another nice way to frame that distinction that there could be this technical problem right off measuring the influence of these model bills on legislation. but the research question is is this happening in the first place That's what's relevant for policy and for social science We there another point right? So he so in this case like the problem is like this technical question of how to to measure this influence. That's great. Okay so the next topic is do mention a reduction and this is something that elementary you as gin to prolyscene and other computer science classes. This picture is from the Aurelian Garden Book which I think that on the syllabus, the Aurelian Garden Book increase on machine Learning Physics learn is really the best way to read about the method side of what we're doing in this class. and the issue is that even if you have high dimensionality, data is not just randomly distributed across these all the dimensions in the data set and this latent dimension. So the Swiss role is an example of one is what you do call manifold and that manofhode can be learned. That's what dimension reduction techniques are designed to control this Swiss role for example or identify whatever other structure there is in other data sense. So what dimension reductions have we already tried? So let's think what we did in week one and week two in already today have we already done some dimension reduction to somebody? Want to think of a mission, an example or two? Ye totally right to the washing function. You go from having a vocabulary you know millions to acabular a tin thousand. Definitely dimension reduction. And actually even just the process of taking a plain text document and representing as a vector is dimension reduction. So the washing is an doing kind of two stages there out the totally year. Yet like that's compression right? So like unmoving software that's like basically taking out the extra noise data from the text even dictionary met this. Its is really extreme dimension reduction. like her you're saying if I want to say this is the bit that what I'm interested in. So what we've already done in the last two weeks and what we will do further down the road In this course most all of it can be understood as dimensioum reduction and the classic approach to dimensioum production that you would learn like to linear average class is a principal component analysis or a singular value deco position in and this is just as in the same way that we no and twenty five stiff Immigrants is like this kind of workforce for document similarity. Pace is its workforce for dimension reduction and the auralian Imaginary Book has a nice discussion of of this if it this new to you. But the most important practical piece is that you can just extract these these directions in the space, these dimensions that explain the most variants using just these three lines incycilar and you can imagine that this is a preventative algorithm that it captures the first dimenso that explains the most variant, takes that dimension out and you' get a second dimension, explain the rest the variants in the remaining data up until you have allowed the data explained and you can imagine. Rather than take the flood matrix of word frequencies in stiff matrix instead, you take the first ten principal components of that and then use that as your data representation. The problem with A is it? or I would say the advantage of Pace is that the distance metrics in the space are approximately preserved. So for many, perhaps most data sex, the tariff similarity, or the stiff distance between the full matrix of immigrants will be quite similar than let's say, the first hundred principal components of of that Natrids. So you could say like, rather than have an hundred thousand columns, you could have an hundred pounds and the distance between observations would be preserved, as that's very convenient if you're doing document distance. This type of dimensionof reduction could also be useful if you're doing supervise learning, so you could actually then use the principal components of the reduced matrix to our procedures, but this might not work well. This has not often done, But the other major problem with Pace is that the dimensions are not interoperable, so it's quite rare. If you lie, look at the first principal component or the tenth principle component. look at the words or phrases that are correlated with that. It usually do not be interpreted as a top. This is why it will. Beginning into other approaches to dimension reductions, there's topic models which do have this nice feature. It's something that is think's quite popular like in recommended systems but I have not easy to but none. Negative Matrix factorization in is it's a more interoperable. usually you get more interpretable outputs within me than in place so even try a that and Pace and Nmf it will dimension reduce your data to like a set of like tin components. Let's say where all of these are continuous and so rather than rather than have a dense one hundred thousand colunmatrix you'll have a dense hundred column trip. If you want your data to be sparse at the end and you want to just separate the data into groups then on the society that's called clustering is recommended for that. So may means clustering. Its an along with them that you take data that could be distributed in various ways with in a space and campaigns. Clustering will automatically find these cluster boundaries and assign your data into different locations based on the geometry. In terms of you knwthe terms of fixing the number of clusters this is a Hyperpremer You have to decide and if you end up using Cinemas clustering you should read this chapter in in the early Hungarian book because you can use something called the Satellite Score to find the optimal number of clusters. So in terms of other clustering algorithms to others that you him I think Kmeans is again we just like this standard approach that often works in different kinds of data but you could also try a medoid clustering which gets it medians rather than than mean asteroids do scan will work better if you have data and this so where you need to identify these continuous regions of destiny. Most of the time you do not know what your data looks like before this so you meet as why the sense came into clustering and to conclude this section on clustering just to give you an idea how this could be used illegal or such as science applications in Ganglemere and Ward law they apply a modified clustering so a clustering approach to different contract clauses and then you can imagine the clustersintroid as like a K of standard contract, the Boiler Plate contracts and then a contract that's kind of far away from this Detroit that's an outlier customised contract clause and so you can imagine that this is useful for a descriptive analysis. They can say you know what types of contracts have clauses that are far away from this indirect are more customised and they find it the larger deals so that the more valuable deals have more customised contracts which that make since deceptively probed and Phillips is a paper where they use in K filings so basically filing with the Securities and Exchange Commission and so this's like kind of finance that they apply a text clustering approach to the business description sections of those contracts and so this is basically a little texting bit describing what the business does and so then they can. if they apply taxes clustering to that, you get these kind of industries or these product groups and so and they use that to basically to analyze how companies are differentiating their products from each other and they do one antitrust analysis to that as well. Only so that's clustering. Were going to take a break now and will come back in eleven minutes at Three Fifteen to talk about top models. Ice are words heano on as a plan who they are inflicted for. Its up it right. We're going to resume with topic models. So as we mentioned a second ago, the one of the problems or the the disadvantages of our seeing pace for example as to mention and is that the resulting topics or factors or components are not interpretable and this is beyond the fact that if you have like an in grain representation of a document which is musiclikea vector or and hundred thousand numbers in it that's not going to be interpreted either and so ideally you'd want to look at us just on it like a single number. To summarize a document like this one is about law. This one is about policy Things like this and topic models are a way to do that. So the core methods and topic models they actually come from basin statistics Yes, Skleed is of pub his. it's I think the might have to go with that that stologay right on Toastis statue and on pins. So even though these topic models these methods were developed in computer science as way for information extraction, statistical representation, documents, marization and they ended up being quite popular and useful in social science. So basically starting off with documents you could then have to set of measurements that are on the topic, shares of different what do people talk about in toilet models will find a nice way to answer that question and this is. what's useful about them is that they're interoperable. So again, the work force relate to the depot model that is often used lady or lean dricklat allocation And this model basically assumes a structural model of language that each topic is a distribution of words the some set of topics out there in a corbus. Each document is a distribution of topics and so then you start off with a document being fifty topic one, fifty per topic, two and use sample words from those topics and that produces a document. And it turns out that mathematically geometrically is just another example of matrix factorization in the same way that pace and if we were factoring matricies and so what we assume is that there's a topics and that's something that you can choose when you're setting up your cot niwl. it's really the only thing you have to choose when you're setting up lady besides pre processing. So that's another reason that this is polar because is pretty simple to get a nice result. and like Pr, Nmf lady will take in your corpus which is represented by one bigmaker its x and factorize it down into two smaller motors is a document topic matrix and a topic term matrix and this is borrowed this scheme from some of branding's credit slides where basically all you a depict is capital kit here the number of topics and and then your label. read through your corpus and infer from the word counts in each document what are the or is the sound working now can you indicate in the chat? Thanks Got all of these details are as idea. you can actually just train kids withlke five lines using payments gensim and in the end you get this nice kind of statistical highlighted tool where you take in a document for example this newspaper article from Henoolix sites and it will identify the words in that document that correspond to these interpretable topics of these constellations of words that try to go occur and usually this works almost all the time. Where you get these topics like this one's like genes, genome sequencing, genetic that'sgens topic right and an organisms survived life that's like a biology topic and the computer numbers competition predictions that's like statistics topic, computer analysis and say they're starting off with try they behave that they spin to right well do this the last time. If it does not work now then I'll just others finish in the room without the zoom and then ultra record a separate version the zoom stream. So as I mentioned like you know so one you have lady trained you can then take out any document and get a a set of topic shares for that document. or you can just take the highest probability, highest probability topic and that is the time for the document and then once you have these paper proportions those then can be variables in a social science analysis. So an example that we linked in the sillabis is any calacatalinack's two thousand and sixteen paper. She ran a topic model on Japanese political speeches and identified topics that were related to national or federal or policy issues freeze topics that are related to local or which will talk about. So everybody on the zoom i'm I'm just going to it's going in and out so I'm just going to finish the lecture in person. In in the report are another version to add to for few guys later. so I realize that that's a bit inconvenient but we'll have a backup next time. |
| 4th row | Well, I'd like to, in these two talks, I'd like to talk about some foundational issues, in particular with the most important ones, I think, namely, what are the fundamental computational operations that enter into constructing syntactic objects and why these and not other ones. And It turns out there's quite a lot to say about that, since the last time I talked here, there were many problems, some solutions, I'll get to this in the course of the discussion as far as I can, but I think it would be useful to begin with a couple of comments on something more general, namely: what are we trying to achieve altogether in studying language, many different ways of looking at it. These Questions I think are in many ways more important than the particular technical results. They raise many questions about what is an authentic, genuine explanation, a genuine solution, and what is a maybe very valuable reorganization of data, posing of problems, often posing as solution, but not really achieving it. These Things are worth thinking through, I think. The Basic issues were formulated, I think, for the first time, quite perceptively, at the outset of the scientific revolution in the 17th century. Galileo and his contemporaries, who were raising all sorts of questions about received wisdom, turned their attention to language as well. And They expressed their awe and amazement at the miraculous fact that with a couple of dozen sounds, it was somehow possible to express an infinite number of thoughts and to find ways to convey to others who have no access to our minds everything that's going on in our minds. So In their own words, which I rather like, I'll quote, they were awed by the method by which we are able to express our thoughts, the marvelous invention by which using 25 or 30 sounds, we can create the infinite variety of expressions, which having nothing themselves in common with what is passing in our minds, nonetheless permit us to express all our secrets and allow us to understand what is not present to consciousness, in fact, everything we can conceive and the most diverse movements of our soul. Galileo himself regarded the alphabet as the most dependence of human inventions because it had these amazing properties and also because, as he put it, it allowed us to express all the wisdom of the ages and to it contained within it the answers to any questions that we might pose,. kind of like a universal Turing machine in our terms. The Port Royal Grammar on Logic, actually which I was just quoting a paraphrase of Galileo, had many insights into logic and linguistics,. it's in many ways the basis of modern logic. There was a rich tradition that developed exploring what was called rational and universal grammar,. rational because it was supposed to provide explanations, universal because it was concerned with what was taken to be common to the common human possession of language, was seeking explanations including descriptions of even the vernacular which was quite surprising at the time, innovative, but mainly explanations and universal Trying to find what's common to all languages. This Tradition went on for a couple of centuries, many contributions. The Last representative of it about a century ago was Otto Jesperson, as he put it, his concern was how the elements of language come into existence in the mind of a speaker on the basis of finite experience, yielding a notion of structure that is definite enough to guide him in framing sentences of his own, a crucially free expressions that are typically new to speaker and hearer. And also beyond that to find the great principles that underlie the grammars of all languages. I Think it's fair to, you have to interpret that tradition is metaphoric, often vague, but I Think it's fair to extricate from it. The recognition that language is the capacity for language, as well as individual languages are possessions of individual persons, they're part of a person, they're shared, was recognized throughout the species without significant variation, and recognized to be unique to humans in fundamental respects. That General Program falls within the Natural Sciences, within what these days is called the bilingualistic program. Of Course, it ran into many difficulties, conceptual difficulties, empirical difficulties,. the evidence was pretty thin and nobody really understood how to capture the notion,. Jespersson's notion of structure in the mind and what is that that enables us to develop, construct in our minds infinitely many expressions, and even to find a way to convey to others what's going on in our mind, that's called the Galilean Challenge, which is still extant. Well, All of this was swept aside in the 20th century by structuralist, behaviorist currents which very typically adopted a very different approach to language, taking the object of study not to be something internal to the person, but some outside thing. So Maybe a corpus, an infinite set of utterances, some other external formulation. And You see this very clearly if you simply look at the definitions of language that were given through the early 20th century by the leading figures,. So for example, a language is a kind of social contract, it's a collection of word images in the community of speakers. For Leonard Bloomfield, its language is the utterances that can be made in a particular speech community. For Harris, it's the distribution of morphemes in a set of sentences. For the philosophy of language, say, Van Quine or languages, as he put it, I'm quoting, a fabric of sentences associated with one another and with stimuli by the mechanism of conditioned response. Elsewhere, an infinite set of sentences. David Lewis, languages also took just languages, a language is some set of sentences which is infinite. Both Quine and Lewis crucially argued that it makes sense to talk about an infinite set of sentences, but not of a particular way of generating them, which is a very strange notion if you think about it because these are the leading logicians and philosophers. You Can't talk about an infinite set in any coherent fashion unless you have some characterization of what's in it and what's not in it. Otherwise, you're not saying anything. But The Behaviorist,. the pressure of behaviorist beliefs was so powerful that the idea that there could be a privileged way of generating that infinite set was, as Quine put it, folly, Lewis put it, something unintelligible. But Whatever any of these entities are, they're outside the individual. The Tradition was completely forgotten, people like Jesperson, the last representative, were literally unknown,. There's good review of this by a historian of linguistic. Julia Falk who runs through the way. Jesperson was disappeared in the first half of the 20th century, and the whole tradition way back also. In Fact. To this day, the even linguistics historical scholarship is pretty thin, it doesn't barely recognize any of the things I've mentioned. So Returning to the forgotten tradition, by the mid-20th century, there were clear ways of capturing the concept, the notion of structure in the mind, Jesperson's concept, touring other great mathematicians that established the tools for addressing the Galilean challenge, something you're all I'm sure familiar with. So Jesperson's notion of structure becomes what's now called the I language, the internal generative system, finite system that determines an infinite array of hierarchically structured expressions that express thoughts insofar as they can be expressed linguistically and it can be externalized in sensory motor systems, typically though we know not necessarily sound, we can call this the basic property of language. Well To meet the Galilean challenge, there are several tasks that have to be undertaken,. The main one, of course, is to try to determine the internal languages, the I languages of speakers of typologically varied languages, a huge task,. Then the question comes of how a speaker selects a particular expression from the internal I language, then how the expression once selected is externalized and the inverse how the externalization is internalized by the here, the last two tasks are both input output systems, we kind of grasp how to study those, and quite a lot has been learned about it over the years. The First of them, how the speaker selects a syntactic object out of the infinite array,. That's a total mystery,. there's nothing to say about it, that's true of voluntary behavior generally,. So actually here at MIT some of the two of the leading specialists on the neuroscience of voluntary action, Emilio Bizzi, Robert Adjemian About a year ago wrote a state of the art article in which they discussed how, what they know about voluntary motion, simple, not language, simple things like lifting your finger, you know, And they said well, They put it as they said, fancifully, that we're beginning to learn about the puppet and the strings, but we can't say anything at all about the puppeteer, so how you select what you're going to do remains the kind of question that you can't even pose intelligibly in the sciences at this stage here, as well. Well, the eye language, keeping to the tradition, is a property of the individual and also the species specific faculty of language, also an internal property, something which allows the eye language to be acquired and it has to meet a couple of empirical conditions, two conditions which are kind of conflicting, the conditions of learnability and the conditions of evolvability,. So whatever the faculty of language is, it's got to be rich enough so that possessing it, a child can acquire the eye language from the scattered and limited data available. And it is scattered and limited. and it has to achieve the internal system, which has all of these rich and complex consequences,. so it has to be that rich. But it also has to be simple enough so that it could have evolved. And now we can be a little more specific about that, because some of the conditions of evolution of language are coming to light and talk about it later. if there's time and the evolution has to meet those empirical conditions. Well, those are the conditions for a genuine explanation,. If some proposed descriptive device satisfies these conditions, then it's the basis for an explanation for addressing the Yellow Land Challenge as it was formulated and developed. In the tradition of rational and universal grammar, the general explanation is always at the level of UG, the theory of the faculty of language, and it has to offer some prospects of satisfying the conditions of learnability and evolvability,. that's a pretty austere requirement, very austere requirement, but it's the right requirement. Anything short of that is short of actually explaining things, it's maybe very valuable, maybe organizing problems in an interesting way and move on from there, but still falls short of general explanation. We Can now, I think, grasp somewhat more clearly what actually is a genuine explanation, something that was really not possible in earlier stages of linguistic inquiry, but again, any device that's introduced to account for something unless it can meet these joint, these dual conditions is short of explanation, maybe very valuable. So Many examples, take a concrete example to illustrate about something I'll come back to later if there's time,. an interesting paper by Djokovic, who you all know, on the coordinate structure and adjunct island constraints. What he points out, is that each of these constraints poses many problems, many mysteries,. but his paper is an effort to try to reduce the mysteries by reducing both constraints to the same constraint, using the device of neo-Davidsonian event semantics, which interprets a junction as a kind of coordination. So You can reduce both of the problems to the same problem of coordination, and then we still have the mysteries,. but now a simpler problem, one set of mysteries instead of two independent ones, tries to show that the problems then reduce this way. Well That's a step forward, it leaves the mysteries in a better position for productive inquiry, but it's not an explanation, he's quite clear about that. And I Think if you look over the field that virtually every achievement, everyone, is a partial step forward in this respect. There's very few exceptions, just barely coming to Light, which I think can count as genuine explanations,. They're important in themselves, and they're also kind of a sort of a guideline into how we should think about proceeding, and they may also tell us something about just how far it's possible to go. It's not so obvious, you can go much beyond what kinds of explanations that are now beginning to come to light,. I'll talk about that. Well, actually the earliest work in generative grammar, tried to meet even more austere conditions. It was heavily influenced by a work of people like Nelson Goodman and W.E.V. Quine, who were working on what they called constructive nominalism. No sets, very austere, just a mere illogical concept of a very limited kind. That was too austere, at least for the present, couldn't get very far that way,. there were several papers about it. So That was kind of dropped, at least for the present, maybe even come back to it someday, and the tension turned to something else, namely the vast range of empirical data from all kinds of languages that was beginning to appear as soon as the first efforts were made to write actual generative grammars. It Turned out that everything was puzzling and complex,. nothing was understood, it was just massive puzzles. Big Change From a few years earlier, during the period of structural linguistics, it was basically assumed that everything was known, everything was solved, we had the methods of analysis, you could formalize them, all that was needed was to just apply them to one or another language. That turned out to be radically false. Well, the first proposals, as you all know, were dual,. There were operations to deal with the problem of compositionality, very structured grammar, and totally different operations to deal with the phenomenon of dislocation, ubiquitous phenomenon, transformational grammar. Both Systems were far too complex to meet the long-term goals of genuine explanation, that was well understood. The General assumption at the time remaining for a long time, often broken up until today, is that the principles of compositionality are natural, you can expect those, something like very structured grammar, But the dislocation is a weird property that languages have, a kind of imperfection that we have to somehow, languages for some reason have this,. formal languages would never be constructed with that property. And That is still a widely held view, I Think it's exactly the opposite of the truth, the opposite I Think turns out to be true, that more recent work suggests that dislocation is kind of the null hypothesis, it's what's expected on the simplest grounds, and it's the most primitive of operations, I'll come back to that. But Let me just take a brief look at the steps that were taken to reach what I think is this conclusion. Well In the 60s, phrase structure grammars were basically eliminated. A Phrase structure grammar is far too rich to be contemplated as relevant to describing languages, so there's nothing in the theory of phrase structure grammar that prevents you, say, from having a rule, you know, VP arrow NCP, let's say, fine phrase structure rule, doesn't make any sense. It was just assumed, you just can't do that sort of thing. But The right theory has to rule that out as unacceptable. And That step was taken by the late 60s, basically led to X Bar Theory. X Bar theory had interesting consequences, which weren't really fully appreciated at the time, they're obvious in retrospect., For One thing, X Bar theory, notice, has no linear order. So Japanese and English, roughly mirror images, have about the same X Bar theory, linear orders on the side somewhere. That was a step towards something which I think is now much clearer, namely that the surface order of expressions is not strictly speaking part of language. It's something else. We'll come back to that. But If you just look at X Bar Theory, it's already a step in that direction. Another Thing about X Bar theory is it forces a theory of parameters. So Japanese and English, say, differ, and they're going to differ in some choice that is not determined by X Bar theory. So Some, the speaker and the hearer, who's using a linear system of externalization, you don't have to use that. But If you are using it, you're going to have to make a choice as to the order in which you're going to externalize the internal system. So X Bar theory itself, first, is a step towards separating a linear order and other surface organization from what we might think of as core I language, the I language that's dealing with the Galilean challenge, constructing the set of linguistically articulated thoughts, putting externalization in some medium to the side. And I Think that picture is becoming clearer. We'll come back to that. Well, There are also, along with the clear progress of X bar Theory, there were very serious problems which weren't recognized at the time. The Main problem is it excludes the possibility of exocentric constructions. Everything has to be endocentric in X bar theory. And That's just false. There are exocentric constructions all over the place, simple things like subject predicate, or for that matter, every case of dislocation, without exception. All of these give you exocentric constructions. There's no way to describe them in X bar theory. Now In order to describe them, many artifices were developed. So For example, if you have a subject predicate construction, maybe it was called a TP or an IP or something like that, or a VP. But That's just stipulation. You could just as well call it an NP. And This runs all the way through the descriptive apparatus. So There was a serious problem not really recognized until a couple of years ago. My Own feeling is it's pretty much overcome by labeling theory, which tells you in a principled way in terms of minimal search, a simple computational principle, when movement, internal merge, may take place, when it must take place, when it need not take place. There are many interesting results and plenty of interesting problems about this, a lot of very intriguing material, most of which I presume you're familiar with. Well, by the moving up to the 1990s, it did seem to a number of us that it's enough had been learned, so it might be possible for the first time to confront the problem of genuine explanation. That's what's called the minimalist program. Pursuing That program. If you want the, if you want a genuine explanation, you want to start with computational operations which meet the conditions of learnability and evolvability. Well, the easiest way to meet the condition of learnability is to say that learnability is zero. It's just innate, nothing to say about it. And The easiest way to meet the condition of evolvability would be to say, let's find a computational principle that had to evolve. There was no way for it not to have evolved. Well, if you look at those two conditions, they're satisfied by the most elementary computational operation, what's been called merge in recent years, which incidentally has many problems that I'll come back to. But Basically just the operation of a binary set formation. It has to be there because the basic property exists. And That means at least, at the very least, the simplest operation must exist, maybe more complex ones, but at least the simplest one. So We know that it has to exist, had to evolve, so it meets the condition of evolvability. That leaves the question of just how it happened and what the neurological implication is. But Whatever the answers to those, this is an operation that had to evolve. And Having evolved, it's innate, so it meets the condition of learnability. So If you can reduce something to that, you do have a genuine explanation. That's as far as it's possible to go. If It doesn't,. if you can't go that far, it's a description. It's not a genuine explanation. Again, this is a pretty austere requirement, but I Think it's the one we ought to have in mind when we're thinking about the goals of our efforts in inquiring into language. Well, So let's, I won't give the details because I think you're familiar with them, but the simplest computational operation, then merge binary set formation, meeting the no tampering condition, least possible computation. You Don't modify the elements, don't add any more structure. Interesting Things to say about this, which I'll come back. There is a good deal of current literature which tries to show that you can reach this operation in steps. That's incoherent. You can't have partial binary set formation. You can't reach it in steps. You Either have it or you don't have it. There's nothing simpler. Again, lots of literature about this, but it's just beside the point. There's actually a recent, interesting recent paper by Rene Heubrichs analyzing some of the recent proposals and showing why they don't make any sense. But If you think about it, they can't make sense. The Simplest case of merge is going to have at least, maybe at most we would like to show, but at least two cases. One of them, external merge when you're taking separate things and forming the set. One Internal merge when you're taking one thing and something inside it, forming the set of those. Those are at least the two simplest possibilities. Notice There are only one operation. There's no two operations, just one operation with two cases. Much Confusion about this in the literature, but that should be obvious if you think it through. Well, notice that this whole program is a program. It's not a theory. The Program is to see how far can we go if we take the simplest possible operation and try to give genuine explanations in terms of it. Maybe That's impossible. Maybe You have to find more complex operations. But In that case, it's going to be necessary to demonstrate how they can be acquired, how they can be learned, and how they could have evolved. And That's not so trivial. You Can't just say, well, natural selection does anything I like. That's not an explanation. You Have to give a real explanation. Very difficult in biology. In The biological literature, it's pointed out that it's a fiendishly difficult standard phrase to give an account of the evolution of almost any trait, even the simplest ones, like having blue eyes, for example., And It's not the kind of thing you can hand wave about. So Either you can try to meet that condition or recognize that you don't have genuine explanations. Well, there have been, I think, substantial achievements in the last recent years in trying to gain general, genuine explanations. They do have problems. I Want to return to the problems later, but I'll put them on the shelf for a moment. The One achievement, which is not trivial, is to unify the two traditional kinds of operations, compositionality and dislocation. They are unified once you keep to the simplest computational operation. So Far from being an imperfection, as was always assumed by me in particular, it would take a stipulation to bar dislocation. If You have no stipulations at all,. you get dislocation. Furthermore, As I mentioned before, that's arguably the simplest case of merge. Actually, you can't have only one and not the other, because once you have merge, you have both. But If you're looking for one that's more primitive, it's probably internal merge. The Reasons for that are quite straightforward. External Merge requires enormous search. To Put two things together that are separate,. First of all, you have to search the entire lexicon. Then You have to search everything that's already been constructed and maybe is sitting there somewhere waiting to be merged. With Internal merge, you have almost no search at all. So One reason for regarding internal merge dislocation is more primitive. It requires a tiny fraction of the search. But There's a good deal more than that. There's some interesting suggestions in the literature. They're not definitive, but they're suggestive. So One was some work that was done by Marv Minsky a couple decades ago. He and one of his students just explored what would happen if he took the simplest touring machines, smallest number of states, smallest number of symbols, and just let them run free and see what happens. What Turned out was kind of interesting. Most of them crashed, either got into infinite loops or just stopped. But The ones that didn't crash, all of them gave the successor function. Now, what's the successor function? Well, One thing the successor function is, is internal merge. So If you take merge and you have a one member lexicon, just run three, you get the successor function. Minsky's argument at the time was that probably evolution,. in the course of evolution, nature found the simplest thing. That's what you'd expect. So It found the successor function. And That happens to be internal merge, not external merge. If You look at other organisms, a way down to the level of insects,. they have, they count. So An ant, say, can count the number of steps it's taken. It's got a counter, maybe a set of counters inside. And If you look at just the mathematics of successive counters, they kind of tend towards the successor function. It doesn't take a big step to move them up to the successor function. So From various points of view, it seems plausible to think that of the core operations, the most primitive one is actually dislocation, contrary to what was always thought. And As you get richer constructions, you have external merge and it gives you richer kinds of languages. We Plainly have it in natural language. It's not just internal merge. An Interesting question is why. It probably has to do with argument structure, which is uniquely related to external merge. We'll come back to that. Well, What's with the unification of internal and external merge, compositionality and dislocation,? what was suggested by X-bar theory, as I mentioned before, becomes much more clear and explicit. So It seems that the generation of the CI interface, sometimes called LF, what gets thematically interpreted, linguistically articulated thoughts, that's, we can call, core I language. And That just keeps the structure. No Linear order, no other kinds of arrangements. So Why is there linear order in spoken language? Incidentally, not strictly in sign language. So in sign language, which we know to be essentially equivalent to spoken language, there's different dimensionality. So You can use visual space. You can use simultaneous operations, the facial gestures and motions. So It's not strictly linear. It makes use of the contingencies allowed by the space that's of externalization. But Speech happens to be linear. You have to string words one after another. So If you pick that particular modality of externalization, yes, you're going to have linear order. But Does linear order have anything to do with language? Well, you know, depends what you think you want to call language. But What it really has to do with is an amalgam of two totally different independent systems. One of them, internal language. The Other, a particularly sensorimotor system, which has absolutely nothing to do with language. The Sensorimotor systems were around the hundreds of thousands, maybe millions of years before language ever appeared. They Don't seem to have been affected by language. At Most, there's very minor suggestions about slight adaptations that might have taken place for, say, changes of the alveolar ridge and clique languages. Very Small things. But Basically, the sensorimotor systems seem independent of language. But If you do externalize the internal system through this filter, you're going to get linear order. But Strictly speaking, that's a property of an amalgam of two independent systems. And In fact, that's true of externalization altogether. And Notice that externalization opposes a hard problem. You have two completely independent systems. They have nothing to do with one another. You have to match them somehow. You can expect that process to be pretty complex and also to be variable. You can do it in many different ways. Also, to be easily mutable, can change from one generation to another under slight effects. Putting All these expectations together, what is a natural expectation? And I Think it increasingly is coming to be imaginable, Maybe true, is that the variety and complexity and mutability of language is basically a property of externalization, not a property of language itself. And It could turn out to be true. It's a goal at the moment that the core I language is really unique, may not vary from language to language. Actually, that much is pretty much tacitly assumed in essentially all the work on formal semantics and pragmatics. It's not assumed to be parameterized from one language to another, or to be learned somehow. It's just there, which means if we ever understand it properly, it will be reducible to elementary computations, which just don't vary. That's the way the internal system works. That should be the goal of inquiry in those directions. I should say, just as a terminological point, what's called formal semantics is actually a form of syntax. It's symbolic manipulation. Technically, something becomes semantics when you relate it to the external world. And That's a tricky business. Even Things like, say, event calculus,. if you think about it, events are really mental constructions. You can't find them in the outside world. You construct them there. And The task of relating what's internal to the external world, dealing with questions of reference, is no trivial matter. A Lot to say about this, but I'll put it aside. But It seems to me we can see a goal for all of this work to try to reduce it to computational operations that do meet the conditions of genuine explanation. Again, a very austere criterion, but I think one that's worth keeping in mind. Well, these are all possibilities that I think are looking increasingly plausibly and the field may go in that direction. It'd be very striking discovery if it really does. Well, let's go on with genuine explanations. One of them is dislocation, putting it together with compositionality. And Notice that that includes automatically the basis for what's called reconstruction. You Keep to the no tampering condition. You Automatically get what's called the copy theory of movement. That's the basis for the complex properties of reconstruction. There's a lot to look into, but that's essentially the basis for it. You Don't need rules of reconstruction. They're just there. That's automatic. Well, of genuine explanations, the most interesting case, I think, is the old principle of structured dependence. This was discovered back in the 1950s. This is a really strange principle of language, which had never been noticed, namely that the rules and operations of language, the ones that yield interpretation of sentences, don't pay any attention to linear order. They Just deal with structures, which is extremely puzzling when you think about it because linear order is what you hear. It's 100% of what you hear. You Never hear structure. Furthermore, at least superficially, it seems that computations on linear order are simpler than computations on structure. From Another point of view, that turns out to be false, but at least superficially that looks right. So What it seems, and what always seemed extremely puzzling, is that the syntactic rules and the rules that yield semantic interpretations don't pay any attention to 100% of what you hear and to the simplest operations, which is a pretty puzzling fact. We Now have a simple explanation for it. It follows from the simplest computational operation. If The entire internal language is based on the computation of the simplest merge operation in its simplest form,. Then you automatically get structure dependence for operations of movement of construal of interpretation of everything else. I Won't run through examples. I Assume you're familiar with them, but that just seems to be a fact about all constructions and all languages. That, if it's correct, is a genuine explanation of a fundamental property of language, maybe the deepest property of language, that the core language just doesn't care about order and arrangement. It only cares about structure. And A child learning language just ignores everything they hear. By Now, there's interesting independent evidence supporting this conclusion. So For studies of language acquisition, which have proceeded in very sophisticated ways by now, have now gotten down to the point where 30-month-old infants have been shown already to observe the principle of structure dependence. That's almost no data, remember, and it's a very abstract principle. There's other work, earlier work by Steve Crane, Nakamura, who's got a lot of evidence. The Three-year-olds have mastered it. Recent Studies have it down to 30 months. If We have better studies, which, as they keep improving, it'll probably be earlier. What that means is you're just born with it. So It meets the condition of learnability, namely zero, and it has the condition of evolvability. You Have to have this particular operation at least, maybe more, but at least this one, because you do have the basic principle. Well, there's also, as many of you know, neuro-linguistic evidence. The studies of, inspired by Andrea Moro of a group in Milan, Musso and others, have shown, many of you know this, that if you present subjects with invented systems of two types, one which correspond to the rules of an actual language that the subjects don't know, the other, which uses things like linear order, you get different kinds of brain activity. In The case of, say, having a negation be the third word in the sentence, a very trivial operation, you get diffuse brain activity. If You follow what look like more complex rules of actual languages, you get activity in the expected language-specific areas, the brain, Broca's area, and so on. That's been, by now, replicated many times. It looks like a pretty solid result. There's also psycholinguistic evidence of other kinds. The Moro-Musso Experiments were actually suggested by work of Neil Smith and E. Anthony Timpley on a subject who they've been working with for many years, a young man they call Chris, who has extremely limited cognitive capacities, almost none, but tremendous linguistic capacities. He picks up languages like Ken Hale, like a sponge, in other words, just picks them up immediately. Neil Smith Tried these same experiments before the neuro-linguistic ones were done. He Just tried it with Chris and turned out when he gave Chris a nonsense language modeled on an actual language, he learned it easily, like every other language. When They gave him the very simple language, things like negation being the third word, he couldn't handle it at all. It was just a puzzle. He Can't deal with puzzles. That's what inspired the neuro-linguistic studies. I Think those are the most interesting discoveries so far in the brain sciences related to language. It's a direction in which other experimental work had gone. Looking Back at this, it seems to be one of these very rare cases where you have converging evidence from every direction, leading to the same conclusion that poor I language just is independent of linear order and other arrangements. You Have linguistic evidence, psycholinguistic evidence, neuro-linguistic evidence, evolutionary considerations, anything you can think about. Now There's a very curious fact. There's a huge literature in computational cognitive science trying to show that somehow this principle can be learned, which is a very weird fact if you look at it. It's like trying to find a complicated way to disprove the null hypothesis. Things Like that just don't happen in the sciences. I Mean, here you have the absolute optimal explanation and a huge literature trying to show, look, there's a very complicated way in which maybe we can reach the same conclusion. It's an enterprise that's kind of senseless at the base of it. Of Course, when you look at the actual cases, it never works. It's not going to work. If It did work, it would be meaningless because it's always asking the wrong question. I Mean, suppose you could show that by a detailed statistical analysis with recurrent neural networks and so on of many layers of, say, the Wall Street Journal, you could find evidence that a child might have used at 30 months old to discover that you have structure dependence. You're not going to find that, of course, even though there's literature claiming it. But If you did find it, it would be completely meaningless. Of Course, the only question is, why is this the case? Why Is it that in every language and every construction, this is the way it works? If You could find a way of showing, well, here's how it might work in this language, tells you nothing. It's answering the wrong question. And Furthermore, as I say, it's trying to find a complicated way to disprove the null hypothesis. The Whole Enterprise is completely senseless. It's actually probably the major effort in computational Cognitive science to try to find a basis for some linguistic principle, huge literature on it, new papers still coming out. A Very strange thing, papers trying to show that, as they put it often, you can get structure dependence without what's sometimes called an inductive bias for structure dependence. But There's no inductive bias. It's just the null hypothesis. Make No assumptions. This is what you get. There's no bias. It's just given. So I Think an interesting question about, many interesting questions about how linguistics is done. But One of them is why things like this go on. I Think it's worth thinking about. Well, there are other successes, but what I'd like to do is turn to problems. There are a lot of problems about merge, and there are some paths to solution. So One problem is what I already mentioned, exocetric constructions. So It takes a NPVP. Let's assume, since Dominique is here, let's in his honor, assume the predicate internal subject hypothesis. So you put together a subject and an NP and a VP. The NP's are often called DP's. I'll come back to that. I Think it's probably a mistake. Let's just call them noun phrases for the moment. You have a noun phrase and a verb phrase. You Put them together. That gives you the basic Theta structure. Well, the noun phrase and the verb phrase have to be independently constructed, which means you have to have some kind of workspace, something that Jonathan pointed out years ago. You Have to have some kind of workspace in which you're constructing these separate things. And If you think it through, the workspace can proliferate, not indefinitely, but can get larger, where you're just doing parallel things and putting them together. So It means that the operation merge really ought to be revised to become an operation on workspaces, not on two elements, X and Y. It's an operation which changes a workspace to another workspace. And Then the question comes, how it does it. Well, I should say I'm very pleased to be back at a nice low-tech institution like MIT with blackboards and no PowerPoint and no projections and any of that stuff, which they have in Arizona. So What we want is some kind of operation that says it's called a capital merge. We'll look at its properties, which takes two things, call them P and Q. The Guys were going to merge on a workspace and turns it into some other workspace. So What's the other workspace? Well, it's going to include the set PQ, the two guys were putting together. In Fact, let me use a different notation for reasons all mentioned. A Workspace is a set, but we want to distinguish it from the syntactic objects, which are sets. So A workspace doesn't merge with something. So Just for convenience, just use a different notation. So The new thing will include the set PQ and a lot of other junk. And The next question is what's the other junk in the workspace? That Turns out to be not a trivial question. A Lot turns on what the answer is. So Let's take the simplest case. The Entire workspace consists of two elements, a column A and B. That's the workspace. And Suppose we decide to merge them. So We get the new workspace, which includes the set AB. And Does it include anything else? So For example, does it include A and B? Well, If we think about the way recursion generally works, it should include A and B. So If you're doing, say, proof theory, you're generating a proof. You Construct a line from axioms and former things. And You can go back to that line, if you like. You Can always go back to anything you've produced already for the next step. But There's a good reason to believe that for organisms, and particularly humans, it doesn't work that way. And You can see that if you think what would happen if you did allow this to happen, suppose you allow this, then you could go on to construct some much bigger object here, including AB as a term. But It could be of arbitrary complexity, any kind of complexity you like. And Then you could take A and merge it with it and get xA. But Then A would be up here. And A would be down there. And There are two copies. And They would violate every imaginable constraint on movement. So If you allow this, you're going to get total chaos. Every constraint on dislocation will be violated. No matter how radical you make the violation. Well, that tells us something. It tells us something kind of surprising, and I Think significant, that the kind of recursion that takes place in human language, and probably organic systems generally, cuts back the number, the set of items accessible to computation as narrowly as possible. Let's give it a name and call it resource restriction, RR for simplicity. It Looks as though a very general, this is merely the first example. If You think it through, it works for millions of things. The Same model of refutation eliminates a whole set of possible extensions of merge that have been proposed over the years. I'll come back to examples. But You notice what the problem is. The Problem is that if you allow the normal kind of recursion, no constraints, no limits, then you're going to find that by legitimate means, you can get illegitimate objects. Now That has to be barred. You can generate all kind of divin expressions. That's not a problem. But You don't want to have legitimate means for generating things that violate every possible condition, descriptive condition. Anything Like that is wrong. Well, in this case, and it turns out in a great many cases, you can bar this outcome simply by limiting the resources that are available. Now, what are the resources? The Resources are elements that are accessible to the operations. So The real condition says limit accessibility. Keep Accessibility As small as you can. We Already have examples like that that we're familiar with. One of them is the phrase impenetrability condition. If You think about what that condition says, basically it says when you're generating something, you get to a certain unit, a phase, talk about what it is. Anything Inside the phase is no longer going to be accessible to operations. That reduces the amount of computational search that's required, but it's a way of limiting accessibility. It says those things down there aren't accessible anymore. Another example, and this may be the only other example, is minimal search. This is what's often called a third factor property. Third factor, for those of you who are not familiar, comes from the just simple description of the elements that enter into computation, into learning. So What enters into acquiring a system is three things, external data, internal structure, and basically laws of nature, which are independent of the system in question. So If you're studying growth of arms, let's say, humans grow arms, not wings, partly because of nutrition to the embryo, and partly, in fact, largely because of internal structure, just genetic determination, and extensively, simply because of the way physical laws operate. There's only certain ways that organisms can develop, other ways just aren't possible. You Put these together, you account for any kind of growth and development. The Same is true of language. There's external data, whatever it is, it's going to determine whether you end up with the Golag or English. Internal structure, which at least includes merge, more, no doubt, but at least that. And In fact, anything that can be explained in terms of that, it does yield a genuine explanation. And Then laws of nature. What are the laws of nature? Well, language is a computational system, kind of unusual fact. That's rarer and organic nature, maybe unique even, aside from counters. But Anyway, that's what language is. So Among the laws of nature that you would expect would be things like elimination of computational complexity, doing things as simply as possible. Several Reasons for that. One of them actually goes back to Galileo again. One of Galileo's precepts was that nature is simple and it's the task of the scientist to prove it, whether it's falling objects or flight of birds or growth of flowers or whatever. That's a kind of a prescriptive hypothesis. You Can't prove it, but it's been extraordinarily successful. In Fact, the whole success of the sciences in the last 500 years is based on that. And That's a good enough reason to assume that it works for us, too. So Reasonable to accept that. There's a general point, which just has to do with the nature of explanation. It's just a fact about explanation that the simpler the assumptions, the deeper the explanation. That's just kind of logic. So There's a lot of, I should say in the case of language, there's another reason to believe it which is unique to language. And it has to do with the conditions on evolution of language. We Don't know,. very little is known about evolution altogether. As I Said, to try to really account for the development of any particular trait is very hard, even in simple cases. In Sort of the evolutionary psychology literature, everything looks easy, happened by natural selection. Why Not? But When you really try to explain something, it turns out to be hard. In The case of cognitive development, it's uniquely hard because you have no fossil records, no tape recordings of people doing whatever they were doing 100,000 years ago. Furthermore, when you deal with language in particular, it's super hard. With Other organic systems, say vision, you have comparative evidence. You Can study cats and monkeys which have essentially the same visual system. And With them, we rightly or wrongly allow ourselves to do invasive experiments. So You can stick a neuron into one cell in the striate cortex and see what's happening and so on. And You learn a lot from that. That's how we know about human vision. But In language, you can't do it because there's no other analogous system. It's a unique system, nothing analogous in the organic world, so there's nothing to test. So It's kind of uniquely hard. And Nevertheless, there's some evidence. The Evidence at Bob Berwick and I have a book reviewing it. By Now there's better evidence than what we had in the book. There's by now genomic evidence that Homo Sapiens began to separate roughly 200,000 years ago. That's when you get the separation of the Sun people in Africa from the rest. Interestingly They have unique forms of externalization. These turn out to be essentially all and only the languages that have complex click systems. There are what apparently look like a few suggestions, exceptions, but they seem to be borrowings or something accidental. There's a kind of very interesting paper by Rini Huyper on this recently. So One thing we know pretty convincingly is that roughly around 200,000 years ago, humans began to separate. They shared the faculty of language at the time. So There's no known difference between the faculty of language of the Sun people and everybody else. Nobody knows any differences, group differences in language capacity. There happens to be a different form of externalization, which suggests, and in fact Rini goes into this in detail in his article, that this particular forms of externalization develop later. As A matter of logic, the internal system had to be there before you can externalize it. That's not debatable, but he suggests there's a gap when the system was there roughly 200,000 years ago and began to be externalized in somewhat different ways later on. When Did Homo Sapiens appear? Well Here, we have reasonably good fossil record which shows that anatomically modern humans emerge appear roughly at that time, maybe 250,000 years ago, which is essentially nothing in evolutionary time. So It looks as though the language emerged pretty much along with Homo Sapiens, with the faculty of language intact. Another Kind of evidence comes from the archaeological record, which gives a lot of information about rich symbolic activity. It Turns out that almost the entire rich symbolic activity that anybody's dug up so far is after the appearance of Homo Sapiens. Well rich symbolic activity has been naturally taken to be an indication of the existence of language. Also more complex social structures, you know, burial practices, all sorts of stuff. So Putting it all together, it looks plausible that language emerged suddenly in evolutionary time. Along With Homo Sapiens, some whatever change gave rise to Homo Sapiens seems to have brought language along with it, and it apparently hasn't changed since. That's independent reasons to believe that whatever is in there is probably very simple, along with the Galilean precept and the general principle. If you want explanation, you want simplicity. So It makes good sense from many points of view to assume that the relevant laws of nature here are avoiding computational complexity, computational efficiency. That's what It means to call it a third factor property. Well One particular case of computational simplicity. Hi, how are you doing? What Are you doing here? You Graduated years ago. Go Back to Tufts. I'm wondering where you put among the three factors: the fact that the computation has to run on neurons. That's third fact. First Of all, it's not necessarily the case. That's a myth, you know. There's a myth that neural nets are what do the computation, but there's pretty good evidence that that's not true. The Neural nets just don't have the computational principle. I'm talking about real neurons. Real Neurons. I'm talking about real neurons. Real neurons may not be the elements that enter into computation. There's by now reasonably strong evidence against that. Randy Galastol's book with William King is a very good case where he gives strong evidence, Randy, that if you look at neural nets, you simply cannot find the basic elements of essentially Turing machines. You can't find the core kind of computational element that yields computational activity. It's just not there in neural nets. So What he's arguing is that people who've been looking for neural net accounts of computation are like the traditional blind guy who's looking under the lamppost for his lost keys because even though he lost them across the street because that's where the light is. So Yes, we know something about neural nets, but happens to what we're looking for somewhere else. There's a lot more evidence. The Speed and scale of computation is way beyond what neural nets are capable of. And by now there's including. Randy's particular proposal is that the computation is actually down at the molecular level, having to do with RNA and so on. There are other proposals by not inconsiderable people like Roger Penrose For example, that the computation is being done by structures that are internal to neurons which have vastly greater computational capacity. There's chemical processes that go on in the brain that aren't captured by neural nets. It's been known as far back as Helmholtz. That speed of transition of neurons is just kind of too slow to be doing very much. So We're going to have to look elsewhere to find the implementation of computational systems. There's something there, and it's going to be a third factor property, something about the brain, clearly. We Talk about this in our book. So Yes, there's surely we're going to try to relate whatever is going on ultimately down to the level of cells. That's science, try to reduce everything. So That's all third factor if you like. But Just talking about neural nets is kind of like talking about natural selection. I Thought you brought up neurons. Something In the brain, yeah. Surely Something in the brain is responsible for this. not the foot, let's say. You Can amputate your leg, you still have language, you have to take your head, you don't. So Yeah, we agree. there's something going on in there. It's a tricky question. It's a very hard question, even for simple traits, not just language. Very Simple traits. As I Said, you look in Technical Studies of Evolution, the phrase that's often used is fiendishly difficult to find the evolutionary basis for even the simplest traits. So To think we're going to suddenly find them for language is a little misleading. There are some interesting ideas. So Angela Ferrichi's book that came out recently, MIT Press book on the state-of-the-art and neural linguistics, gives interesting suggestions about what might be involved in the probably small change in brain rewiring that led to the appearance of merge or something like it. You Know, closing a certain circuit in the dorsal and ventral connections. It's an interesting proposal, but certainly not going to be trivial. You Know, it's a hard problem. Yes, but yes, that would be third factor. So Now where was I? He's always been very disruptive ever since he was a student. Well, maybe I'll just kind of end here and go on next time. But One principle that we're going to expect for many reasons is computational efficiency. Minimal search is the strongest form of computational efficiency. Search As little as possible. And There is a case of restricting accessibility that we're familiar with, which reduces to minimal search. That's the case of successive cyclic movement. So Suppose you've taken a Wh phrase and you've moved it up to here and then you've moved it up to here and you keep going. And Suppose both of these, and neither of these, let's say, is blocked by the phase impenetrability condition. Suppose That only blocks things here. Well, the next thing that's raised is this one, not that one. We Just take that for granted. Nobody talks about it. But If you ask why, it's again a minimal search question. Whatever's selecting the operation, argue about what it is, that goes back to that mystery I mentioned is going to take this guy because that's the one that's going to find by minimal search. So We have at least two cases already that we're familiar with of limiting accessibility. One PIC, which is pretty broad, the other minimal search, which we've just taken for granted. And Maybe that exhausts it. But Now I think there's a broader general principle which says just restrict resources. That will have a lot of effects. I'll come back to more examples of that next time. There's a temptation at this point to relate, restrict resources to something that Ray and I were just talking about. The Fact that the brain is just slow, doesn't work fast, works quite slowly. And There's many domains in which that is the case. So In many ways, the most striking one is vision. If You look at the sensory motor system, the visual system, The cells of the retina are actually responsive to single photons of light. They're maximally, they give you a maximal amount of information. The Brain doesn't want all that information. It's just the way overloaded if it ever got that kind of information inside. So Whatever the visual system is doing is, the first step it's doing is throwing out almost all the information that's coming from the retina. And Apparently every sensory motor system, every sensory system, is like that. The First thing it does is throw away just about everything and try to get down to something limited enough so this slow brain up here can deal with it somehow. That looks very much like a general property of which Resource limitation of the kind That says, don't do ordinary recursion, but restrict the resources of which this is some special case. All seems to converge kind of plausibly. We're very familiar with this in the study of language acquisition. So As you all know, an infant acquiring a phonetic system is basically throwing away information. It's throwing away tons of information in the first months of life, and maybe in about nine months or a year, saying these are the only things I'm going to pay attention to. The Same thing goes on through the rest of language acquisition. If You look at something like Charles Yang's general approach to language acquisition where you're just shifting, you start,. the child starts with all possible grammars by languages. And Then it changes the probability distribution of them as data comes along, reducing the probability of things that don't have evidence for them so that they become fundamentally invisible. It's also a matter of throwing away lots of information and converging on just a little bit. The Development of the brain is constantly losing neurons because you don't want all this junk around. You Want just what you need. And The resource limitation fits pretty naturally into that system. I Think I'll stop here and try to come back to more detail examples next time. Unless Somebody else wants to disrupt. |
| 5th row | Well, last time I Talked about a number of things and got up to the point of beginning to discuss the problems that exist with the concept merge that was developed back in the 90s and has been used in many ways since. There's a kind of a simplest version of merge, which was the original intention, which just had the two special cases, external and internal merge. As I mentioned last time, the more primitive of the two is actually internal merge. But Because of the fact that language has exocentric constructions that can't suffice, I mentioned some of the things that you can explain on the basis of merge, and also wanted to make the point that a genuine explanation in linguistics will, if we're viewing the study of language as part of the study of nature, basically the bio-linguistics program, which I think has roots back to the 17th century, as I mentioned last time,. although you have to skip the structuralist, haveless period, if we're engaged in that enterprise, then a genuine explanation will always have to meet these austere conditions of learnability and evolvability, which are very hard to meet anywhere in biology and in particular here. There's some reason to think that they might be, conditions might be attainable here because of the specific conditions of human evolution, which I mentioned briefly last time. If That picture's correct,. there's some antecedent reason to believe that there might be success in the enterprise, which is rare in the biological sciences. Well, the concept merge does happen to meet those conditions. It meets the condition of learnability because there's nothing to learn. It meets the condition of evolvability because since, in fact, the basic problem, the basic principle does exist, there had to be something to evolve the computational procedures that yielded it. And It would obviously be at least the simplest one. So We can be secure with explanations based on the concept merge. But Anything else is problematic. It's a very austere condition, but it's one that really has to be met. Well, I then started in on, talked about some of the examples where you can get an explanation, some interesting cases. But There are problems. The Problems are that the concept was very loosely defined and all sorts of other applications, the implementations have been given, which kind of more or less fall within the original loose definitions, but I think are probably illegitimate. I'll talk about that today. And I Think If we think through the matter carefully, we end up with just allowing what was originally intended and none of the extensions for good reasons, which leaves us with many problems, some of which have, I think, potential solutions, others look quite mysterious. Well, I mentioned last time, something which is a kind of a paradigm for many cases, the simplest case when you have only two elements. First Of all, since we do have exocentric constructions, it's going to be necessary for the operation merge to actually operate on a workspace, not on elements, because you're always changing the workspace every time you apply merge. So We have some sort of a definition that we call a capital merge, which says take two things that you want to merge and a workspace that exists and form a new workspace, which will include at least this, and then other things you don't want. And I Suggested a notation, which I sometimes will forget. Sam Fixed it last time. I will use square brackets for the workspace and curly brackets for the syntactic objects. There is a crucial difference between them. The Workspace is a set, but it's not an accessible object for operations. So We'll just distinguish them by that notation. And The notation actually means something. So For example, if suppose the workspace consists of just x, we want to distinguish that from X. So The singleton set is different from its member, because the workspace is not a syntactic object, and x is. On The other hand, for the syntactic objects, we want the opposite convention, namely that the singleton set is the individual element. There are good empirical reasons for this, which go back to Phrase Structure grammar. So In Phrase Structure grammar, you just, by convention, didn't allow rules like, say, Np arrow Np or E arrow V. That's assumed not to be a reasonable rule. That's essentially saying that a singleton set is identical with its member. Now, this was fudged often in the use of phrase Structure grammar. So There were allowed rules were allowed like this. When You move from phrase structure grammar, which is totally unacceptable for language for myriad reasons, as was recognized since the 50s,. When you move from that to x bar theory, then this becomes meaningless because there's no Vp. if it's only V. Actually, despite the fact that it's meaningless, it is used. So For example, if you try to implement Richie Kane's LCA, try to get it to work, you're forced to have rules like this, which is a serious problem in the LCA system, I think. You Have to argue that if you have a verb object structure, the object, even if it's a pronoun, still is complex. Otherwise, you don't get the right ordering. But That's technically illegitimate in an x bar theoretic structure. And Here, the analog of that illegitimacy is this convention. So We want to accept this convention, which has a number of consequences for syntactic objects and this convention for workspaces, which are different kinds of things, though they're all sets. Well, In the simplest case, we just have a workspace consisting of these two guys. And We merge them. Merge AB gives us a workspace which contains the set AB, which we've merged. And The question is, what else? Now, if it was normal recursion, like say, proof theoretic recursion, you would have here A and B. But You can't have it for language, for reasons which I mentioned last time. And This is a kind of a paradigm that applies to a great many cases of the extensions of merge. The Reason it doesn't work is that this can be built up to an object of arbitrary complexity. This, since it's accessible, can then merge to it. That gives you a relation between the thing that's merged up here and the thing down here, which violates every imaginable condition of movement. So That's illegitimate. And We do not want to have legitimate operations which yield illegitimate conclusions. That's elementary. So Therefore, we conclude that, surprisingly, if these things aren't here for language, recursion for language is different from general recursion. Namely, it has the property that I hold last time restricting resources. And What computation for language, and presumably for organisms generally, is doing, is trying to keep the resources as limited as possible. You Have to get something new, or you don't generate anything. But You want to generate as few things as possible. This really turns into a number of subcases. One Subcase is limiting accessibility. So Accessibility means something's accessible if the merge operation can see it and do something to it. We Want to limit accessibility. If We allowed general recursion, we'd have too much accessibility here. So We want to limit it to the minimal amount. It's tempting, as I mentioned last time, to try to relate this to a more general property of brain computation. Namely, the brain is pretty dumb and slow. So What it does is throw out tons of data that are coming in. In Fact, that's its main activity, is to get rid of lots of stuff that's coming in. So In the visual system, the sensory part of the visual system is essentially perfect. So You get a cell responding to a photon of light. Can't do better than that. But That's pouring into the brain tons of information that are going to totally overwhelm a computation. So The sensory system throws out almost everything, get down to just the limited part. Same True in language acquisition. The Phonetic system is throwing out just about all the noise that comes in, picking only very limited kinds of phonetic properties. And Even those are being thrown out very quickly in early language acquisition. That's the main part of language acquisition. Charles Yang's model for general language acquisition kind of exploits this generally. Notice, incidentally speaking, of Yang that if you have this property, you infer determinacy. It Turns out that if you think it through, when you limit the resources available, you're also going to force determinacy, meaning the operation will be uniquely determined by what it's looking at. That's not a trivial property. It wasn't true, for example, of standard versions of phrase structure grammar. So In a phrase structure grammar, if you reached a point where you had something of the form NP, the NP, and you have a rule expanding NP, which one you apply it to is indeterminate. It's a normal part of phrase structure grammar. But If you have in this much narrower system, resource restriction, you get determinacy. And That's kind of important, because Charles's work on the price of productivity are very important work, I think. That work depends, personally, on the assumption that the operations are determinate. So If you want to get those rich consequences, we want that property. It's one of the few examples I know of of the work in computational, statistical or computational linguistics that has real consequences, very rich consequences, very important work. So Resource restriction makes sense, as a property has a lot of interesting consequences. Lots follow from it. Now, resource restriction is going to have two components, sort of. One of them is restrict computation. The other is restrict resources. Restrict computation means limit yourself to the minimal kinds of computation that are possible. Well, merge is one case. It's the least possible computational operation. But It also wants to operate in the most limited way. So One of the consequences would be keep to minimal search. Don't use deep search if minimal search already works. Many Empirical consequences to that. Notice That limit accessibility already includes things like PIC, the phase impenetrability condition is one general condition that limits accessibility. Minimal search is another. It means, for example, in successive cyclic movement, when you move to the next stage, you don't look down. You only find the first thing that you find. That raises many interesting questions about possible ambiguity. Are There ambiguous cases? I'll come back to that. quite interesting question later on. But Those are the things that are in the background. Now, if you look at the original definition of merge back in 1995, it actually had this property inadvertently. It wasn't noticed, particularly. But The operation of merge, as it was defined, was defined basically as replace, which says, don't keep what you already had, but get rid of it. That was not particularly noticed. But It's a property of the original definition. And There are good reasons for it. We Can now see good reasons for it. If You don't accept it,. you do get legitimate operations, which yield illegitimate conclusions, the clearest sign that something's radically wrong. This is a paradigm case, but it extends to many others. Go into that. Well, So for example, take something that I presume no one's ever proposed. I'll draw trees. But Let me make, I should make it clear that trees are very misleading notations. One should be aware of them. For One thing, a tree notation suggests that this exists, that the root exists, but why does the root exist if you just have merge operations? In Fact, what the root is understood to be is something that comes from some other source, namely projectability. But That should be a completely separate property. Projection seems to have nothing to do with the compositional operations. The Tree notation kind of sticks them all together and misses many questions about what projectability is,. a particularly interesting case in exocentric constructions. So What's projected? Well, that has interesting consequences. Whole theory of labeling deals with that. I Assume you're familiar with it, or I'll put it aside. But We don't want that. The Other thing that tree notations allow you to do is draw lines in all kind of complicated ways. Draw A line from here to here, and that seems to mean something in a tree, but it means absolutely nothing in a merge-based system. And You really have to be very careful about that. So I'll use trees for exposition, but with a condition that you don't take them seriously. Well, let's take something that I presume nobody's ever suggested. Suppose That you have a structure like this, and you decide to merge these two. The Original definition doesn't say you can't. So That would mean you're forming the set xp, yp. But It has nothing to do with this original set, just something added on. And Notice that it has exactly the properties that are barred here. It adds accessibility. It's adding new accessible items, which will then be subject to exactly the problem I already mentioned, namely: this object can be made as complex as you like. You could then merge one of these guys to it, and you violate all conditions. So We're not allowed to do this. As Far as I know, nobody ever proposed this, but there's a good reason why you can't do it. However, There are things that people have proposed, and I think they're all ruled out for the same reason. I'll kind of leave it to you as an exercise to work out why, but if you think about things like parallel merge, which is usually written in trees like this. Again, the case where the tree notation seems to be saying something, but it's not because there's no way to construct this. But If you think about what parallel merge is doing, it's increasing accessibility runs into this very same problem. Parallel Merge is the basis for many, a lot of work in the literature, which yields multidimensionality. The Idea of multidimensionality goes back to the 70s, work by Jim McCauley and others. But If you try to reconstruct it in a merge-based system, you can draw the trees with funny lines like this. And The way of constructing it is through parallel merge, which has this lethal property that it yields illegitimate consequences from legitimate operations. So Everything That, if you look up the handbooks of contemporary syntax, there's a chapter on multidimensionality, which has many interesting consequences about ATB, across the board movement, parasitic gaps, and so on. But None of them work, because they're all based on an illegitimate operation which has this efficiency. Same True of sidewards merge as the same problem. The Same is true in spades this time of late merge, which is widely used. I've used it a number of times. Many others have. Late Merge, first of all, has this problem. It's creating a new, when you draw a tree, it looks as if you can do it. You Just add a line to the tree, and you've got late merge. But If you try to spell it out in terms of the merge operation, you're first creating a new object, which is bad enough, because that's already illegitimate. But Then you're adding a new operation, a substitution operation, which inserts what you've just created at just the right point inside the tree. That's a pretty tricky operation. Try to formulate it. It's a new, complex operation. So Late merge has a double problem. One, the problem of not restricting accessibility. Second, the problem of invoking a new operation, which is really unformulable. It's quite a complex operation, if you think about it. And It's way out of the framework of anything we're talking about. There's many very interesting results that follow from late Merge, very much like multidimensionality. But I Think the way to look at those results is as problems, problems that have been constructed in an interesting way. So We have organized data instead of chaotic data, which is a step forward. But It's only a step towards eventual explanation in terms of something that meets the austere conditions that we're interested in. And that sets interesting empirical problems to address. I Think there are some answers in some cases, which I'll. OK. OK. I've been to that question a little. Somebody, yeah, that's right. Figure out what's going on. Let There be light. OK. Is that the second day of creation, I think? Thanks. You're divine. OK. So Where are we? Sorry. We Need a different God. OK. OK. Are You guys doing that? OK. Well, There's a lot that follows from all of this. I'm kind of leaving it as an exercise to think it through. But If you think it through, what you'll find is that all of these applications, extensions of merge, including the kind that nobody's ever thought of, including others that have been used widely, all have the same problem. They all have exactly the problem that you see with the simplest case. And The problem, again, crucially is that they are constructing what are alleged to be legitimate operations. But When you apply them, you get illegitimate conclusions. And That is the sign that there's something seriously wrong. Obviously, that can't be. So All of those, the entire literature, big literature, that yields very interesting array of results is not legitimate explanation. It's a proposal of problems. It's posing problems that are interesting, important, big step forward. It's useful to have organization of data instead of just chaotic data. But That's not explanation. That's not the goal of linguistics, at least as a science. Well, All of this suggests a kind of a research program. First, determine which subclass of the loosely characterized operations of merge, which subclass of them are, in fact, legitimate. That's a research problem. If You run through it, I Think you think through it, case by case. Again, I Believe this is an exercise. I Think what you end up finding is the originally intended ones not defined properly, but the originally intended ones are probably the only ones. The Rest of the extensions are not legitimate. Interesting Consequences, but not legitimate. The Next problem is to formulate merge. So It gives you just the right ones and then try to explain why that definition of merge is the kind that should be reached on general considerations, general conditions that any linguistic operation ought to meet. And This should be deducible from them, along with third-factor properties, minimal computation, minimal resources. Well, I Won't run through the cases, but we get something that looks like this. And We have to put various conditions on X1 to Xn. So What are the conditions? Well, I Won't bother spelling it out formally. I'll just say it intuitively. Don't lose anything in WS. Spell it out. It means if Y is in WS and Y is distinct from P and Q, it's got to be in the X's. So You don't lose anything. We Don't want something to just disappear in the course of the operation. Second Condition, limit accessibility. In Fact, limit it to one. It has to be at least one because you're constructing a new object. Otherwise, you're not doing anything. But Don't do anything beyond that. So No new accessibility should be permitted by merge. And A third condition is x1, Xn should be minimal. In Other words, don't throw in some new junk that has nothing to do with the operation. Well, what we want to do, of course, is get rid of these. We Want a definition of merge, which has no conditions. But Notice we basically already have that the first one that follows from the no tampering condition. The No tampering condition has to be revised now so that it doesn't apply to syntactic objects, but to the workspace because all the operations are on the workspace. The No tampering condition says if something's in the workspace, don't change it. Well, the most extreme form of changing is to delete it. So You can't delete it. So For many reasonable interpretation of the general condition, NTC, which is part of the strong minimalist thesis, it follows. You're not going to lose anything. This One, we've gotten rid of. It follows from resource restriction, which is a special crucial property of organic computation, it seems, at least for language, but probably quite generally. Probably Related to the general brain activity of massively reducing the data available for computation. So We've gotten rid of this one. But This one implies this one. If You're going to add any more junk, it'll increase accessibility automatically. So Therefore, we don't need that condition. So Therefore, we can get rid of all of these. They All follow from plausible, in fact, necessary conditions on general computational procedures for an organic object. So That gives us the best possible definition of merge. And If you think it through, on principle grounds, and if you think it through, it restricts it just to the original intention, which was never captured by the actual formulations, just kind of in mind. Turns Out that what was in mind was actually created correct, and all of the extensions have to be barred. Now, I should say one word about one of the general conditions, kind of a meta-condition, descriptive adequacy. I've just been assuming that we want them. We Want the operations, of course, to be descriptively adequate. But That's not an innocent notion. You Don't know from data whether they're the right data. Descriptive Adequacy, and this is not just linguistics, all through rational inquiry, all through science, is a theory-determined notion. It's not innocent. You Get a lot of data in, say, chemistry. You Don't know. is this real data or not. There are two kinds of problems the data could have. One, it might involve too many variables, lots of other factors that you're not interested in. In Fact, if you just look at the phenomenon of the world, they're just worthless. Too Many things are going on. So You don't develop physics on the basis of just observation of the phenomenon of the world. If You're in. Silicon Valley, that's the way you do it. But I'm talking about science now, not Silicon Valley. So What you do is try to get rid of the data that doesn't really matter. It doesn't have to do with what you're interested in. But that's a theory internal notion. The Other problem is that you look at the phenomena that are around you. They Usually don't include the relevant data. They don't include the critical experiments, the kind that matter. These are all problems that were faced in the early days of the scientific revolution. And They were sort of settled for the sciences. But Linguistics and the soft sciences haven't really internalized it. So If you go back to, say, the 17th century, the Galilean effort to try to rebuild science on firm grounds, throwing out the neo-scholastic occult properties and so on,. what was important. And He had a hard time convincing the funders, the aristocrats, not the National Science Foundation, convincing the funders that there was some point in this. So It was very hard for the aristocrats to see why. You should care about what happens when a ball rolls down a frictionless plane, which can't happen. Why Should you care about that and not leaves blowing around in the wind, which you see all the time? That's a big move, actually. And If you think about the problem, it's not trivial. So Why is the rate of fall independent of mass? I Mean, if Galileo had done experiments, they wouldn't have worked. Too Many other things would have happened. So What he did was just a thought experiment,. neat thought experiment. Suppose You have two masses which are identical, two objects which are absolutely identical. And Suppose they fall. Well, obviously, they'll fall at the same rate. Suppose You bring them a little bit closer together. That's not going to make any difference. They'll still fall at the same rate. Suppose You bring them so close together that they actually touch in a point. Well, that can't change anything. But Now it has double the mass. So We've proved the theorem without an experiment. Most of Galileo's experiments, if you run through the dialogue and so on, are really like this. So For example, another problem that was bothersome is what happens if you have a sailboat sailing through the ocean. And You drop something from here. Where Is it going to fall? Is It going to fall here? Or Is it going to fall here? Aristelian Physics says it'll fall here. Sailboats Moving forward. So Of course, it'll fall behind where you dropped it from. Galileo Wanted to argue that it's going to fall here because the mass is getting accelerated with the sailboat. Suppose He had done experiments. I Leave it to your imagination to see what you would find about where the thing is falling. You get junk. So You don't do experiments. You Just do critical experiments, often just thought experiments. And For linguistics, that happens all the time. So You read the literature these days, linguistic papers, cognitive science papers, stuff coming out of Silicon Valley, Google. One Of the great achievements heralded is to be able to parse 95% of the sentences that you find in the Wall Street Journal. Suppose You could parse 100% of the sentences and get the right result with training. It would mean absolutely nothing. A sentence in the Wall Street Journal is just an experiment. Is This an acceptable sentence or not? You Don't care if you can match 100% of random experiments. That's of no interest. First of all, a lot of the experiments have the wrong data, too many variables. The Other thing is, they don't include the critical experiments, the kind you're interested in. Can You get parasitic gaps, for example?. Can You get garden path sentences? Well, It turns out when you look at the critical experiments, they fail almost totally. They get maybe 95% of the data, but that's a result of absolutely no interest. And A lot of the field is sort of going off in that direction. Even In the linguistic literature, you find that anyhow, without going on, this concept is not a trivial concept. It's not an innocent concept. A Lot follows from trying to understand what descriptive adequacy means as a theory internal notion. Anyway, we certainly want to be able to achieve the level of understanding that was reached in the 17th century in the sciences. I Think that's a fair goal to try to achieve, understand that there's something significant and serious and theory internal about what we call descriptive adequacy. There are other conditions that have to be satisfied, one of them. Most of them we just take for granted, but one of them call it stability. By That, I mean in the course of a derivation, a syntactic object can't change its interpretation. So For example, if you topicalize, say, Mary's book, I want to read, in the internal system, the non-externalized system, it's going to be Mary's book. But These two objects have to have the same interpretation, like this one. You Can't be saying, I want to read the book that Mary owns, but I'm talking about the book that she bought, let's say. That's sort of taken for granted. Same for ellipsis. If You say, I read Mary's book, and so did Bill, What Bill read is the same Mary, and if she owned it, it's ownership in both cases. So You have to have a general principle that's telling you that anywhere through a derivation, you can't change the interpretation of the expression. Doesn't matter for right now how you express this fact, but it's got to be somewhere in the deep inside the theory. And That has a lot of consequences. We'll come back to that. At This point,. notice at this point, we're getting into a very interesting area of the area where we have to identify what are called copies and repetitions. So Here, the two cases of Mary are copies of each other. Copies are symmetrical. The term is a little misleading, but you have to recognize it's symmetrical. So These are basically the same entity. They have to have a precisely the same interpretation for ellipsis for any operation. They could be copies. Like If I say John saw John, then these two are repetitions. If You look at the generation of the sentence, you had the same formal object, but they were generated independently, and they have nothing to do with each other. This one might as well have been Bill, let's say. That Distinction between copy and repetition is a tricky one. There's an interesting paper by Chris Collins and Eric Groat, which goes through, it's in Ling Buzz, I think. I Don't think they published it, which goes through a lot of problems in trying to distinguish copies and repetitions. I Think we can cut through all those problems in a non-trivial way. And Again, I'm going to leave it as an exercise for you to work it out. But If we appeal to a general, a very general principle, which seems overwhelmingly true, looks false in some cases. But In that kind of situation, it's reasonable to assume that if we understood enough, it would always be true, the principle that's sometimes called duality of semantics. What It comes down to saying is that argument structure, Theta theory in particular, is determined by pure compositionality. So In fact, the strongest conceptual reason, I think, for Dominique's predicate internal subject hypothesis, is that the subject gets a Theta rule. Dominique and Hilda have a lot of other arguments for it. But The basic conceptual argument, I think, gets a Theta rule. So Therefore, it ought to be determined in the general VP system. And In fact, anything that gets a Theta rule ought to be in this system. What About things that are out? And That includes functional categories. They have an argument structure, but they're going to be determined by just essentially by external. What About internal merge? Well, that's always yielding things which have no independent interpretation. So Internal merge is sort of determining aspects of semantics, which have to do with discourse with information structure and so on, not argument structure. But That looks like a very sharp distinction. And If we accept it, then it follows that when at the phase level, when interpretation is trying to determine what's a copy and what's a repetition, it can stand on a pretty high ladder. If Something's in a non-theta position, it's a copy. If It's in a theta position, it's a repetition. That cuts very sharply. Now, It does leave possible cases of ambiguity. If You think through the possible cases, there are some that seem not to be determined by this. But Here, I'll just give you a thesis and ask you to prove it in your spare time. It Turns out, I Think, that there is a kind of a conspiracy of other principles that solves the ambiguities. The Things that seem to matter are Verneu's abstract case theory, which makes distinctions that you don't see sometimes in the morphology. And That turns out to be quite important. Another is Ricci's left periphery theory, which assumes that there are actual positions, like topic, focus, and so on, that a raised element moves to. And Third, connected with this labeling theory, which tells you when those movements are legitimate, when they give you a real interpretation, when you have to move on, when you don't have to move on. I Think If you put all these together, it probably solves the ambiguity problems. But I'll leave that as another exercise. Very Interesting question. You Might think through. One of the tricky cases, which is not easy to deal with, is small clauses. So You might want to think about that. So It has interesting consequences when you try to think how that would work for this. But Again, I'll just leave that as a problem to solve. But If we can solve it this way, then we can solve the copy repetition problem simply by looking at internal merge and external merge. We'll say that every merge operation yields a copy. Nothing else yields a copy. In The case of external merge, the copies that are yielded disappear under the replace interpretation of merge. So You don't happen to see them. With Internal merge, they remain. And Then this phase level algorithm, based on duality of semantics, along with the conspiracy that language is kind enough to preside us with, should resolve the ambiguities of interpretation. That's the general picture. You Can think about it and try to fill in the details. Will? No. I Still don't understand this basic AB tree and the BC tree. The Two trees you put down right at the start, AB and BC. But Parallel merge. Yeah. OK, So let's. I don't understand why that's. You Seem to say that that was ruled out by something. Yeah. So Suppose you have AB and C. And. Now you copy B to C. The Way people draw it is like this. But We're not allowed to draw trees. What you're actually doing is forming a new object, BC. So Now you have two objects, AB and BC. Now You take this one. You Make it arbitrarily complex. Anything You want, you're allowed to merge this. This is accessible, remember. It's a copy of that. So You can keep merging it again. You Merge it to this. You Now have a copy relation here. But It violates every condition. It's back to the initial case. All The cases follow from the simple case. So Parallel merge is out. What's the part of the definition of merge that forbids the original? It blocks this. Not Adding accessible things. If You do this, you're adding this thing, which is accessible. And You're also adding this, which is accessible. And In fact, this, which is accessible. So You're adding three accessible things to the workspace. Actually, this one. These are all new, remember. They're not the ones we started with. There are copies of it. They're new objects. And They're now accessible. This One, personally, because this is the one you've moved. And Now you have two copies of it, both of them accessible. Plus the pair, which is accessible. So You're violating resource restriction, which is the crucial condition. Most Everything is following from resource restriction, which I think is probably a deep property of organic computation. Same is true of side-bridge merge. Collapses for the same reason. Notice In duality of semantics, that has consequences that may be objectionable. So For example, it seems to rule out Norbert Hornstine's theory of control. Norbert's interesting theory of control relating control to raising raises a controlled element to a new theta position. So that's giving internal merge to a theta position, violating duality of semantics. So Here we have a problem. Either Norbert's theory is wrong, or duality of semantics is improperly formulated. OK? Is This the same reason why side-bridge movement is ruled out? Side-bridge movement, same reason. Same Reason, OK. Even More reasons, because at least this, it does form a new object with more accessibility. But There's also the question about how you connect these two separate things, which is another problem. But At least it has the problem of more accessibility. That One runs across all of the extensions of merge. Look Through them, all of them have this property. So They all basically reduce to this very simple operation that I had down here somewhere. In the simplest case., That serves as a paradigm for just about everything. So We're now down to, with various problems hanging around on the side, like what about Hornstein's theory of control? We seem to be converging on exactly what we want, some the simplest possible operation, conceptually justified, which gives us exactly the cases that don't yield illegitimate operations and a limit to the cases that do. Well, what about all the problems that are left over? So It takes a across-the-board movement. The Parallel merge and multidimensionality gives you interesting ways of describing ATB. Notice describing, because none of them are legitimate. But At least you can kind of draw graphs that kind of look as if they're doing something, even though they're not. So What do we want to? A Clarification question. Yeah. Is The point,: was it so doing parallel merge or whatever similar side we're going to merge, you want that to be bad due to the constraints imposed by whatever gives us a restriction of resources? There's not enough stuff accessible to do that. The Way that we restrict resources is just via phases, I guess, if I understand things right. Does That mean that there's a phase-y explanation for the line? You can't have any operate. As Soon as you look at the operation, if it adds accessibility, it's out. Because There's a meta-condition on linguistic operations, probably on general organic computation, but at least on linguistic operations, which says you can't add resources. When You have an operation, it can create one new accessible object. The Object that you've constructed can't construct any others. Is Agree Also sensitive to this constraint? Agree is an operation, is he also sensitive to it? Agree has the same property. So It's also a big position. I'm going to generalize it. It holds also for Agree and labeling and other things. But The interesting cases merge. But None of the operations should add new accessible items. This should be a general property. You have the two AB and BC. Why are the BCs the same? Parallel merge case, yeah. So You end up with two element sets, which you see, AB and BC. Why Do you call the elements of the BC? Why Do you call those accessible? Accessible, There's no reason why they're not accessible. Unless You stipulate somehow that they're not. Why Are they accessible? If This one was able to move from here, then it's able to move from here. I Mean, it's the same structure. Unless You impose some other rule that says I'm not allowed to move. But There's no reason for that. Remember, this is a copy in IM. Copies are always allowed to move. So Unless you simply stipulate, unless you simply stipulate, I'm not allowed to move, which gives the game away, then you're not allowed to have the operation. Because There's no basis for that stipulation. It was originally, notice that it was originally allowed to move. That's the whole point. So Therefore, it should still be allowed to move. It's the same structure. OK? Both Move now. There are two of them. Yeah, either one or the other. So I Don't think there's any way to solve it. It's just out, like all of them, and late merging for further reasons. Well, let's go on and take a look at the things that the empirical cases that were described by these illegitimate operations, like, say, ATB. So What did John buy and read? Well, how do you generate this, first of all?, Well, it would start with, I'll forget the reduced order, and all that stuff, simplify it. What John Bought, What,? let's remember we're talking about the internal core language, not the external language, and what. John Read, what. That's the thing that we generate internally. Well, why doesn't that give the interpretation, standard answer? Because The thing that he read might be different than the thing that he bought. So It's not giving it the right interpretation. However, think it through for a second. This is a copy. This is a copy. This, of course, is a copy. This is a copy. What is the property of copies? Property of copies Is they all delete, except for the top one. That's a general principle of computation. So Like in successive cyclic movement, you delete all the copies. There are languages where you lead some residue somewhere, but we'll forget about that. Basically, you delete all the copies. That's general computation. So That means that in the externalization that we delete this, of course, we delete these. Well, notice that that gives you the right form. It gives you the form. what did John buy and read. Does It give you the right interpretation? Well, In fact, it does, because of the principle of stability. Remember This principle. It says in ellipsis or topicalization or anything, in the CI system, you can only delete if you have absolute identity. Otherwise, you just can't delete. General Property of interpretation at CI. Deletion requires perfect identity. But Notice that that gives you ATB automatically. Nothing else to say. If You think about parasitic gaps, essentially, it works the same way. There's actually an interesting paper by Rene Huyberg. It's not in print, unfortunately. But It's circulating around the internet, which goes through the details of it. But In fact, the basic idea is pretty straightforward. So ATB and parasitic gaps get straight. Notice Also, in the case of parasitic gaps, it follows that A movement won't yield parasitic gaps, because you don't have the initial WH phrase, which will allow the second one to be a copy. So Parasitic gap with A movement would be something like John made a sandwich, which book did he file, or something. Professor? I Don't understand. I Understood you to be saying that in the case of what did John buy and read, it's the same thing that he bought and read. Has to be. Otherwise, you can't delete it at the CI level. Right. But We can say things like, what did John buy and married? What did he read? Oh, that's fine. There's nothing wrong with this interpretation. It's just not ATB. So How do we derive, what did John read and marry each? You Just decide not to call this a copy. Call this something. Call this a repetition. You can always call something a repetition. The book was John saw John repetition. Generate them separately. If You don't say anything, one of the options is ATB. Another option is the other interpretation, which is all you want. And Notice that if you use the ATB interpretation, then you're deleting. So If you're giving the ATB interpretation, you're reading it as ATB with a deletion, because the deletion is dependent on the stability. That's quite understood. I Didn't understand a sentence like, what did John buy? And what did he read? No. What did John buy and marry self? That's fine. Well, it can be different things. It's just totally different sentences. What did John buy and Tom went to the store? You can generate that. But There's no question in the second part. But It just has the interpretation it has. But Isn't that what you just said? Doesn't that force the meaning that there is a thing that John bought and married eight? If You have two WHs, which are copies by definition, because they're in IM non-theta positions. So Therefore, there is the option for them to be interpreted as copies. And if they're identical to delete, which gives you the correlation between the ATB semantic interpretation and the externalization with deletion. You don't have to delete it. You could be saying, what did John buy and what did Tom eat? That's a fine sentence. Just Nothing happened. I have a follow-up question. So I think if one person asks, what did John buy and Mary read, it's possible to answer that question. John bought a book, and Mary read a journal, something like that, no? Not if it's ATB. It has to be the same thing. That's the point about ATB. If You delete, if you delete, you can't get the different interpretation. It just feels like a normal dialogue. And Then it would mean that the first was not an ATB interpretation. What did John buy and Mary read a book? No, no, no, no. What did John buy and Mary read? It's the same thing. Is It possible to answer if John bought an apple and Mary read a book? That's a fine sentence. But There's no doubt. And you have an answer to the ATB question. No, if either it's ATB or it's two independent constructions, that's what ATB is. Think about it. Otherwise, there wouldn't be any ATB phenomenon. No, I Think that the facts are, at least I Think that for many people, what did John drink and Mary read is fine as a question. Well, for those people, they don't have ATB. We're talking about ATB, the phenomenon, right? If Somebody says my language, sorry. But We're talking about people who have ATB, which has this interesting property that you delete the things and interpret them identically. That's the whole point of ATB. So. We're not going into the question of, does somebody have some different language? We're talking about the languages which have a problem to solve, namely ATB. If. You don't have the problem to solve, you don't do multi-dimensionality either. Are There speakers who have both ATB and the meaning ATB meaning and the one which you're saying is the non-ATB meaning? So What did Mary drink and John buy? For Me, that's ambiguous. It could mean that, obviously, for some bizarre reason, Mary drank and John bought the same thing. Or It could be John bought a Cadillac and Mary drank a Pepsi. I Mean, if that's your interpretation, then you don't accept the multi-dimensionality analysis either. Oh, no, I don't. OK, fine. That's fine. I Don't care. I Mean, if somebody has a different system, we can try to describe that one. But I'm talking about ATB. OK, if you don't have it, fine. No, you don't. You have something different. You have something that is different from what is described in the ATB literature, OK? So Then we can talk about that, yeah. But I'm trying to say that what is captured by parallel merge in the multi-dimensionality interpretation, in fact, is yielded automatically with nothing, OK? Same with parasitic gaps. If You look at parasitic gaps, there's a million different problems about them. And This doesn't address those problems. It's just saying the basis, the very basis for parasitic gaps, we already have without anything, including the fact that they are conditional on WHO movement in the first sentence, not A movement. All of that follows right away. Then You get into the morass of problems about different kinds of parasitic gaps, OK? A different kind of question. Yeah. I'm sorry, one more question. I Believe in ATB. It's not the same question. So Do we end up with a system where, if you're in a Theta position, if you have two things in theta positions, they are either copies or repetitions, and you get to choose which one? No, see, that gets back to this conspiracy that I just talked about. If It's John likes John, then they're repetitions. What I'm saying, what I'm giving you as an exercise is to show the following.: And I Think it works. That If you accept duality of semantics, the general principle, and you look at these properties of language, I Think you end up resolving the ambiguities, determining what are copies and what are repetitions. That's a claim, OK? It's up to you to falsify. So A movement, no parasitic gaps, but A movement is claimed to show across the board movement. So This is just dealing with ATB with WH movement, OK? There's other questions about A movement. Yeah, but I'm not talking about this. The Main ones are the ones in the multi-dimension element. The line of reasoning that rejects parasitic gaps with A movement. That explains why you don't get it with A movement. You Don't get it with A movement because the element in the gap, the operator in the gap, has nothing to be a copy of. So These two things are just totally independent of one another. It's like John ate a sandwich before reading. This doesn't mean anything. But It can do it in the case of across the Board. Like, would that also lead us to expect that there is no across the board A movement by the reasoning provider? Across The board A movement is a different problem. It's not dealt with in the multi-dimensionality literature of the kind that I'm discussing. I think it works, but it's basically the same. That's a fair question. We should look at that. Well, this question of the. I have a question over there. Sorry. I have a qualification question about parasitic gaps. So in a sentence, what you don't read after buying? What Did John read after buying? What did John read after buying? The gap in buying, what is the higher copy of it that you would like? Well, that would be something like, what did John read before what John buying? And The two what's are copies. So You get the same phenomenon. If You want the details, look at Rini's paper. But That's basically the structure. And It gives you the core properties of parasitic gaps. It leaves lots of questions open about different kinds of them. Well, let me. Any more? Yeah. The Question might have been had in mind was to do a comment on a sentence like, what two graves were one friend buried in on Monday and the other friend buried in on Tuesday? This doesn't go into that. And There are many other questions about interpret. Anybody Who's interested in the kinds of questions that Barry's raising should? look at the Book, the longest book in the literature with the shortest title. It's called, it's called, And, and the author is sitting over there. It gives the hundreds of pages of very interesting examples of ways of interpreting conjunction and complex structures based on a kind of a neo-Davidsonian event calculus. And There are tons of problems there that are really interesting to solve. But What we're interested here in is asking, what is the basis in the structures for yielding those interpretations? Barry actually doesn't go into that question in the book. I Think you tell me if I'm falsifying it, that Barry describes it fundamentally as a kind of conjunction reduction. But If you think about that formally, it can't mean, I Think, this is a hope, a friendly amendment to the book. It Doesn't mean that you first generate all these huge conjunction things, infinitely many of them for a short sentence, and then get the syntactic structure. It Must mean, that's the thing I'm going to get to next, that you have a syntactic structure. And There are some kinds of interpretive rules that we have to figure out that give you this mass of amazing stuff that you find in there. That's the challenge to face. And If the event calculus approach is the correct one, that will yield the event calculus interpretations of the structures that you generate. Unfortunately, Barry tells me he's now working on another book called, But I Hate to Think About That. OK. Question Here, how do we handle copies, not copies, resumption under this? How Do we handle resumption? How Do we handle resumption, resumptive pronouns? We don't in this system. That's something else. A Lot of things I'm not talking about. OK, yeah, it's a fair question, but not talking about. Not That there's any other way of talking about it, right? It's just that if it exists, we're going to want to have an explanation of it in terms of operations, which meet the austere conditions of explanation. That's the general point. Whatever Problem You're working on, whatever it may be, phonology, semantics, syntax, morphology,. If you want an actual explanation within the context of a program that regards language as part of the natural world, if that's your framework, you're going to have to have explanations into terms that meet this highly austere condition of learnability and evolvability. And About the only thing we know that meets those conditions is merge. So If we can account for things in those terms, like, say, ATB and parasitic gaps, we're in business. Otherwise, we have problems. OK, let me turn to another one, which is problematic and is related to what Barry was just raising to get to it. Well, let me just make one more comment about this. I Won't bother spelling it out. You'll notice that this is a sketch. I Haven't really formalized it, but you can figure out how to formalize it. But When you're left with this definition of merge, the simplest one, and the one that I think is principled, then you can reformulate it in the usual style of transitive closure, friggin'' ancestors. Like, If you're characterizing the set of integers, the standard way of doing it is to say the set of integers includes, say, one, and it's the least set containing one and the successor of any member of the set. That's the standard transitive closure ancestral definition. The Analog here would be the set of workspaces for a given language is the set, not the least set. The Set We leave out least, which includes the lexicon and merge of any triple. That's the set of workspaces. We Don't have to say least, because that's already incorporated in resource restriction. Otherwise, you get the standard recursive definition of the set of workspaces. So It sort of fits the norm. You Can work out the details pretty straightforward. Instead Of that, let's go to something else. There's pretty good reason to think that in addition to merge, which maybe we've now got in the optimal form, there's probably another operation, at least one other operation. That is asymmetrical merge. There are strictly asymmetrical structures, like, say, Young Man. Young Man, the structure is a noun phrase. The Two elements in Young Man are inaccessible. You Can't extract this and leave this the other way around. So We have an asymmetric structure where young is attached to man, and the whole result is still basically man,. adjuncts, essentially. All Adjunct structures, I Think, require pair merge, which is the next operation to look at. Now, There's a very interesting property of pair merge, which has been a thorn in the side of all generative systems since the 1950s, namely unbounded, unstructured coordination. So Things like young, happy, eager, go to Harvard, you can have an unbounded, unstructured coordination. This is a real problem. You Can't generate it by phrase structured grammar. Even Unrestricted rewriting systems, which are universal, in the standard interpretation, don't give you the structures for unbounded, unstructured coordination. Now, notice that since they're universal, you can code it, but that's not interesting. They are universal Turing capabilities. You can find a coding for it, but that's not of interest. If You look at the generation by Phrase structured Grammar, you need an infinite set of rules. It was thought for a while by George Miller and the papers back in the 50s that you could get around this with generalized transformations. But Howard Lasnick had a paper showing that the same problem arises. You'd have to have infinitely many of them. So You can't do it by phrase structured grammar. You can't do it by transformations. There's no way of generating it. It's been a big problem all along. But Notice there is a way of dealing with it in terms of paramerge, namely super-multimensionality. So You have, say, man, and you link to it with any number of adjuncts. They're all on different dimensions. But There's no limit to the number of dimensions. You can paramerge to the element. There's no reason to believe that, just because blackboards are two-dimensional, so is the mind. It does whatever it does. So It could have any number of possible dimensions attaching to a particular point. So For simply, that would give you something like unbounded, unstructured coordination. This can incidentally become extremely complex. Here We get into Barry's type of questions. So For example, one of the conjuncts could be a disjunct, could be John is young, angry, either going to Harvard or to go to MIT, so on. So You can have unbound, and the disjunct could also be unbounded. So You can have unbounded, unstructured disjunction inside of unbounded, structured, unstructured coordination. And This can yield incredibly complex structures. I Leave it to you to give the semantic interpretation of it. But In the nature of the system, you can see that this is possible. Well, instead of trying to, actually, the formalism is not very difficult, so I won't go into it. But Just take the simplest case and take a look at that unstructured. If We can deal with unstructured, unbounded coordination, Then the simplest cases of adjuncts are just automatic. They're the case where there's only one element instead of an unbounded sequence of elements. So We will get simple adjuncts if we can handle the unstructured case. So Let's take a look at that. That's the essential case. Notice A few properties of it. For One thing, it matters what the order is. So If I say, if the order of the adjuncts is young, angry, that's different from angry, young. The Reason is because of something that Jim McCauley noticed back in the 70s, the word, respectively. So If you think of structures with, respectively, then the young, angry man ate the turkey sandwich, and the young, angry men ate the turkey sandwiches and the chicken sandwiches, respectively. The Order of the adjuncts determines the nature of the interpretation. So Somehow, the object that we have in unbounded coordination is actually a sequence. Furthermore, it can have iterations. Like You can say, John is young, angry, young, tired, and so on. You can iterate them. So Basically, the problem is we have unbounded coordination or disjunction either, which has a sequential structure with possible repetitions. And That sequence is interpreted both at the CI level and the externalization level. Now, this does not tell us that linear order enters into syntactic operations. It just tells us there's some object being constructed, which is going to be interpreted in terms of its order and spelled out that way. So We're not crossing the barrier into believing that externalization feeds CI. That's important, even though there's order involved. So What we have to have is something that works sort of like this. If You just think of the general properties, you have to be able to pick out a set of things that are going to be adjuncts. And You have to form from that set a sequence where the elements of the sequence are drawn from the set, but in any possible way. That requires an operator. It's actually an operator that's familiar in logic. It's Hilbert's epsilon operator. In Hilbert's formalization of metamathematics, the core operator that he bases it, is this epsilon operator, which says that out of a set, you can pick an element. Basically, an. It's like indefinite articles. So We need an operator like this, which tells us that given a set in the generation of an expression, given a set, you pick out a sequence. And Then somehow, the elements of this sequence link to something. Each of them is going to link to it independently. So If I say young, angry man, the man is both young and angry. So Independently, they're going to link to something. So What we're getting out of this is a set which, first of all, it'll have to be identified as either coordination, or conjunction, or disjunction. So We have an element here, call it k, which will be plus or minus. Conjunction will be one or the other. And Remember, they can be interspersed, but that's just more formalism. And This sequence will include the pair-merged elements, y1 and some link that it's linking to, all the way up to Yn, and a link that it's linking to. And These links have to be identical all the way through. Like If one of them is a Wh phrase, they all have to be Wh phrases. If You think about unbounded coordination, you can't stick a question in one of the positions. And Of course, you can't have different links. So We have an object that looks like that, and that has to be merged into the general expression. That's the formal problem of dealing with a junction. I Won't bother spelling it out. It raises interesting questions. So For example, one question is, what do you actually link to? So Suppose you have a coordinated noun phrases. John, Bill, Tom, Mary, the guy I met yesterday, et cetera, et cetera. Each of those things is going to link to something. What Is it going to link to? Well, the natural interpretation would be that the individual items here, Y, K, L, if it's a noun phrase, they should all link to whatever is common to noun phrases. Some Thing, call it N. Notice Here that I'm not using the DP analysis, which I think is a mistake. I won't go into it here. But It seems to me that nominal phrases should be regarded as nouns, not as determiners. Determiners are probably something that hang off the outside. And Definiteness is probably a feature of the whole noun phrase. So I'm assuming that the Semitic is the universal language, that the determiner is just a feature on the noun phrase, which distributes somehow differently in different languages, depending on externalization. So You have a feature of the noun phrase, specific, nonspecific. You have a structure which is basically N, with determiners hanging on somewhere, adjuncts that you don't care about. What Is this N? Well, Here we get back to something that Hageet suggested years ago, that the basic structure of language is, again, kind of like proto-Semitic. You have roots which are unspecified as the category. And Then you have categorizers that determine what they are. So For example, N and the root probably paramerged, probably in the lexicon. That's probably the first operation, going back to the paper of yours. The First operation is probably a lexical operation. There are many operations inside the lexicon that involve merge-type operations. One of them is probably categorization. But Notice that this N that's determining that this root is a noun can't be the same as this one. Same with the verb. The V That's determining that something's a verb has to be distinct from what we usually call V or V star up at the phase level. Those are just different elements. They shouldn't be confused, I Don't think. In Fact, these, the ones at the phase level, I think should simply be regarded as phase markers, independent of category. Categories Decided down below, probably in the lexicon. At The phase level, you have something saying Amaphase, a phase marker. And Notice that if you take a look at noun phrases and verb phrases, they have some interesting similarities, some differences, but some similarities. One Similarity is that both noun phrases and verb phrases can be either call it strong or weak with regard to extraction. So The complex noun phrase constraint, it's well known, is strong for specific noun phrases. So It's basically inoperative for non-specific noun phrases. That's the same, sounds like the same distinction as between strong phases and weak phases. So transitive verbal phrases are strong with regard to extraction. You Have to move to the edge. Weak Ones, you don't have to move to the edge. It looks like the same property as weak noun phrases. So Possibly, what we have is something like this. Going Back to classical Greek grammar, philosophy, linguistics, didn't distinguish, we have the notions substantive and predicate, which gives us a four-way classification of things that are substantive, non-predicate. Those are the nominal phrases. Substantive, predicative, adjectival phrases, non-substantive, less predicate, verbal phrases, and then non-either, which is all the junk, prepositional phrases, and so on, some structure like that. And The phase, the crucial phase operations seem to be restricted to those that are the perfect elements, pure substantive, pure verbal, with either the strong or weak property. Now, one of the curious distinctions between noun phrases and verb phrases, which has kept, prevented the thinking of noun phrases as phases, is you don't get the normal escape hatch. But I Discovered a couple of days ago, thanks to somebody, that Uli has an escape hatch for noun phrases. So that fits in the gap that we were worrying about. So Let's say Uli and proto-Semitic are the four languages, and that would have put noun phrases and verb phrases together, ending the idea of using the same notion for the categorizer and the phase marker. They're probably different notions. Adam, yeah. I Would like to go back to the definition of merge, as you mentioned. The Original definition. They're the first. No, the one you put on the board, the first one is not. Per Merge or merge. And you said PQ And the work space. And This is the stuff I Saw. The Original one. Yeah. If You run the clock backwards to the first merge, the stuff you talked about. And Q then becomes the work space, right? Because In that definition, Q equals the work space. There's nothing. The Work space is empty until you put in there. If You start with just two things. No, you start with one thing, right? Before You start merging, there is a work space which is empty. Well, there's no work space until you put some. The Work space is a set. OK, so the work space is the set. PQ in this definition. Because Why do you have three elements? This is the confusing part. The Work space can't have PQ unless they get into the work space. You Merge them. So How do you get them into the work space? OK, So in this recursive way, Q is the prior work space, right? So At some point, you finish merge, and you have a work space. Before You start another set of merge, right? I Don't follow. Before You do any merge, you have nothing. You Just have a lexicon. But In order to put P, let's say, and Q even, into the work space, the work space has to be prior of P and Q being there. It just means there is the option of creating a certain set, which you can put things in if you want. OK, So the set, the work space is essentially something that comes out of PQ being put together. Yes? The Work space has nothing in it unless you put something in it. OK, And then the question is, so the work space, at some point, there's Q, though, if you have another. Is Q equals Q, right? No, only if you've put Q into it. Yes, So the next step, let's say. And You have another merger, which is, let's say, P1, Q1. Then Q1 was the previous work space. I Don't follow. Let's imagine that we haven't. We're just beginning the computation. We Take P and stick it into the work space. Now The work space is the singleton set, including P. Now We put Q in the work space. Now It has two elements. Then We decide to merge them. We Get a new element, set PQ. But not P and Q. There is an interesting empirical question here. How Do you start? And There are various options, and they have lots of consequences. One Possibility is that the only thing that goes into the work space at the beginning is things you've already merged in the lexicon. So If the first, remember, inside the lexicon, there are constructional operations going on, like the words in the lexicon already have structure. Part of the structure, if Hageet is correct, and I'm assuming she is, is taking a category like N, V, maybe broken down into substitution and predicate, and categorizing a root as one thing or another, an operation inside the lexicon, which is giving you a pair, like the pair V hit, let's say, the verb hit. Then That thing can be put in the work space. It already has two elements paired. Then You could put something else in the work space, begin to merge them, put more things in, build up the work space, and so on. Is That the account for the increasing accessibility constraint? No, it doesn't. Because You're basically pre-packaging things to make them accessible for merge. Yes, but you're going to have to. Every Operation that you carry out is going to create something. That's what an operation is, create something. Now, the resource restriction says, don't create too much. Create As little as you can, at least the thing that you're forming, but nothing else. OK? On The accessible? Yes, because I Noticed that when you put in the pair root verb, neither is accessible. Because An adjunct structure, That's what I said before, if it takes a young man, you can't extract young, you can't extract man. Notice, incidentally, that this approach, I mentioned earlier, less time, that there's a way. I've mentioned a paper of Jelko Boskovich, pointing out a way of putting the adjunct island and coordinate island problems in the same package, making them essentially the same problem based, again, on the idea of event calculus, which is a somatic event calculus, which treats a junction like coordination. So It kind of unifies the problems. And That happens automatically here. The Pair merge structure gives you both the islands of conjunction and the islands of adjunction. Now, notice that it leaves the mysteries, just like Jelko's paper does. So If you look at, say, the adjunct island effect, which Jim talked about years ago, it has interesting properties. There are some languages where you can't extract the adjunct. There are other languages where you can't extract from the adjunct and other distinctions of that sort. Those are interesting problems that remain. Furthermore, if you look at adjuncts, they're not uniform. There are some kinds of adjuncts which you can extract from. There are other kinds which you can't extract from. So The notion adjunct is too diffuse. We Have to sharpen it further to find different kinds of probably different kinds of pair merge. Those Problems all are sitting out there, more problems to solve. But You begin to get a unification of the problems. The Adjunct Island, conjunct island effects do reduce the same structure, this sequence that you pick out by the epsilon operator. Now, As far as the point that you're making, as I understand it, it does raise, It leaves open some questions about how you get things from the lexicon into the workspace. There are a couple of ways of thinking about this. And They have different consequences. One Approach is just to take something in the lexicon, insert it in the workspace, and then go on from there. Another Approach is to take something from the lexicon, and to merge it with something that's already in the workspace that's formally slightly different. It has different consequences when you spell it out. You may need both. But Those are questions that you want to resolve, certainly. But I Think it's plausible to believe that the whole system of operations begins by just forming categorized roots inside the lexicon, then building up from there. I Don't see a problem with that. I Don't see the problem you're raising. These categorized roots are completely invisible to syntax. Yes, because the categorized root itself is, but not the part. You can't just raise the root and leave the categorizer. Or conversely, because they're parametric. And That's essentially the idea of development. They don't have to be just moved. They can be targeted by a Greek, for instance, or other things. Targeted by a Greek, for instance, or other things. So They're also invisible to any operation, not just moved. There's other sort of syntactic operations, which will look at the root and move. That raises other questions. Here, we're talking about accessibility to merge. There are questions you can raise about whether you can have agreement into an adjunct, let's say. That's a different question. Here, we're talking about accessibility to merge. Lots of other questions. Let Me: just get in kind of late. So I Mentioned a couple other things that you might deal with in terms of paramerge. There's lots of interesting questions hanging around that have a potential, I think, paramerge analysis. So Let's take one that's been a crazy problem for a long time. There's a strange restriction on extraction from causative-type verbs and perception verbs. So If you look at structures like John Sawville walking down the street, you can passivize this. You can say, Bill was seen walking down the street. On The other hand, if this was a bear verb, walk, then you can't do it. You can't say, Bill was seen, walk down the street. This also holds for the kind of causative-type verbs, verbs like let and make. They don't have the full paradigm, but they have part of it. I saw John, I let John walk down the street. But You can't say, John was let walk down the street. Now, there's a long problem in the literature about how to deal with this. The Only partial solution I've seen is a paper by Norvin. I don't know if it's in print, even, left. The Paper by Norvin in terms of contiguity analysis, which gives a description of how you could lock back. OK, I mean, come back in. I'm insulting you. So The only paper I know that says anything about it is Norvin's paper, which this is the blocking of pacifization out of perception, verbs and causative verbs in terms of contiguity theory, which is an interesting description. But It doesn't cover the whole set of data, because the fact of the matter is, you get the same property without extraction. So For example, in English, these structures are a little bit odd. In Other languages, they're normal. But Things like, there were seen last night, three men walking down the street. And You can't say, three men were seen. You Can't say, they were seen last night. three men walk down the street. So Even without extraction, you get the same property. So It can't be based on extraction. It's got to be based on blocking pacifization. Now, if you think about it, with the let-make type verbs, you can think of those as being basically causatives. They have essentially a causative structure. And In fact, the verb cause itself is kind of resistant to pacifization. So John was caused to leave and that sort of thing. So Suppose we think of the let and make as being essentially causative affixes, the kind that show up in many languages. That would mean that they're pair-merged with C, probably in the lexicon. Well, that gives a unit that's invisible to the operation of pacifization. It's a pair-merged element, which is resistant to whatever we think pacifization is, maybe eliminating the case structure. That would block both the in situ cases and the raising cases. In Fact, it's the only way I know of dealing with that. Now, it's natural for let and make, because they are kind of quasi-causative. But There's a very interesting question that goes way back as to why perception verbs should act the way the causative type verbs do. Actually, Jim Hagenbotham has an interesting paper on this in the 60s, 70s, 80s, I guess, in which he tries to argue that the complement of the perception type verbs is basically some kind of nominal expression with a there verb. Maybe That's an avenue to explain it. But At least using the device of pair-merge, you have an opening to try to account for this strange phenomenon. I Don't see any other way of dealing it. Another Kind of case that's quite interesting is head movement. Head Movement has always been a terrible problem. It doesn't have any of the right properties. It doesn't fit anywhere in the movement system for all sorts of reasons. There is an approach, an interesting approach, by Pisa Kitahara, as the paper is on this, in terms of pair-merge. I'll just give the simplest case. Take T to C movement. So You have a structure of C, T, V, whatever. And At some point, this moves here. How Does that work? It's one of the cases of head movement. Notice That the thing that's moving is really not T. It's V. This is an error of the traditional head movement analyses. But The thing here is usually described as a T with a V adjoined to it by an adjunction operation. But It's actually the other way around. It's a V with a T adjoined to it. One Of the reasons the traditional adjunction operations just don't give you the right result. There are many reasons. What Pisa suggests is that when you get to this point, you've created this object. You have a C. And Then the next operation is to form C, T. Notice That the elements of C, T are not accessible because that's a pair-merged adjunction structure. So You've only added, you've actually enlarged the workspace. But You've only added one. Actually, you haven't even because you've taken C and added T to it. You've kept the workspace the same size. But The only accessible thing you've added is this. So That's permitted by resource restriction. Then The next operation is, you've got this thing. And This thing is just to merge them. When You merge these two things, you get what you wanted. The Structure, C, T, with T down here, V. So That gives you a possible way of looking at head movement. Notice It has a problem. It has the same problem as all of the examples. back to the original one of resource restriction. What happens if you then make some new thing here, x? You start building it up. It ends up being of the form T, V. And Then you decide to merge this to this one. That's crazy. Doesn't make any sense. What blocks that? That's the same paradigm we always have. Now, this gets kind of complicated. But I Think there's a way out of this problem. And I'll leave it to you with something to think about. There's a way out of it by sharpening the notion of restricting computation so that it tells you at each point to add as little as possible to the workspace to still continue. If You think about that, it gives an interesting direction into perhaps blocking this option. It amounts to a condition that will say, you're going to have to merge this one before you create something new. Now, you can't make that too strong, or you won't be able to build up exocentric constructions. You have to put conditions on it that allow just the right ones to block the wrong ones. I'll leave that as another exercise to the reader. HISA has a paper which doesn't go into this. It just gives the proposal. Well, there's a lot more that could be said. I Think I'll stop at this point. These are the kinds of problems that arise when you try to give a principled approach to the nature of explanation. You get some interesting results, get a hoard of problems. The Problems may be presented in an organized form, which is helpful, but we want to go on to try to find real explanations for them. Sometimes You can, as in the case of unifying compositionality and movement or structure dependence. or the basis for reconstruction, things like ATB and parasitic gaps, maybe some of these things. But There's a mass of problems out there to try to deal with in a principled fashion. So That's why it's an interesting field. CHEERS AND APPLAUSE. |
Common Values
| Value | Count | Frequency (%) |
| Welcome everyone! This is natural Language processing for Law and Social Science. Thanks for joining remotely today. It still is a bit up in the air how we will do the hybrid verses in person versus Zum format. This term, you hear, I'm a little stuffy today. This is true. Avoid nineteen case I Caught it from my daughter who caught it in daycare. It's very mild so I hope if any of you catch it, it's not worse than this. It's definitely manageable. You can see I'm here to teach even though I have it so it's not that bad. I'm a little congested and we might end a bit early today. I Hope that's all right, but going forward. I Would like to do the course hybrid where we have some impression meetings at least, but before text money you'll hear from me about that. Thank you for all of you who filled in the students survey. There was a broad agreement that we should have an online aspect or at least should be recorded so well. we will work with that. So I have a few slides to introduce the course and then we'll have a chance to answer any sex questions about the format. So this course is a applied Natural Language Processing. It's not a course where we will just start with different texts, data tools, or different help models and learn how to come them up in there. We care just as much about the applications of those methods in the law and in social science and this is in the news all the time. Here is an example from a recent legal product called Clarity which uses in all tools to to analyse contracts and for example terms of use to highlight different clauses that are unusual. You also hear these really exciting ideas such as the World's First Robot Lawyer I'm excited about this, I think everybody should you. I think that this technology is is improving rapidly and dramatically and there is scope for many interesting things happening in the law and inoup world. But there also is a lot of hype and we will take a skeptical view of of these strong statements such as the World's First Robot Lawyer And in turn I Think that while there is a lot of interesting about tools coming about for law and social science and other applications, I Do not think we're close to having a judge to be replaced by a contact. some other reasons to be skeptical or to be concerned about the arrival of these legal inopetuls is that they can be biased So northwest. One of the classic examples is different languages have different degrees of being rendered having notions of gender, mail and female in the language and if you translate from English such as she is a doctor he is a nurse to Turkish which does not have notions of gender pronouwns and then you translate back the gender switches so basically they they have since fixed to this in google translate but it used to be where if you to see as a doctor translated to Turkish and then translated it back it will change to him as a doctor just because similarly he as a nurse would be transformed she as a nurse and this is just because. Theories this basis this statistical correlation in language where doctors tend to be male and nurses tend to be female and statistical language models and translation systems will capture that bias. These issues are based. the language models are as the technology comes more powerful, these issues become more intense to get more intense benefits but also more more intense risks and good. It is now a few years a few years old but this is language whose hole it came out in the Tousadand nineteen that could. It was basically among many other things a fake news production engine and it could produce a lot of context appropriate prose. So you can imagine to know Twitter and email. can the news being filled with all this machine produced speech that would drown out all other speech and I think that those those concerns are still relevant. but now that got two has been out for it for three years and there's an even better version called just There that's been out for a year and we have not seen the internet employee. That means that maybe it was not as bad as we thought and so in this course we want to know. Can we take legal Gptto illegal? Get there to help judges in their work? So this is the course. It's natural Language processing for law and Social science and our engineering goals are doing these kind of two pieces. This course we're going to develop skills in applied natural language processing which will include machine analysis, interpretation, generation of documents and those could be on news articles or contracts or judicial opinions or political speeches. And we want to also take a social science approach where we do not just care about sequencing language in different ways, we care about relating it to attend data, and to understand the social forces that underlie these documents. What are their determinations and what are their outcomes and so you knowsome. Of the examples that we will frequently come back to are: what are the motivations for judges? Can we analyze judicial opinions and see what's driving their decisions? Can we look at the writings and the speeches of politicians and to the end where they're coming from And I Think this is kind of the broader agenda or the goal in this research area is Knpowders language matter in society in human relationships s and what can help do to help understand that? So what we will do. We're going to read text documents as data so there's you know many ways to do this and will go over many of them. We will use supervise learning techniques for dimension reduction, topic modeling, groups interpretation, supervise learning for text regression and text classification can be predict from a speech. Is this form a left wing politician or a right wing politician will get at World embeddings, document embeddings, a lot of exciting technologies there for producing these learned representations of language of words and concepts in geometric spaces. and towards the end will get into disclosure analytics. So this is where the linguistic side of natural language processing, cynicism, and a summarization question answering I'm checking. These are all really relevant to legal applications for example. So some course logistics or beating times that will be two even fourteen to it in an sixteen so we'll have a ten minute break in the middle going back to what I started mentioning at the beginning. These are the survey results for the the course format and there were only a handful of you who would be register if there only online and everybody else wanted some online aspect or the indifferent. The based on these surveys we will certainly have a online component. like everything in the class will be durable online. but I'm hoping that we can have some impression component as well. so there's a question in the chat I have the the chat year so I'm sorry but it in general how keep track of the chat so you can always ask questions three were asked to. We need to have knowledge about law I said are to be a good in the class. The answers no no not at all so you do not to have any knowledge of it, you need to open to learning about it. So if you have some interest in social science applications or legal applications of help it will make the class much more enjoyable. but there will be no substantive aspects of health or social science that will be tested. and so given that we will have this important online component to the class I Want to make the most of the hybrid learning. The lectures will be recorded by but in some view that contacted me about different special cases which is fine but if you can it's going to make the class more fun and and more better for everyone if everybody comes in if you're going to be absent let me or the tea now and if you have questions or comments from online you can just type them in the chat as you as the doing or you can use the the raise and front in which a no monitor so help asks are people who either know Pytha nor beeligible for this course so aim going to get to that in a little bit. But the short answer is if you've never seen Python I do not recommend taking this course, it will be too difficult. I mean you can try to stay for the first two weeks and see if it's manageable for you. but in previous versions of the class, people who were new to Python and people who had never done any text analysis it was frustrating for them. and so I do not recommend the course for anyone who's sure asked and well tell you that as some emails if you'regoing to be absent for a lecture, a email email after the tea to let her know and if you have to miss a significant number of courses the email of cause you might have to do an additional assignment. So ya so relax. If you're anyone who is kind of the new to Python or has not done any ex data and turnout sure let me know you can talk about it so avoid asks and can homework only be permitted in Python or can we draw also try to this in or sure yeah you're welcome to try it in our for me I should. I wouldbe great if if anyone is interested in converting the course materials to war that would actually be a great course project so we can arrange to get extra credit for that report. asks what's the do registration deadline? There is not an official one I'm not sure exactly. I think it's varies a bit by department but I do not have an official de registration that line. If you're going to be just for for our grading purposes, it's better to do it before the first response essay which is like six weeks in five or six weeks in because others take a long time for grading and so I would rather you deregister before that. So I would say I think by five or six weeks you should know whether you will stay or not. So smart. Asks if we attend the lecturers and only watch the videos. there will be additional projects yes, so mandatory. The live attendance is mandatory and so if you're going to just watch the videos then you have to do another assignment. but I have not decided what that is yet. Okay so yes, so this is related to newly keep track of course participation through in class activities. So young asks, do you recommend someone who also general machine her knowledge but just to experience with help. If you're actually pretty comfortable machine learning with Python then this course actually wopolity Fine. So if I think that if you're doing the first two assignments, the first two home work assignments and they're not taking you a long time to do if you can finish them within of hours then then your on track. but it mainly do not recommend it. I mean it, if you're quite new to Python then I do not recommend it if you have some machine learning experience than that's good. But as I said some text analysis or snap experiences is recommended. So we have course syllabus I've already sent by email and I be in oppose to league to it again so also asks why this course worked for people who intend buillier of it judge course at so if you took my course in the spring and you l in off if you've done if you finish the assignments of the course in the spring then this course will find freedom so there's there's a little bit of overall. So I would say that he saw in the fall course it was say in the fall course ably report judge it would be fine as a prerequisite for this course. If you've done that then this should be fine for you. So those links are a bit is going to change to it screenshare to the syllabus so I just pose I did a link to this in the home Here's my contact details. Here's area's contact details: the lecture schedule that sessions which I'll cook you a bit in a second but those are at ten a man on Fridays they're not mandatory but these will. They also be recorded and Afro will go over the example coding notebook for the week and then also go over the previous week's homework. This is our daughter's the structure, All the slides including Iardaploi today slides here there in the slides thotfolder notebooks. These are the simple notebooks for learning the material and so before before doing the assignment. You should read the notebooks so you can see. You can kind of skim though you can read through these, ensure you understand them and everything that's in the homework which is here under homework. The homeowners will follow will have similar content to what's in the notebook so you can see we fill in part of the notebook and then you have to add something in a new column text which contains the lower case, title and lead. So here's lead. here's title and so No Nights is an example and here you can just like to nowtype lower so thiswill be how to get lower case. So this is just to show you that these the notebooks and the homework are designed off for self motivated learning of all the coding as aspects of the course. so find asked how do the homers omissions work? So there's there's a homework every week and so it's like this: homework Here you download this Jupiter notebook and fill it out and finish it and then you upload it to add you that it's not working yet but it's going to be on the course model. There's a submission system or using coal ufous up load it on the website and going to be due. The homework are done on thousands but the first homework is done next. Thursday So I can not actually show you if you scroll down so everything you need is going to be highlighted here. So for example, do this week. next week the homework one is done on Thursday fin is that what you were asking? I'm going to come back to that as let me know if you have some other questions about it. So here's me. I'm going to put in the correct system still working on this but camera acts are all homework mandatory if you want it mean you lose point if you do not do them but they are. The homework are a completion grade so you know we're not grading them. We're not grading all the answers but if will check that you did like you tried to do every piece. and if you say you get full credit and basically so in terms of the grading it's thirty percent for the programming homeowners and I do not go eleven homework so mistake. three points per homework or that's thirty percent and so for every lecture will have the slides. I'm going to post the links to the recordings here so like after it today, you'll be able to get the recording for for everyone here. there's going to be a tea session on free, there will be a recording link heiress about. So unique asks what can we think, what the response essays are. Can you be a bit more specific? like do you want to see an example? Okay, well get to that. We'll get to that next week. It may be the We attributes. You do not have to do one of those for a time until a month from now. time is talking about the response essays. Whether's some information here I'll provide some example response essays from previous previous years, but it was not going to get into that into detail today because it's a more complex topic but you can read about them here. But basically what it is is reading a paper one of these papers in writing a response as I about it. Like a review here, I have a link to the bibliography of references. So these are all of like the materials that the slides are based on so you do not. Someone of these are required readings but it's worth skimming through this just to see where things come from. and if you are interested and you want to to go back to add to fill in your knowledge from the slides then you can read these the other. the other required assignment is there is going to be there required readings for example in week for you have to read one of these papers and then we'll do an inner class activity about them but it's going to be. We will form groups and do short presentations to each other about these press but I'm going to provide more information to that in week there before we do that. So the the three pieces of the assessment are the homework on the coding homework which I showed you the response essays which I mentioned or reading a paper and writing a report about it and in third there's a end of there's an end of the course assignment and its in the you So we would call them an exam but I think here you would just say it's an end of course assignment where you have a few days to do an assign. For those of you who were in my class in the fall you know this is like it's a questionbasicly a question about every like sure in some questions about the required readings and so that the end of the course assignment is one the things that we will cover in the lecture are covered their of sotthat's how the course will be assessed I'm going to cover some of that information again now just in the slides so it mentioned awards the is the first that session will be on fairly area's here as well. After do will introduce yourself sure here one man after I'm a packed student at an centre and I hope to see you in the first session. So in these is sessions it's what would expect far will go over the notebooks they code note books from the bathtub and then the last week's homework after you've submitted it and then usually there will be some time left over an area can turn the recorder off and you can ask some office hours time questions. I'm going to pose course announcements on Model and if you were registered for the course this morning you should have gotten an email about that if you did not send me a note after class and so we can try to figure out your muddle registration. it's not ready yet but I'm going to work with airfare to post it but we will have this to in a forum on model and so you can post questions there before next week's class and I'll go over them at the beginning of class or I'll just answer on the model. So I wanted to make a note about the course work load because this is not like other science and Perspectives classes like it's not much work to the extent that I've gotten negative. I mean I just want to say expectations I have got a negative feedback on the class because people thought it was too much work and so the thing is, it's actually a lot of work for me too because I have degraded the response essays. So it's actually easier if there's fewer students. So if if you're worried about the course load, then there's other classes you can take that do not take as much time, but according to it, would increase. The number of credit points at I is not the credit system is the maximum for a Science and Perspectives course, but the number of hours for most people If you satisfied the course prerequisites such as having some Phantom background and a little bit of blip background. the amount of work in this course is less than ninety hours. And so it's twelve lectures, eleven programming assignments. There required readings to response essays, and then the final assignment. So that's about sixty hours of time just actually doing these things. And so that includes three more hours. So that includes the tea sessions and then study time if you are new to pythem especially, but if you're new to help then it will take longer. So I just wanted I Want to say expectations about that beforehand? Also, if you were interested in this topic of applied Up for Social science then I would highly recommend you also sign up for the two additional credits for the course project so we'll talk about that more after class next week. So if your interested in it, just stay after it you. This is simply recommended for people who might want to do graduate research after because the previous course projects have turned into conference and journal publications. two of the projects were part of into Swiss startups as well. So if your interested in legal tracker or other entrepreneurial projects based on applied help then the course project could be interesting for you so then asked one where doing for the submission of the project. there's there's a link from the syllabus on course projects that has the rough deadlines. Basically you need you haveyouhave to pick a topic within the next month and then have an outline within the next month and then the full draft of the paper is then day until remember September first so you can work on it over are so a related system of what we've talked about already. Thank you to everybody who filled out the course survey. if you registered since I said this out, send me a note, email me a no because it send you a link to the course survey. Oab'll just send out another link so you can fill it out as well if be curious who else has joined. It's about half master students and few old students and then the rest bachelor students and mostly computer science some data science. He's actually zero people from law which is somebody asked do we need substantive law background so if we did not, we would lose all these students. So we do not require that so that two you guys are So I Already went through this a bit in the syllabus, but the required readings are indicated in the syllabus schedule. In addition, there's the bibliography of references that has additional readings if you want to complement the slides in the link related to the response essays. there's the list of applications papers for response essays, which will talk about more next to be. So I wanted to just give a quick outline of some of the books that were mentioned in the references list. Again, none of these are required readings, but for those who want a deeper understanding, I do coming these books. So Natural Language Processing with Perception is the book that accompanies the Natural Language Tooloquate, which is just this classic blip trouble with kind of more standard classical like old school machine, old school natural language wing tools. If you want to learn machine learning, this is my favorite book for earmachine learning with Physicist learn and wood courses and Monster Flow. It's more generic and's not about Inop specifically, but there are a lot of top applications and for those of you in my course in the Fall you would have already seen this and this is available on oil through the Earth Library. you should be able to get it. This is a free book if you're interested in more of the theory and guess for natural language processing more product than mathematical formalization. If you want to do graduate work research in Blip, then I really recommend this book the Your Goldberg book. I think this is available for download as well on the Earth Libraries. If you can not find it, let me know I can get you a Pdf even though it came out in to this and seventeen. It's actually a little bit out of date unfortunately, so it basically has everything up until Transformers, which as we'll talk about have kind of remade inilp. but a lot of the issues and approaches here are still quite good. Another kind of classic textbook is necessary in Martin. Its kind of more than the does not really focus on neural nets inalp and is kind of more than the older generation of help research, but it's very good for some of the more linguistics oriented in semantics oriented part of the course, so this came up a bit already. Python is a course prerequisite see here for the example notebooks and you know I'm sure many of you as I said, Python is a country register, so you should already have it set up on fairy affairs can provide some help in setting things up, but we would trust everybody. Everybody should be able to get their own another environment running. As a prerequisite to this course. these are the main piping packages that we will be using. As I mentioned in all to is this broad collection of older Inalp tools. Finish is great for topic models and award embedding. Spicy is another kind of generic tool. It's great for named in any recognition, parsing in reference resolution, things like this as well as a library of pre trade world factors. and then as it mentioned this new inilp this new neural net architecture called Transformers in particular large pre train transformer models that are trained on large corporate these have really remade how help is done and hugging base transformers as the hugging base system is the standard for that. To provide an overview on the course, here are your objectives: seventy percent if you want to learn how to use help tools and fifty their parents if you want to learn how to apply opinion tools for law and social science so this is great. We're going to do both in this course which are the followings best Matches your goals for learning in top: sixty percent want to learn it for Engineering in Software development Thirty seven percent for social science research and fifty three percent for computer science research. This is good. We're going to be doing all three of these goals are going to be covered in this course so avoid asks if we need to into processor to no no and maybe you're asking if you know like at you for the assignments. The short answer is no and you do not need any special computer equipment the yeah so we you should be able to the examples on the books and assignments. We use kind of small corporate things so you do not need you do not need any specialized competition for that. If you have problems you can use a Google collaps right and so Afro will cover that in the tea. sure you can just those Google could for everything. So why this course right now we live in this really amazing time I Think for language processing where with our lifetimes there's been these new social structures that have remade the data landscape. the Internet Social Media digit join efforts by governments and Google Books for example just as amazing amazing initiatives for digitizing tons of text at the same time as having these huge crops are. We also have had this amazing increase in computational resources as from cheap disease to efficient databases, solutions for quarrying all of those corporate and then having cups give rise to go Pus and then also tips for training these gigantic volunteers and in particular for natural language analysis. We have these really interesting tools in machine learning, a blip and casual inference for the legal and the social science applications of these tools. And for me I Think it's fair to say that at least from a research perspective a lot of as these trends are especially amplified in the law and Legal language. Political Science and Political Language Here many doors that are being opened in these old fields by these new technologies and so we care about legal and political institutions such as what judges write in their opinions, what politicians say speeches, what's written in patents or in newspaper articles about the law or in legislation and regulations. Those are all millions of lines or millions of documents of unstructured texts. and there's no way that humans could read them even if they wanted to add. So this is why bring in these tools for computers to read them is so exciting. So manual asks, could you share the response to the questionable students background acknowledging presence is up. I do not have that in the slides but if all talk about that next week. I don' think there's anything not notable from that. or do have a specific question manual right? All talk about that a bit next to be but but you do not need to worry about that. So here's an outline of the course and actually I would say let's will all just go through this and they will take them break. So I know this is like an eyefool and I made this last year, but we're actually going to follow basically the same format and justice. but you can visualize everything that we're going to learn in this chorus from this gap. And you what? what we're starting with as raw data today. and next week we go from raw data to segmented documents that are pre processed in different ways. And once you have these documents, you can only use these in some social science analysis just to say oh well, how long are the documents you know? How many bonus do they use? What's the word link to the sentence Link This public a measure of reliability or sophistication in language. The second would be dictionary accounts. and I Think if you're a example researcher, a computer scientist, the fact that you should just go and count different words and count the number of the times the word good shows up as a measure of sentiment that just seems so primitive. it's like the stoneage it. But it I think that we should consider those models cape seriously and I'll give you a good reason at the end of today why dictionary methods are are not to be ignored. And so next week we'll get into tocanization. So the different ways that documents are split up in to sendances and words and looking at part of speech things like this. Once we have documents as these end up being teprimitives for all the models that we will analyse including in gram. So that's converting tokens to phrases. Topic models that's converting documents to distributions over topics so you can imagine in illegal groups there's a crime in contracts on tutors and patterns and things. Each stations are left wing politician I Think my internet might be unstable but I'll give it a second. Can you go hear me now? Can you put it in the cash? I back market thank you So think asks do do we take a look at how he's methods roughly work or do we may learn to use them or what were weregoing to do both rateboth so we will in the notebooks in homework. In the tax sessions we're going to be learning how to do these things in Python we will implement them, but in the lectures were going to be focusing on whether's the purpose, how whatever, going to help us understand the judges or the lawyers things and's so after machine learning will get into neural nets and a particular if we'll learn about word embendings which is a way to represent language in a low dimensional space we'll get into passing, getting at syntax, the relationship between subjects and objects, agents and patients. This is getting into linguistic sides things. We will then get into Transformers and that part of the course which will lead to are ample language modeling knowledge graph's entitlement. So this is getting into asking a computer does a empty public, does sentence A empty B or unknown We do information and extra going to extract relations from a corpus and learn about what the corpus is about and towards the end will be getting into these of more global semantic aspects of summarization. Question answer, automatic claim checking, casual inference from documents, identifying casual relations in documents and a lot of this gets way past that social scientists are using right now. But I think these technologies are improving rapidly in the case of the legal domain at least the there going to be clear applications that really have not been done yet but or will be running right to the frontier in this course to take a look at this if you like will be following us as long as we go in the course. Okay so I know we've already gotten abunchab logistical questions but I wanted to take break now and give everybody a chance to ask a question and then we'll take a quick coffee break. So are there any questions at the moment that have not been covered? You can put them in a chat or the race hand function so manual ask how relevant is reinforcement learning to snap. That's a very interesting question You next's not universal, but it is certainly relevant. There are many very interesting reinforcement learning applications help for example the recent paper that cannot used as reinforcement learning to to improve the quality of summaries and I have a paper with a old student which actually came out of this course using reinforcement learning for attractive summarization. So if you understand reinforcement learning, there's a lot of interesting applications and I can provide some more resources on that. So so fun. Also, memory can note of that area. can you make it note of that? They set up environment script that must have been something from last year so we'll fix that thank you find so report access. Is it possible to do the course this semester near the party next year? Sure it the mean things change right but I'm planning to teach the course again next year so should be fine. Thank you for asking that. So think asks on right, think is asking. theatre's not a simple yes or no. answered to that question Sometimes yes he mean it depends on it mean we're not. We're not going to be like trying to implement them and see Plus Pals or something so but you knew will be talking about you know will have some review of Nstarcastic Radiant Descent. You know how volunteers learn. You know it is not as it said it and this is not a substitute for machine learning in our Pop course so we will do some. but if you need that then you had take a different course or take both of right? So we're going to take a breaknow and will return in nine minutes at there fifteen. I'm also going to turn the recorder off now So if you want to ask a question the recorder please do of really's We really started the content of the course for the remainder of today so nfuzum in on this on the course outline We get to the Corpera. so text data is a sequence of characters called documents and the set of documents is the corpus which we will call them in this course. And what makes text data challenging but also interesting is that it's structured. It's just this stream of characters and the information that we want is induced in the symbolic information in the those characters and it's mixed up with a lot of information that we do not need for any and task and so a lot of what we're going to do in this course is its information extraction or if you prefer information disposal. we're trying to get rid of the information we do not need and will be the main focus of what will do next week and a week after. And to recap something that was mentioned earlier, this course is appealing in place and we do not care about that much of the documents by themselves. What we care about is related into better data and that could even just be like the time Theatre document is published. So we might say well, how syntimate towards a impose social group, How a sintimate towards immigrants changed over time And we can. we can. make a systematic measure toward immigrants and end show that How that's evolved over the last ten years less one hundred years. And the important point there is that we have met a data on the time and so just to say that it be more specifically new might start off with some positive negative syntimate capital by and judicial opinions. And that by itself is not that interesting for a lawyer or for such a scientist. But what if we had the dynamite in opinion is by Judge J at time It and so we will often use this type of notation with these subscripts for the indexing corresponding to the meta data that the document corresponds to. And we can say you know how to sintimate very over time. Or let's say we have the information on the political party of the judge are they do. They have more negative sentiment towards defendants from groups Go. So let's say that go is a dummy variable enjoying one for cases where the defendant is an immigrant and so then we can is information in our data set to say. Well, the right wing judges. They have more negative sentiment towards immigrants for example and you can see that one you relate the text features to meet data. There's many interesting questions that one can answer and a precursor question to you. This type of analysis is what's the document. So now often have this big data set. If we might have a bunch of books on what do we do with them, we need to zoom input to some degree to find out which document, which observation is useful for our meditative variation. And so you know if if we do not want to just arbitrarily make the documents really small because they do not, they do not help us answer a research question such as you know our republican judge is more negative towards of immigrants. If we made a data seat at the sentence level for that, the sentence would both abstinence. data would be super high dimensional, but we would not have any additional variation comparing the right wing judges. did the left win judges. And so this is going to be one of our first Islands activities. what should we use as a document in these contexts? So I Want to take about five minutes six minute to answer these questions? We're going to set up them, going to set up breakout groups to put you guys into pairs and this is just a way too please pull up the slides on your own computer because you're going to lose it when we stop sharing and I'm going to put everybody into pairs and this is just a way for us to start to know each other. So you going to be. You're in a breakout room of two or three people. introduce yourself and said that your major and what you are interested in the chorus and answer these three questions together. What counters the document and were is going to open the breakout rooms for six minutes and will be back at Fifteen Twenty seven so only so handle in corporate. So we have some examples in the notebook about some breed processing to especially if our working on a project not just in this course but you knlater on in your career in life. It's good to think about from given questions or given to ask the data sets out there that have been compiled and so far for court cases like in the U is for example court listeners good but in social media there is really excellent data from Twitter and Facebook for example. for Red it is also for Wikipedia All these data that's are really useful for such a science analysis. This is not in part of the course necessarily, but you know it will be useful for you later on to learn about queueing ships running web scalpers doing processing to remove home markup and there is also the hugging face hub and hugging face system. They have a lot of amazing data sets so it's good to just kind of produce through that ski that a bit can fu access should learn or should have learnt so it would say it mean that should learn because it do not be tested in this course but it will help you to know it for you for other things. So I recommend learning it but you can do it after that. We do it in the summer so all everything that we talk about this course is kind of language agnostic I'm a native English speaker so everything is going to be focused on English. but for your course projects and things you're welcome to do things in other languages. After one of area's special teas is multilingual implants so she can help you get started with it. And there's also just really great machine translation So Elevthos asks why should we use Cyllinium is not cycle evercovtalate that it depends on what what you trying to do then I think that doing a webscraper with that you are a Lib is like a nice start. but if you keep going on up for social science for data science then we will come to a point where you will need a crime driver or something on those lines. So how can we use the quantity of text as data So you know most of this course is going to be about the semantic or conceptual or symbolic content of language. but there is also interesting things to be learned just from the service features just the statistics in the characters in documents and one thing that me and old students in my group had looked at is how the style of writing changes over the life cycle for judges. and one of odd or curious thing we found is that older judges that use shorter words but longer sentences and so whether this is like better or worse writing thing is kind of subjective but it shows that there is. It seems like there is this of biological component of writing style for judges. Another relevant debate in the case of legal quantity of text, the law is on legal complexity where this depends on what country you're from but like in the U's and Finance for example they're always complaining about the law being too complex on but using help we can actually try turn this into an empirical question and ask how complex is the law and certain bedroom which is one of the applications. Papers linked to the syllabus is about measuring the complexity across different areas of the legal code and they produce the number of words in a code title which is a section of the code and they also produce a word invite measure which is basically it's the tiny of the probability distribution across words and what they show is that Public health and the Tax code for example is very complex. It has a a lot of words in it but if you look at the codes if you look at the code titles that have high word intern so a more diverse vocabulary it's quite different and so you. The Commerce in Trade or Public Health and Welfare scores highly on both measures of complexity and so from a tax lawyer perspective. This is kind of interesting because it shows that even though the Internal Revenue Code is complex in terms the number of words, it's very repetitive so it might not be as complex as if sounds. So the next set of methods that will talk about today our dictionary based method which is counting words or patterns and so in the notebook we show some examples of using regular expressions which is this really powerful string matching system and this is going to be depending on what your research, question or engineering task is how you would want to to do this. So one theme or question in the law or judicial behaviour is do judges? Are they making decisions based on justice or morality or are they doing it based on cost and benefit in analysis or efficiency? And so you can imagine using a dictionary method to go through to what different judges say and say, do they say justice demands or do they say efficiency demands And so a dictionary based method could help you get it that. There are also general dictionaries such as Wardnet and Luck which will talk to you about in a second. One example from Economics and this is also one of the applications readings is Baker Bloom and Divas where they wanted to know what is the amount of uncertainty in the macro economy and this is like if you're a Mcroeconomister, you're going to finance a big beam of when there's more uncertainty in the markets, they're more volatile. That could be costly. And so they built this data set of newspapers and they use this kind of simple dictionary method. Does the article that the word uncertain does it have a word related to the economy and then does it have some words related to the government and then they ploddedin that and this is actually just the main result in the paper is just showing that this policy of uncertainty index based on the newspapers, it spikes during these kind of stock market shocks. So like Black Mundane or the Wars elections, nine even. This is the financial crisis to two thousands and nine the Euro Death Sealing crisis. And so you can see that these kind of intuitive moments of market uncertainty are captured by this newspaper based measure. There are are some problems with that to be sure if you're curious about using this in like financial applications is recommend to keep that all paper. another set more fundamental dictionary methods are are available in World That which is this nice python package with a database of it's a big dictionary with all the words in English and how it related to each other and so you can see for example, all all the different senses of the word bass or bass are located here. so it's like music or voices or the fish. There's many meanings of the word base or bass and the the word that it captures these synonym sets of groups of beer anonymous and has information on the anthiponyms and also parts and holes and then also all towns are organized in this categorical hierarchy and so you could have employers which are the higher word in a category and then symptoms are the members of a category and so you. There's many ways this could be used, so if you want to do definition reduction, you could replace words by their hypernem for example. And if you keep on going up Twilight of Categories word that has twenty six lexacer, gray graphic suppresses and so like action, animal partifact person relations shape things like this. I think it's pretty cool that these linguists have gone together and really tried to organize all of language in this and it's all automated now and a Python and so they did the same thing for verbs as well. So this is for bonus and this is for verbs and so you could take a corpus in, say, well, which books are mostly about perception versus possession for example. and you can imagine there's a kind of digital humanities or cultural analytical types applications for these methods. Some other general dictionaries which are relevant for us include lists of function words so thinking of words like four rather than also these get at the style features of text so most of them have you going to drop them because they're not going be very useful for topics, but they can be used to get it nonpolitical dimensions. So if you want to identify authors of texts without that being correlated with the topics, then using software is a good way to do like. or luck is it's kind of the standard licenses for linguistics and Social Psychology and the like team. They have seventy lists of word's of category relevant words including the commotion, cognition, family, positive, negative. We will see some applications using Luck later on in the course and in more recently there is these big lists of words that are initiated by people on crowd scouring platforms. So Mohammed and turns have joy and sadness, anger, fear, trust, disgust, anticipation, surprise and can warmer at all. They could fourteen thousand words along violence, arousal to dominance dimension. So these last two on kind of emotional content. Those are part of this broader set of tools in blip on sentiment analysis and this. You'll see this in many papers, but also in an industry like in advertising digital marketing, they really care about determine right for obvious reasons and we want to know for a given like a review of a movie. Is it positive, negative or neutral And the standard approach could be licensed. base research for the word good, but it's easy to break that right. Like what if somebody said though the movie was not good or the movie was not very good and so just like amends other words totally transforms the meaning. and so it means that just counting words often is not going to work very well. And the more moderate approach is to use pre trained transformer based syntimate models in area I Think is going to add an example of this into this week's notebook to show how you can download a pre trained sentiment analysis from the Hugging Faced Model hub and apply that to documents and you should be careful about these off the shelf scores through these are trained models because they might be trained on a corpus that's quite different from yours. So if you take a corpus that's like initiated tweets and then you apply it key contracts, that problem will work right and so you have to be careful about that. Check that it works in your new domain and there is also some methods which will get too in the world embeddings for making some domain specific licences so you can start off with some positive negative words and then you might find that in read it. the positive words are different than the ones in Instagram for example. So I wanted to just point out a specific issue with syntimate awareness that are based on initiated data. So if you take a syntimate analysis like from hugging bass or something and this is from from some slides by crime car I saw on his Twitter feed. so I do not know exactly which model this is but he made this where if you just take let's go get initial food versus let's go get medical food this has a very positive sentiment and this has a low sentiment. This is bit so just changing the word from relation to Mexican but just that soon as by itself the sentiments exactly the same right? but they changed it from relation to medical and the sentiment went from positive to almost negative and then this is an even kind of more striking example. If you say my name is email, you get this positive sentiment. but if you say my name is unique while you get negative sentiment and so this is like really kind of striking and important and concerning issue in all Pi systems and you want to ask the mean, Is this sentimental model racist? What's what's going on here? How did that happen? And this shows how you have to be careful about using symptomatic scores that are trained on initiated data sets because they're also going to be learning this correlated information in the data set. so that's not unsintimate, but then it will learn it and apply it when you use it in a new domain. And so this is just part of this broader problem or challenge or issue in an Apple for social science. But also this is going to be relevant to many things for products as well that we care about. some true sentiment in language. It but what we get is a systematic scorer. and the model that we trained to learn a sentimate score from language has all these confounding factors. So you nonwhite examples for medical food versus Italian food. You can kind of see how that might have happened right where initial restaurants. maybe they tend to be more of scale or like thecritics like them more. and so because the language model or the competent classifies trained on these biased data sets. that food this, let's say, food critics like Italian food, then that gets baked into the intimate model even though it has nothing to do with even though it's using these words that are syntimate neutral. So this is not easy to fix me. You know, because there's not going to be, you can not just make data set that's neutral, like every data set going to have some biases in it. And so I Think that trying to fix it is this really important and interesting area of upon research that's happening and and this is not unique to determine either. This is a universal problem that I want you guys to be thinking about Throughout this whole chorus is that we are trying to measure this true quantity in language, trying to extract it to learn about science or to learn about to solve a legal task to make a prediction and whizzl we care about. But we get this. We can only make a measurement of its indirect measurement and it's going to be confounded by other language features and sometimes non language features like where it comes from or the large age or style and supervised models are just by construction how they're built. they learn features that are correlated with the label being initiated and unsupervised models you in a topic model or world embodying there also going to learn those correlations and so you like. A classic example is like the similarity between the word banana and the word yellow is actually pretty low. but the similarity between the word banana and the word green is actually really high. And it's not because bananas are green, but it's because if a banana is yellow you we just never mention it right and so you have to be very. This is just some examples of these these problems or limitations with these language models you have to be careful about when you're interpreting their their outputs. But and this is what I mentioned at the beginning dictionary methods. They do have these other limitations But They very significantly mitigate this problem. And because the researcher is familiar with the context, they know what would the words mean and so they can always regularize out any serious surroundings. And so if I'm like trying to make a sentiment analysis and my model tells me that captain means high positive, I can easily fix that with a dictionary method, right? And so this is like a defense for dictionary methods. potentially. And I think it's why economists in particular and just other empirical social scientists. they might still use dictionary methods because of this reason. And I mean they have these other limitations which you you can not. Those are difficult to deal with. but we have to be careful with these blackbox models. Okay, so let's wrap up so the first assignment is already put on the gatehub as well as the coding example that Afro will go over on. Friday. So this came up. We explained it a little bit earlier actually, so I'm just going I'm going to stop for a second and answer elithereosthis question sorry is missed that earlier Those elyptherias asked are those general dictionaries useful any more since we can not easily measure the similarly between words. and also such dictionaries require a lot of human labor to be kept up to date. So I think that's an important point. So I mean the general dictionaries. They are built up over time by these research groups. but I think they have limitations and I mean I think they should be updated over time but meant I think of now all the methods that we'll talk about in this class in terms of dictionaries and classified things. They're all going to have kind of prose and cons and we's want to identify those problems and think of ways to address them. So the way that time the timeline will be in this course about the coding practice and the homework is that for last week to so like week one, the notebook, the coding notebook for week it is got to be reviewed in the tea session on fairy week it and the home work for a week it is going to be done on Thursday week to plus one. So for example, week one notebook is going to be reviewed this fairly homework one is going to be done next Thursday arch third and then reviewed in the week to to session. the notebook two will be reviewed in the week to that session and so on. All of this is in the syllabus in terms of the days, so you now something is confusing. We'll cut you guy some slack in the first few weeks of course, but I think it'll be self explanatory as we move forward. That's the first week we have five minutes left for any questions. | 1 | 5.3% |
| medical_ data_ science Ezurich Lecture Machine Learning for Health Care" (261-5120-00L) Basics of Natural Language Processing Gunnar Ratsch, Rita Kuznetsova Biomedical Informatics group, Institute for Machine Learning; Department of Computer Science DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Outline for today Introductionlmotivation Basic preprocessing steps Basic text features LDA algorithm Embeddings: from BoW to word embedding models POS tagging Language Modelling DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Towards Comprehensive Patient Models for Precision Medicine Distill Data Embed Data Connect Data Clinical Trial Design Precision Medicine Health Records Pathology Images Genomic Data X=w( p(phenotype x, t) Drugs Predict Treatment Mobile Health = argmax p(survival X, DINFK Gunnar Ratsch 15. 3. 2022 3 medical_ data_ science Ezurich Usefulness of the clinical texts Classification of the clinical notes Binary: mortality Multiclass: phenotyping Sentiment analysis of the clinical reports is diagnosis confirmed, suspected or excluded Topic modelling diseases, treatment etc Medical Named Entity Recognition diseases, drugs etc Text generation medical report generation (for example based on images) Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 Latent Representation DINFK medical_ data_ science Ezurich Problems with clinical texts Ungrammatical, has misspellings and concatenations. Contains short telegraphic phrases, acronyms, abbreviations, which are often overloaded It can contain many things that can be typed or pasted, such as long sets of lab values or vital signs Institution-specific template-use is common Pervasive fear; misunderstanding, and confusion around security, privacy, de-identification, and anonymization => significant barriers to get access Some sections might be long and detailed, other sections might be empty or only contain some captions DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 5 medical_ data_ science Ezurich Basic Text Processing Most NLP tasks needs to do text preprocessing and normalization: a Segmenting tokenizing words b_ Normalization C. Stop-words removal d. Punctuation removal e Lemmatization Istemming Disclaimer: Natural Language Processing is a huge topic on its own and we can only cover the absolute basics. Check out lectures on Natural Language processing and understanding offered by colleagues in the department: DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Basic preprocessing steps Tokenization "This is a cat:' 99 "This" "is" "a" "cat" 66 9) Issues: "whatre, Im, isn't" "What are, |am; is not" Ways to do: spacy, nltk, custom tokenizer Normalization The indexed text and the query terms must have the same form: e. g., the US, USA & U. S. A. are the same; could be done with RegExp lowercasing Stop-words removal from nltk. corpus import stopwords ['out' on 'off' 'over' under' 'again' 1 further' then' 'once here there'when'where '.. ] Punctuation removal from string import punctuation DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Stemming vs Lemmatization Stemming reduce terms to their stems in information retrieval: Is crude chopping of affixes automate(s), automatic, automation automat Consult; consultant; consulting, consultative, consultants consult Lemmatization have to find the correct dictionary headword form; use of a vocabulary and morphological analysis of words: Lemma: same stem, part of speech a lemma is the dictionary form of a set of words (headword): cat and cats = same lemma (cat) run; runs, ran, running = same lemma (run) Reduce inflections or variant forms to base form Car; cars, car's, cars car Am, are, is 3 be DINFK NLP Stanford Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Bag of Words (BoW) How to map the term representation into vector space? One way to do this is to represent documents as the bag (multiset) of its words, disregarding grammar and even word order; but keeping multiplicity "No, my brain did not hurt 9} "Perhaps it was more exasperating this way than if it had."66 would have preferred it to hurt me. We assign every word w from vocabulary Wa one-hot vector: Vw [0, 0, 1, 0,., 0] e RII W 0 The document is represented as d = {w_1, w_2, "J w_n}, then we could assign the vector for the document d: Vd = Vw wed vocab size W O1oo, 000) 0 1 "dog 0 DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 medical_ data_ science Ezurich Bag of Words (BoW) & Term Frequency How to map the term representation into vector space? One way to do this is to represent documents as the bag (multiset) of its words, disregarding grammar and even word order; but keeping multiplicity "No, my brain did not hurt 9} "Perhaps it was more exasperating this way than if it had."would have preferred it to hurt me. Term Frequency (TF): raw count of term in document: number of times that term t occurs in document d. {"no" 1, "my": 1, "brain": 1, "did": 1, "not": 1, "hurt": 1} {"Perhaps": 1, "it": 2, 66 was": 1, "more": 1, "exasperating": 1, "this": 1, "way": 1, "if: 1, 66 "had': 1} {": 1, 66 would": 1, "have' 1, "preferred": 1, 66 "it": 1, "to": 1, "hurt": 1, "me": 1} W) Easy to use bag of word representation for vanilla ML models (like SVMs) as inputs (example) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 10 medical_ data_ science Ezurich TF-IDF (term frequency-inverse document frequency) TF suffers from a critical problem: all terms are considered equally important: In fact, certain terms have little or no discriminating power in determining relevance. How to scale down the term weights? (Simple idea: scale down the term weights of terms with high frequency, defined to be the total Inumber of occurrences of a term in the document collection: Document Frequency (IDF): the number of documents in the collection that contain a term t TF-IDF: assigns to term t a weight in document d that is highest when t occurs many times within a small number of documents (thus lending high discriminating power to those documents) lower when the term occurs fewer times in a document; or occurs in many documents (thus offering a less pronounced relevance signal); lowest when the term occurs in virtually all documents DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 11 medical_ data_ science Ezurich TF-IDF (term frequency-inverse document frequency) nt TF (t, d) k nk nt the number of times that term t occurs in document d. Denominator the total number of terms in document d. IDI IDF(t, D) = log Kdi EUR D lt e di } Usage: similarity computation feature vector for the classifier (as a baseline) IDl total number of documents in the collection Kd; e D l t e d } number of documents from the collection, where the term t appears, "t = 0. TF-IDF = TF(t, d)* IDF (t, D) TF-IDF values could be obtained with sklearn feature extraction. text. TfidfVectorizer DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 12 medical_ data_ science Ezurich Latent Dirichlet Allocation (LDA) Topics What is the Topic Modelling? geae 8. 82 genetic 0. 01 Word: element in vocabulary set; W Document: collection of words, d (BoW assumption) life 0. 02 evolve 0. 01 Corpus: collection of documents, D organism 0. 01 Documents Topic proportions and assignments Seeking Life's Bare (Genetic) Necessities COl SPix HARBOR; Vnl FORK nalI "Fectak Hou {mcnWthe 72 Lnme"the dunl Saks t lun "Wuak Urt #itctnt ArFr 'ch'Fnentol Cl Wi U cuulelre: Wetd. Ju" IItIun ( #urh tT= t cnr canluku Wa mtd 4 ~llte Ieuh ~ua I Thetulie he c#l Uk nce MTT Ute T" Arcu: Muhun: nen htum7e et > Icuu Fuky' JeN R Mult RuuueeI k a IntruE tkut Iueuntit R theMnln/ Cmfirin: Go J Fkns m Juthe wat Alhw_h ohe nmker: dent 1 J UEuch Mtl. I~ Frehcts Topic: collection of words (elements from vocabulary set) Document is represented by latent mixture of topics: p(wld) p(wlz)p(zld) where z topic. Rearon 8. 82 nerve 0. 01 Gonomo appinp and SoqvancCcld Sprng Hor Ncx Yok Mey 8 lo 12 Stripping down @omeuter ana y vieis @nes = Iatr ofina Minimum Nocain and &n0 EURnt Denonio; data 0. 02 nunber 0. 02 computer 0. 01 C'IENLE Ml : Mi Each topic is a distribution over words Each document is a mixture of corpus-wide topics Each word is drawn from one of those topics Material from Tutorial by David Blei, 2012 For given corpus, we learn: 1_ Topic: from full vocabulary important subsets 2_ Topic proportion: for each document what is about? Vocabulary ~ basis vector extraction (topics) represent d in topic space (topic proportion) (topic proportion could be used as a feature vector for downstream tasks) Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 13 DINFK medical_ data_ science Ezurich Latent Dirichlet Allocation LDA is a generative probabilistic model for collections of discrete data such as text corpora. Proportions Per-word n EUR RV, a EUR RK are the model parameters parameter topic assignment Per-document Observed Assume there are K topics across all the documents: topic proportions word Material from Tutorial by David Blei, 2012 Topic parameter Topics For k in (1, K): choose per-corpus topic distribution Bk e RV ~ Dir(n) For d in (1, D): choose per-document topic proportion 0d EUR RK ~ Dir(a) 04 Bk K n Za;n Wa, n N D For each word w_n: choose topic Zd, n EUR Lx Multinomial(0a) choose word Wd, n eZv Multinomial(Wd, nlzd, m Bk) p(B, 0, 2, wla, n) = K N P(B;In) [ [r(oala) Mp(zd, n| Oa)p(Wd, nl B1. K, _ Zd, n) d=1 a is the parameter of the Dirichlet prior on the per-document topic distributions B is the parameter of the Dirichlet prior on the per-topic word distribution is the topic distribution for document m mn is the topic for the n-th word in document m W is the specific word mn DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 14 medical_ data_ science _ Ezurich From basic text features to word embeddings Goal: to map each word to the vector: BoW and One-hot: Easy to build; Big dimensionality; Do not reflect the relationship of words in the text: Distributional hypothesis: words that are occur in the same contexts have similar meanings: DINFK Harris (1954). Distributional structure Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 15 medical_ data_ science _ Ezurich NeurIPS 2013 Distributed Representations of Words and Phrases and their Compositionality word2vec Tomas Mikolov Google Inc Mountain View mikolov@google com Ilya Sutskever Google Inc. Mountain View ilyasu@google com Kai Chen Google Inc _ Mountain View kai@google com arXiv 2013 Greg Corrado Jeffrey Dean Google Inc. Mountain View jeff@google com Efficient Estimation of Word Representations in Vector Space Tomas Mikolov Google Inc.. Mountain View, CA tmikolov@google com Kai Chen Google Inc. Mountain View, CA kaichen@google com Greg Corrado Jeffrey Dean Google Inc, Mountain View, CA gcorrado@google com Google Inc, Mountain View, CA jeff@google com DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 16 https: / /code. google. com/p/ordZvecl medical_ data_ science _ 1 Ezurich word2vec The basic idea is to predict a missing word from a context: What is context? Example: "the quick red fox jumps over the lazy brown dog' 1_ Continuous bag of words (CBOW) the 'context' is the sum (or mean) of vectors appearing in the sentence. Based on this context we predict the central' word. dog fox brown quick lazy the jumps red the over 2 Skip-gram the 'context' is each word from the surrounding central word. Based on the central word we predict each word from this surrounding: jumps the jumps quick jumps DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 17 dog medical_ data_ science _ Ezurich CBOW VS Skip-gram Continuous bag of words (CBOW) Skip-gram INPUT PROJECTION OUTPUT INPUT PROJECTION OUTPUT w(t-2) w(t-2) w(t-1) w(t-1) SUM w(t) w(t) w(t+1) w(t+1) N log P(w;b i-1 wi-k' wi+k) i=1 N k log P(Wi+ilw;) i=1 j=-k j=O w(t+2) w(t+2) Objective: maximise log-probability DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 18 medical_ data_ science _ Elzurich Skip-gram model Back to the optimisationl Maximise this: N k log P(Wi+jlwi_ i=1 j=-k j=O This prime is importantl T W0 UwI exp p(wolwi) W T w=1 exp vW Uw[ For each word, J learn two representations: 1. as the context jumps {the quick red fox over the lazy brown dog} Distributed Representations of Words and Phrases and their Compositionality Mikolov; Sutskever; Chen, Corrado, Dean, NIPS 2013 Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 19 2. as the word itself DINFK medical_ data_ science _ Ezurich Paradigmatic relations This distinction is importantl "the quick red fox jumps over the lazy brown dog" "the quick red fox leaps over the lazy brown dog 'leaps' and 'jumps' are similar because they fit the context: We don't expect to see them occurring together in a sentence, because they have a paradigmatic relationship. (This means they're somehow interchangeable:) The Distributional Hypothesis Magnus Sahlgren, PhD dissertation (2006) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 20 medical_ data_ science _ Ezurich Country and Capital Vectors Projected by PCA China Beijing Russias Japans Moscow Ankara 3 Tokyo Turkey 2 1. 5 0. 5 Polandk Germany France Warsaw Berlin Paris0. 5 Italy Greece Athens Rome Spains Madrid Lisbon1. 5 Portugal 22 2150. 5 0. 5 1. 5 2 figure 2 from Mikolov et al. II Distributed Representations of Words and Phrases and their Compositionality NIPS 2013 Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 21 DINFK medical_ data_ science _ Ezurich So does it work? Tested vectors using an analogical reasoning task A : B as X : Y'? (e. g: 'puppy is to dog as kitten is to cat) This means asking: vec(A) 5 vec(B) 2 vec(X) vec(Y) Or 2 vec(A) 5 vec(B) + vec(Y) vec(X) Where, as mentioned before, the 'similarity' here is cosine similarity. DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 22 medical_ data_ science _ Ezurich So does it work? Created a test set with ~9k semantic questions and ~11k syntactic Examples: calm calmly safe safely Athens Greece Japan big bigger small smaller old oldest best Poland zloty Hungary forint move moving fly flying Austin Texas Honolulu Hawaii Ireland Irish Egypt Egyptian dancing danced saying said girl uncle aunt man men cat cats (these are actually most of the categories) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 23 Tokyo good boy medical_ data_ science _ Ezurich So does it work? Table 4: Comparison of publicly available word vectors on the Semantic-Syntactic Word Relationship test set; and word vectors from our models: Full vocabularies are used Model Vector Training Accuracy [%] Dimensionality words Semantic Syntactic Total Collobert-Weston NNLM 50 660M 9. 3 12. 3 11. 0 Turian NNLM 50 37M 1. 4 2. 6 2. 1 Turian NNLM 200 37M 1. 4 2. 2 1. 8 Mnih NNLM 50 37M 1. 8 9. 1 5. 8 Mnih NNLM 100 37M 3. 3 13. 2 8. 8 Mikolov RNNLM 80 320M 4. 9 18. 4 12. 7 Mikolov RNNLM 640 320M 8. 6 36. 5 24. 6 Huang NNLM 50 990M 13. 3 11. 6 12. 3 neural net Our NNLM 20 6B 12. 9 26. 4 20. 3 Our NNLM 50 6B 27. 9 55. 8 43. 2 language model Our NNLM 100 6B 34. 2 64. 5 50. 8 CBOW 300 783M 15. 5 53. 1 36. 1 Skip-gram 300 783M 50. 0 55. 9 533 yes! this takes 2 weeks DINFK on 180 cores! 2. 5 days on 125 cores Efficient Estimation of Word Representations in Vector Space' Mikolov, Chen; Corrado, Dean, arXiv 2013 medical_ data_ science _ Ezurich Analogical reasoning The oft-cited example of successful analogical reasoning is: vec(king) vec(man) vec(woman) = vec(queen) Intuitively; this means vec(king) vec(man) vec("sense of royalty") Which is a pretty fascinating idea: What if. _ vec(gleevec) vec(leukemia)~ vec("treatment")? Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 25 DINFK medical_ data_ science _ Ezurich Analogical reasoning Then we could do.. vec(melanoma) vec("treatment") = vec("?? 2") This would certainly be useful for a medical Jeopardy. _ It's also the kind of information we want our embeddings to encode 9 for enabling medical language processing: So we ran wordZvec on some MSKCC text. DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 26 medical_ data_ science _ Elzurich An example vec(melanoma) [vec(gleevec) vec (leukemia)] =? 22 Top 10 closest results to melanoma: Top 10 closest results to gleevec: Top 10 closest results to Zeukemia: 3 me Lanoma 340 ~2. 22044604925e-16 gLeevec crel 2148 1. 11022302463e-16 Leukemia 1129 ~2. 22044604925e-16 dfsp 12007 0. 156969851964 dasatinib 4299 0. 0436408836789 itp A3744 0. 216802690111 neurotropic 17723 0. 18892856986 imatinib 2596 0. 0444753136031 myelodysplasia 8348 0. 220480414542 neurotrophic 22261 0. 193823458626 nilotinib 5211 0. 0565038545667 cll 1270 0. 229098946201 SCC 4457 0. 199310695303 anagrelide 6961 0. 0806562061213 aml 2236 0. 232815740704 amelanotic 10967 0. 205920966083 hydroxyurea 3921 0. 0824481117079 cmmol 8659 0. 236774032333 choroidal 9357 0. 208689368781 gleevac 16087 0. 0843472720016 mds 2858 0. 23788044386 fibrosarcoma 8679 0. 223705333644 ruxolitinib 11279 0. 0845686271568 coexisting 16242 0. 241202871701 eccrine Jl13344 0. 22791107347 "ieeeie nexavar 7350 0. 0862700069714 Leukemia/sll 35544 0. 245014293364 fibrohistiocytoma 11045 0. 239171736259| hydrea 6445 0. 100871239337 Igl Bxa10616 0. 246571161984 cancer/ 27891 0. 243011714361 afinitor 10465 0. 10846339025 hypogammaglobulinemia 6544 0. 249632640917 Top 10 closest results to UNM+ G-L and we get:.. CMLIALL ponatinib 14236 0. 42982079986 diascopy 23369 0. 435802215172 #I#th 20379 0. 44739503288 eruption 3763 0. 447999880188 gleevac 16087 0. 448643975007 nexavar 7350 0. 452329959825 hive 18576 0. 455971559975 pustule 11842 0. 455989743044 gleevec Ae2148'0. 458117608185 dabrafenib 10448 0. 459866794241 desatinib 32409 0. 46007721604 typo :) sorafenib (kidneylliver cancer drug) (BRAF-mutated, metastatic) MELANOMAI DINFK CMLIALL medical_ data_ science _ Ezurich Other embedding techniques 1. Glove [1] aggregated global word-word co-occurrence statistics from a corpus. 2. FastText [2] ~ training process with character n-grams, helps with OOV (Out of vocabulary) problem: 1 GloVe: Global Vectors for Word Representation; Pennington, J. Socher; R,, & Manning; C. D."EMNLP 2014. 2. Enriching_Word_Vectors_with_Subword Information; Bojanowski, P, Grave, E., Joulin, A"& Mikolov, T. TACL, 2017. DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 28 medical_ data_ science Ezurich Part of Speech (PoS) Tagging Given a sentence W. Wn and a tagset of lexical categories, find the most likely tag t-tn for each word in the sentence Penn_Treebank PQSTags] SecretariatINNP isIVBZ expected/VBN toITO race/VB tomorrow/NN Problem: Many of the words may have unambiguous tags But enough words are either ambiguous or unknown non-trivial task Brown corpus is a general corpus in corpus linguistics (500 samples of English texts, ~IM words). Most words in English have only one Brown Corpus tag: unambiguous (1 tag) 35, 340 words Many of the most common words are ambiguous (over 40% tokens are ambiguous) Obvious strategies may be suggested based on intuition: to/TO race/VB thelDT racelNN Methods: simple baseline with unigrams hardcoded rules vs supervised unsupervised ML approaches. NNP Proper noun, singular; VBZ Verb, 3rd person singular present; VBN Verb, past article; VB verb, base form, NN Noun; singular or mass: DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 29 medical_ data_ science Ezurich Simplest strategy Choose the most likely tag for each ambiguous word, independent of previous words assign each token the POS category it occurred as most often in the training set This strategy gives 90% accuracy in controlled tests Which POS is more likely in a corpus (1, 273, 000 tokens)? race: NN: 400 VB: 600 P(NNlrace) = P(race, NN) / P(race) by the definition of conditional probability P(race) = 1000/1, 273, 000 =. 0008 P(race, NN) = 400/1, 273, 000 =. 0003 P(race, VB) = 600/1, 273, 000 =. 0004 SO we obtain: P(NNIrace) = P(race, NN)P(race) =. 0003/. 0008 =. 375 P(VBlrace) = P(race, VBYP(race) =. 0004/. 0008 =. 5 DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 30 medical_ data_ science Ezurich HMMs: A Probabilistic Approach We want to find the "best sequence' 9) of PoS tags T=T_ T for & sentence 1' W=W_ 1"W n' where T; is the PoS tag for word W In other words we want to find a PoS tags T that maximizes P(TIW) Using Bayes' Rule, we can say P(WIT) * P(T) P(TIW) P(W) We want to find the value of T which maximizes the right hand side. note: denominator can be discarded (same for every T) Find T which maximizes P(WIT) * P(T) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 31 medical_ data_ science Elzurich Independence Assumptions Assume that current event is based only on previous n-1 events n 1 P(T1, _, Tn) IIP(TIT;_1) i=1 assumes that the event of a PoS tag occurring is independent of the event of any other PoS tag occurring, except for the immediately previous PoS tag from a linguistic standpoint; this seems an unreasonable assumption, due to long-distance dependencies 2_ P(W1.. WnITi. Tn) II P(WAT;) i1 assumes that the event of a word appearing in a category is independent of the event of any surrounding word or tag, except for the tag at this position: Linguists know both these assumptions are incorrect nevertheless, statistical approaches based on these assumptions work pretty well for part-of-speech tagging: Hidden Markov Models (HMMs) is widely used in both PoS-tagging and speech recognition, among other problems DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 32 medical_ data_ science Ezurich PoS Tagging Based on HMM Problem: Find T which maximizes P(W | T) * P(T) Here W=W, 1W n and T=T Tn Using the HMM; we get: Transition probabilities (prob. of transitioning from one stateltag to another): n P(T1, Tn) IIP(T;IT;_1) i=1 Emission probabilities (prob. of emitting a word at a given state): P(W1. WnITi.. Tn) II P(wt;) i] We want to find the value of T_ 1"~T which maximizes: n n II P(T;) * P(TilTi-1) {=1 DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 33 medical_ data_ science Ezurich POS probabilities: transitions P(T1, _ '' 1 Tn) ~ IIP(TIT;_1) 0. 8 i=1 1. 0 PRP MD 04 NN 0. 6 0. 3 "He will race J} Possible tag series for T = T1'T2'T3 T =PRP MD NN T = PRP NN NN T = PRP MD VB T =PRP NN VB POS bigram probabilities from training corpus can be used for P(T) 0. 2 NN 0_ B PRP: personal pronoun MD: modal (auxiliary verb) NN: noun VB Verb, base form Which is the most likely path? DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 34 medical_ data_ science Ezurich Lexical generation probabilities n From the training corpus, we need to find the T; which maximizes: [[ P(WAT;) * P(TilT;_1) C E i] 0. 4 0. 8 MD NN A B 1Q PRP 0. 6 0. 3 0. 2 NN VB 0_ C Emission Probabilities 0. 4 willMD 0. 8 racelNN 0. 4 MD NN VB PRP 0. 8 A B 0. 6 he S | $ 10 helPRP 1. 0 will 0. 8 0. 2 0. 3 0. 22 0. 7 race 0 0. 4 0. 6 willINN 0. 2 racelVB 0. 6 Note 1: sum over column should sum up to 1, the MLE of the emission probability is how many times this word W appears as this tag t, divided by how many times we saw the tag t in training data: DINFK Note 2: the whole table extends further down than is depicted, as there are other words in the dictionary whose emission probability we're not interested in for this sentence. Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 35 medical_ data_ science Ezurich Using Dynamic Programming To find the most likely sequence of categories for a sequence of words, we don't need to enumerate all possible sequences of categories. Due to the Markov assumption, if you keep track of the most likely sequence found so far for each possible ending category; you can ignore all the other less likely sequences: multiple edges coming into a state, but only keep the value of the most likely path, i. e. use of DP The algorithm to do this is called the Viterbi algorithm: DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 36 medical_ data_ science Ezurich Viterbi Algorithm recap (informal) 1_ Assume we are at state h in the HMM: a States H, _H m all come into h 2. Obtain a the best probability of each previous state H_Hm b the transition probabilities: P(hIH;); P(hHm) C the emission probability for word w at h: P(wlh) 3_ Multiply the probabilities for each new path: Best(l) = Maxh <h [Best(H)* P(hlH)]* P(wlh) 4_ One of these states (H_ 1"H m) will give the highest probability Only keep the highest probability when using h for the next state E 0 willlM D racelN N 0. 4 0_ 8 A B 1Q helPR P S | 0 03 0 0_ 2 Find the most likely sequencel DINFK willIN N 82 0. 7 racelV B 0. 6 F medical_ data_ science Ezurich Introduction to the Transformer model Transformer is the encoder-decoder architecture with self-attention: The key component of the Encoder is Self-Attention mechanism. Layer normalization Input data: X e Rnxd Encoder overview: Q-queries, K keys, V values: linear projections of the input data X, d dimension: # QKT Latent Representation softmax )V Residual 1 Va connections Add & Normalize Feed Forward Feed Forward Projection layer / Add & Normalize Self-Attention Attention weights Positional embeddings POscoDMG Self attention Positional Embeddings is used to learn the order in the sequence. Residual connections mitigates vanishing gradients. Layer normalization helps with numerical stability: POSITIONAL ENCODING EMBEDDINGS X1 Pictures from The Ilustrated transformer_Jay Alammar DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 38 Vaswani, Ashish, et al. "Attention is allyou need"Advances in neural information processing systems 30 (2017). medical_ data_ science Ezurich Language Modeling Deep Learning Language Model (LM) computes the probability of p(w_1, w_n) for any W_1, _ w_n = V (vocabulary) How we do LM in modern world? state-of-the art Transformer-based models BERT Input [CLS] my is cute [SEP] he likes play ##ing [SEP] Pretrained WordPiece embeddings Token Embeddings E [CLS] E my E "dog E cute E 'play [SEP] ~likes [SEP] A marker indicating Sentence A or Sentence B Segment Embeddings EA EB To learn about the position in sentence Position Embeddings Eo E Ez E4 E5 E6 E7 Eg E1o 1 [CLS] token is added in the beginning a special classification token, the final hidden state corresponding to this token used as the aggregate sequence representation for classification tasks. 2_ [SEP] token is inserted at the end of each sentence. One sentence also can feed as input. 3_ Both sentences A and B are encoded with the help of Byte-pair encoding (BPE): BPE is a simple form of data compression in which the most common pair of consecutive bytes of data is replaced with a byte that does not occur within that data. ZabdZabac DINFK Devlin et al. 'Bert: Pre-training of_deep bidirectional transformers for language understanding arXiv preprint arXiv:1810. 04805 (2018). Z-aa dog 8 &ne E#ring 4 4 4 4 4 8 8 8 8 S 8 medical_ data_ science Elzurich Training details 1. Masked-LM: replace n% words in the input by special [MASK] token predict these tokens variants altering Structural DNA [MASK] can modify gene function by [MASK] transcript sequences 2. Next sentence prediction: binarized task; from the [CLS] token need to predict whether one sentence follows another in the text. Class 1: Cancer is a group of diseases involving abnormal cell growth It has the potential to invade or spread to other parts of the body: Class 0: Cancer is a group of diseases involving abnormal cell growth A bear is sunbathing: The basis of the BERT architecture = encoder of the Transformer model Way of using BERT: have pretrained model finetune for the required task on the specific corpora DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 40 medical_ data_ science Ezurich BERT in biomedical domain Data: MMC I clinical notes, PubMed abstracts, papers from Semantic Scholar etc Tasks: Named Entity Recognition, Q&A, Sentence Similarity etc Hugging Face Search models, datasets, users__ Models Datasets Spaces Docs Solutions Pricing Tasks Models 296 bio Fill-Mask Question Answering dmis-lab/biobert-base-cased-V1. 1 Updated Oct 14, 2020 1. 01M Summarization Table Question Answering Text Classification Text Generation Text2Text Generation 88 Token Classification emilyalsentzer/Bio_ClinicalBERT Fill-Mask Updated 16 days ag0 533k 14 Translation Zero-Shot Classification 218 Sentence Similarity +14 microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext Fill-Mask Updated Sep 22, 2021 121k 19 Libraries PyTorch TensorFlow JAX 4 24 microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract Fill-Mask Updated Sep" 22, 2021 85. 2k Datasets wikipedia common_voice squad dmis-lab/biobert-V1. 1 Feature Extraction Updated May19, 2021 bookcorpus c4 glue conll2003 73k 0 4 dcep europarl jrc-acquis 828 bionlp/bluebert_pubmed_mimic_uncased_L-12_H-768_A-12 Updated Sep 24, 2021 62,. 6k Gunnar Ratsch & Rita Kuznetsova Languages DINFK 22. 3. 2022 41 medical_ data_ science _ Ezurich Summary & Take Home messages Clinical reports contain relevant information about patient's health state, but are written in challenging, specialized language Standard basic text processing is often applied: tokenization, stemming, lemmatization, Latent Dirichlet Allocation (LDA) is a probabilistic model for text content using topics Word embedding are a recently developed powerful tool in NLP Can be learned unsupervisedly and represent semantic in a vector space Part of Speech Tagging is an import task in NLP that is often solved with HMMs Recent NLP techniques use deep learning and have shown great promise DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 42 | 1 | 5.3% |
| <b><!-- </b>if (window!= top) top.location.href=location.href <b>// --> </b> Sunset Boulevard SUNSET BOULEVARD Charles Brackett Billy Wilder D.M. Marshman, Jr. March 21,1949 SEQUENCE "A" A-l-4 START the picture with the actual street sign: SUNSET BOULEVARD, stencilled on a curbstope. In the gutter lie dead leaves, scraps of paper, burnt matches and cigarette butts. It is early morning. Now the CAMERA leaves the sign and MOVES EAST, the grey asphalt of the street filling the screen. As speed accelerates to around 40 m.p.h., traffic de- marcations, white arrows, speed-limit warnings, man- hole covers, etc., flash by. SUPERIMPOSED on all this are the CREDIT TITLES, in the stencilled style of the street sign. Over the scene we now hear MAN'S VOICE sirens. Police squad cars Yes, this is Sunset hurtle toward the camera, Boulevard, Los Angeles, turn off the road into a California. It's about driveway with squealing five o'clock in the brakes. Dismounted motor- morning. That's the cycle cops stand directing Homicide Squad, com- the cars in. plete with detectives and newspaper men. A-5 PATIO AND POOL OF A murder has been re- MANSION ported from one of those great big houses in the The policemen and news- ten thousand block. paper reporters and You'll read all about photographers have it in the late editions, jumped out of the cars I'm sure. You'll get and are running up to it over your radio, the pool, in which a and see it on tele- body is seen floating. vision -- because an Photographers' bulbs old-time star is in- flash in rapid suc- volved. one of the big- cession. gest. But before you hear it all distorted and blown out of proportion, before those Hollywood columnists get their hands on it, maybe you'd like to hear the facts, the whole truth... A-6 FLASH OF THE BODY MAN'S VOICE Angle up through the If so, you've come to the water from the bottom right party... You see, of the pool, as the the body of a young man body floats face down- was found floating in the ward. It is a well- pool of her mansion, with dressed young man. two shots in his back and one in his stomach. No- body important, really. Just a movie writer with a couple of "B" pictures to his credit. The poor dope. He always wanted a pool Well, in the end he got himself a pool -- SLOW DISSOLVE TO: only the price turned out to be a little high... Let's go back about six A-7 HOLLYWOOD, SEEN FROM months and find the day THE HILLTOP AT IVAR when it all started. & FRANKLIN STREETS It is a crisp sunny I was living in an day. The voice con- apartment house above tinues speaking as Franklin and Ivar. CAMERA PANS toward Things were tough the ALTO NIDO APART- at the moment. I hadn't MENT HOUSE, an ugly worked in a studio for Moorish structure ofsat a long time. So I stucco, about four there grinding stories high. CAMERA out original stories, MOVES TOWARD AN OPEN two a week. Only I WINDOW on the third seemed to have lost floor, where we look my touch. Maybe they in on JOE GILLIS' APART- weren't original MENT. Joe Gillis, bare- enough. Maybe they footed and wearing no- were too original. thing but an old bath- All I know is they robe. is sitting on didn't sell. the bed. In front of him. on a straight chair, is a portable typewriter. Beside him, on the bed, is a dirty ashtray and a scattering of type written and pencil- marked pages. Gillis is typing. with a pencil clenched bet- ween his teeth. A-8 JOE GILLIS' APARTMENT It is a one-room affair with an unmade Murphy bed pulled out of the wall at which Gillis sits typing. There are a couple of worn-out plush chairs and a Spanish-style, wrought-iron standing lamp. Also a small desk littered with books and letters, and a chest of drawers with a portable phonograph and some records on top. On the walls are a couple of repro- ductions of characterless paintings, with laundry bills and snapshots stuck in the frames. Through an archway can he seen a tiny kitchenette, complete with unwashed coffee pot and cup, empty tin cans, orange peels, etc. The effect is dingy and cheerless -- just another furnished apartment. The buzzer SOUNDS. GILLIS Yeah. The buzzer SOUNDS again. Gillis gets up and opens the door. Two men wearing hats stand outside one of them carrying a briefcase. NO. 1 Joseph C. Gillis? GILLIS That's right. The men ease into the room. No. 1 hands Gillis a business card. NO. 1 We've come for the car. GILLIS What car? NO. 2 (Consulting a paper) 1946 Plymouth convertible. Calif- ornia license 97 N 567. NO. 1 Where are the keys? GILLIS Why should I give you the keys? NO. 1 Because the company's played ball with you long enough. Because you're three payments behind. And because we've got a Court order. Come on -- the keys. NO. 2 Or do you want us to jack it up and haul it away? GILLIS Relax, fans. The car isn't here. NO. 1 Is that So? GILLIS I lent it to a friend of mine. He took it up to Palm Springs. NO. 1 Had to get away for his health, I suppose. GILLIS You don't believe me? Look in the garage. NO. 1 Sure we believe you, only now we want you to believe us. That car better be back here by noon tomorrow, or there's going to be fireworks. GILLIS You say the cutest things. The men leave. Gillis GILLIS' VOICE stands pondering beside Well, I needed about two the door for a moment. hundred and ninety dollars Then he walks to the and I needed it real center of the room and, quick, or I'd lose my car. with his back to the It wasn't in Palm Springs CAMERA, slips into a and it wasn't in the pair of gray slacks. garage. I was way ahead There is a metallic of the finance company. noise as some loose change and keys drop from the trouser pockets. As Gillis bends over to pick them up, we see that he has dropped the car keys, identifiable be- cause of a rabbit's foot and a miniature license plate attached to the key-ring. Gillis pockets the keys and as he starts to put on a shirt DISSOLVE TO: A-9 EXTERIOR OF RUDY'S GILLIS' VOICE SHOESHINE PARLOR (DAY) I knew they'd be coming A small shack-like build- around and I wasn't tak- ing, it stands in the ing any chances, so I corner of a public park- kept it a couple of ing lot. Rudy, a blocks away in a parking colored boy, is giving lot behind Rudy's Shoe- a customer a shine. shine Parlor. Rudy never asked any quest- ions. He'd just look at your heels and know the score. PAN BEHIND the shack to GILLIS' CAR, a yellow 1946 Plymouth convertible with the top down. Gillis enters the SHOT. He is wearing a tweed sport jacket, a tan polo shirt, and moooasins. He steps into the car and drives it off. Rudy winks after him. A-10 THE ALLEY NEXT TO SIDNEY'S MEN'S SHOP ON BRONSON AVE. GILLIS' VOICE I had an original story Gillis drives into the kicking around Paranount. alley and parks his car My agent told me it was right behind a delivery dead as a doornail. but truck. PAN AND FOLLOW I knew a big shot over HIM as he gets out, walks there who'd always liked around the corner into me, and the time had Bronson and then toward come to take a little the towering main gate of advantage of it. His Paramount. A few loafers, name was Sheldrake. He studio cops and extras are was a smart producer, lounging there. with a set of ulcers to prove it. DISSOLVE TO: A-11 SHELDRAKE'S OFFICE It is in the style of a Paramount executive's office -- mahogany, leather, and a little chintz. On the walls are some large framed photographs of Paramount stars, with dedications to Mr. Sheldrake. Also a couple of framed critics' awards certificates, and an Oscar on a bookshelf. A shooting schedule chart is thumb-tacked into a large bulletin board. There are piles or scripts, a few pipes and, somewhere in the background, some set models. Start on Sheldrake. He is about 45. Behind his wor- ried face there hides a coated tongue. He is en- gaged in changing the stained rilter cigarette in his Zeus holder. SHELDRAKE All right, Gillis. You've got five minutes. What's your story about? GILLIS It's about a ball player, a rookie shortstop that's batting 347. The poor kid was once mixed up in a hold- up. But he's trying to go straight -- except there's a bunch of gamblers who won't let him. SHELDRAKE So they tell the kid to throw the World Series, or else, huh? GILLIS More or less. Only for the end I've got a gimmick that's real good. A secretary enters, carrying a glass or milk. She opens a drawer and takes out a bottle of pills for Sheldrake. SHELDRAKE Got a title? GILLIS Bases Loaded. There's a 4O-page outline. SHELDRAKE (To the secretary) Get the Readers' Department and see what they have on Bases Loaded. The secretary exits. Sheldrake takes a pill and washes it down with some milk. GILLIS They're pretty hot about it over at Twentieth, but I think Zanuck's all wet. Can you see Ty Power as a GILLIS (cont'd) shortstop? You've got the best man for it right here on this lot. Alan Ladd. Good change of pace for Alan Ladd. There's another thing: it's pretty simple to shoot. Lot of outdoor stuff. Bet you could make the whole thing for under a million. And there's a great little part for Bill Demarest. One of the trainers, an oldtime player who got beaned and goes out of his head sometimes. The door opens and Betty Schaefer enters -- a clean- cut, nice looking girl of 21, with a bright, alert manner. Dressed in tweed skirt, Brooks sweater and pearls, and carrying a folder of papers. She puts them on Sheldrake's desk, not noticing Gillis, who stands near the door. BETTY Hello, Mr. Sheldrake. On that Bases Loaded. I covered it with a 2-page synopsis. (She holds it out) But I wouldn't bother. SHELDRAKE What's wrong with it? BETTY It's from hunger. SHELDRAKE Nothing for Ladd? BETTY Just a rehash of something that wasn't very good to begin with. SHELDRAKE I'm sure you'll be glad to meet Mr. Gillis. He wrote it. Betty turns towards Gillis, embarrassed. SHELDRAKE This is Miss Kramer. BETTY Schaefer. Betty Schaefer. And right now I wish I could crawl into a hole and pull it in after me. GILLIS If I could be of any help... BETTY I'm sorry, Mr. Gillis, but I just don't think it's any good. I found it flat and banal. GILLIS Exactly what kind of material do you recommend? James Joyce? Dostoosvsky? SHELDRAKE Name dropper. BETTY I just think pictures should say a little something. GILLIS Oh, you're one of the message kids. Just a story won't do. You'd have turned down Gone With the Wind. SHELDRAKE No, that was me. I said, Who wants to see a Civil War picture? BETTY Perhaps the reason I hated Bases Loaded is that I knew your name. I'd always heard you had some talent. GILLIS That was last year. This year I'm trying to earn a living. BETTY So you take Plot 27-A, make it glossy, make it slick -- SHELDRAKE Carefull Those are dirty words! You sound like a bunch of New York critics. Thank you, Miss Schaefer. BETTY Goodbye, Mr. Gillis. GILLIS Goodbye. Next time I'll write The Naked and the Dead. Betty leaves. SHELDRAKE Well, seems like Zanuck's got himself a baseball picture. GILLIS Mr. Sheldrake, I don't want you to think I thought this was going to win any Academy Award. SHELDRAKE (His mind free-wheeling) Of course, we're always looking for a Betty Hutton. Do you see it as a Betty Hutton? GILLIS Frankly, no. SHELDRAKE (Amusing himself) Now wait a minute. If we made it a girls' softball team, put in a few numbers. Might make a cute musical: It Happened in the Bull Pen -- the story of a Woman. GILLIS You trying to be funny? -- because I'm all out of laughs. I'm over a barrel and I need a job. SHELDRAKE Sure, Gillis. If something should come along - GILLIS Along is no good. I need it now. SHELDRAKE Haven't got a thing. GILLIS Any kind of assignment. Additional Dialogue. SHELDRAKE There's nothing, Gillis. Not even if you were a relative. GILLIS (Hating it) Look, Mr. Sheldrake, could you let me have three hundred bucks yourself, as a personal loan? SHELDRAKE Could I? Gillis, last year some- body talked me into buying a ranch in the valley. So I borrowed money from the bank so I could pay for the ranch. This year I had to mortgage the ranch so I could keep up my life insurance so I could borrow on the insurance so I could pay my income tax. Now if Dewey had been elected - GILLIS Goodbye, Mr. Sheldrake. DISSOLVE TO: A-12 EXT. SCHWAB'S DRUG STORE (EARLY AFTERNOON ACTIVITY) GILLIS' VOICE After that I drove down MOVE IN toward drug store to headquarters. That's and the way a lot of us think about Schwab's Drug Store. DISSOLVE TO: Actors and stock girls and waiters. Kind of a combination office,Kaffee- A-13 INT. SCHWAB'S DRUG STORE Klatsch and waiting room. Waiting, waiting for the The usual Schwabadero gravy train. crowd sits at the fount- ain, gossips at the cigar-stand, loiters by the magazine display. MOVE IN towards the TWO TELEPHONE BOOTHS. In I got myself ten nickels one of them sits Gillis, and started sending out a stack of nickels in a general S.O.S. Couldn't front of him. He's get hold of my agent, doing a lot of talking naturally. So then I into the telephone, called a pal of mine,name hanging up, dropping of Artie Green -- an awful another nickel, dialing, nice guy, an assistant talking again. director. He cquld let me have twenty, but twenty wouldn't do. GILLIS' VOICE (Cont.) Then I talked to a couple of yes men at Twentieth. To me they said no. Finally I located that agent of mine, the big faker. Was he out digging up a job for poor Joe Gillis? Hmph! He was hard at work in Bel Air, making with the golf clubs. Gillis hangs up with a curse, opens the door of the booth, emerges, wiping the sweat from his forehead. He walks toward the exit. He is stopped by the voice of SKOLSKY Hello, Gillis. Gillis looks around. At the fountain sits Skolsky, drinking a cup of coffee. GILLIS Hello, Mr. Skolsky. SKOLSKY Got anything for the column? GILLIS Sure. Just sold an original for a hundred grand. The Life of the Warner Brothers. Starring the Ritz Brothers. Playing opposite the Andrew Sisters. SKOLSKY (With a sour smile) But don't get me wrong -- I love Hollywood. Gillis walks out. DISSOLVE TO: A-14 THE BEL AIR GOLF LINKS On a sun-dappled green edged with tall sycamores, stands Morino, the agent, a caddy and a nondescript opponent in the background. Gillis has evidently stated his problem already. MORINO So you need three hundred dollars? Of course, I could give you three hundred dollars. Only I'm not going to. GILLIS No? MORINO Gillis, get this through your head. I'm not just your agent. It's not the ten per cent. I'm your friend. He sinks his putt and walks toward the next tee, Gillis following him. GILLIS How's that about your being my friend? MORINO Don't you know the finest things in the world have been written on an empty stomach? Once a talent like yours gets into that Mocambo- Romanoff rut, you're through. GILLIS Forget Romanoff's. It's the car I'm talking about. If I lose my car it's like having my legs out off. MORINO Greatest thing that could happen to you. Now you'll have to sit behind that typewriter. Now you'll have to write. GILLIS What do you think I've been doing? I need three hundred dollars. MORINO (Icily) Maybe what you need is another agent. He bends down to tee up his ball. Gillis turns away. DISSOLVE TO: A-15 GILLIS IN HIS OPEN CAR GILLIS' VOICE driving down Sunset As I drove back towards town towards Hollywood. He I took inventory of my pros- drives slowly. His pects. They now added up to mind is working. exactly zero. Apparently I just didn't have what it takes, and the time had come to wrap up the whole Hollywood deal and go home. Maybe if I hocked all my junk there'd be enough for a bus ticket back to Ohio, back to that thirty-five- dollar-a-week job behind the copy desk of the Dayton Evening Post, if it was still open. Back to the smirking delight of the whole office. All Gillis stops his car at right you wise guys. why don't a red light by the main you go out and take a crack at entrance to Bel Air. Hollywood? Maybe you think Suddenly his eyes fall you could -- Oh-oh! on: A-16 ANOTHER CAR It is a dark-green Dodge business coupe, also waiting for the light to change. but headed in the opposite direction. In it are the two finance company men. They spot Gillis in his car and exchange looks. From across the intersection Gillis recognizes them and pulls down the leather sunshade to screen his face. As the light changes. Gillis gives his car the gun and shoots away. The men narrowly avoid hitting another car as they make a U-turn into oncoming traffic and start after him. A-17 THE CHASE to A-21 Very short, very sharp, told in FLASHES. (Use locations on Sunset between Bel Air and Holmby Hills). The men lose Gillis around a bend, catch sight of him and then -- while they are trapped behind a slow- moving truck. he disappears again. A-22 GILLIS He is driving as fast as he dares, keeping an eye out for pursuit in his rear-view mirror. Suddenly his right front tire blows out. Gillis clutches desperately at the steering wheel and manages to turn the careening car into A-23 A DRIVEWAY It is overgrown with weeds and screened from the street by bushes and trees. Gillis stops his car about thirty feet from the street and looks back. GILLIS' VOICE Was I far enough ahead? A-24 THE OTHER CAR shoots past the driveway, still looking for Gillis. A-25 GILLIS He watches his pursuers GILLIS' VOICE shoot past and out of Yeah... sight. He opens the door and looks down at I had landed myself in the the flat tire. Then he driveway of some big mansion looks around to see that looked run-down and where he is. deserted. At the end of the drive was a lovely sight A-26 DRIVEWAY WITH GARAGE indeed -- a great big empty garage, just standing there An enormous, five-car going to waste. If ever there affair. neglected and was a place to stash away a empty-looking. limping car with a hot license number... A-27 GILLIS He gets back into his There was another occupant in car and carefully pilots that garage: an enormous the limping vehicle into foreign-built automobile. It one of the stalls. In must have burned up ten gallons the adjoining one is a to a mile. It had a 1932 large, dust-covered license. I figured that's Isotta-Fraschini propped when the owners moved out... up on blocks. He closes I also figured I couldn't go the garage door and walks back to my apartment now that up the driveway. In idle those bloodhounds were on to curiosity he mounts a me. The idea was to get Artie stone staircase which Green's and stay there till I leads to the garden. could make that bus for Ohio. CAMERA IN BACK OF HIM. Once back in Dayton I'd drop At the top of the steps the credit boys a picturepost- he sees the somber pile card telling them where to of pick up the jallopy. NORMA DESMOND'S HOUSE GILLIS' VOICE It is a grandiose -- It was a great big white Italianate structure, elephant of a place. The kind mottled by the years, crazy movie people built in the gloomy, forsaken, crazy Twenties. A neglected little formal garden house gets an unhappy look. completely gone to This one had it in spades. It seed. was like that old woman in Great Expectations -- that Miss From somewhere above Haversham in her rotting wed- comes ding dress and her torn veil, taking it out on the world be- cause she'd been given the go- by. A WOMAN'S VOICE You there! Gillls turns and looks. A-28 UPSTAIRS LOGGIA Behind a bamboo blind there is a movement of a dark figure. WOMAN'S VOICE Wlly are you so late? Why have you kept me waitlng so long? A-29 GILLIS He stands flabbergasted. A new noise attracts his attention -- the creak of a heavy metal-and-glass door being opened. He turns and sees A-3O THE ENTRANCE DOOR OF THE HOUSE Max von Mayerling stands there. He is sixty, and all in black, except for immaculate white cotton gloves, shirt, high, stiff collar and a white bow tie. His coat is shiny black alpaca, his trousers ledger-atriped. He is semi-paralyzed. The left side of his mouth is pulled down, and he leans on a rubber-ferruled stick. MAX In here! Gillis enters the shot. GILLIS I just put my car in the garage. I had a blow-out. I thought -- MAX Go on in. There is authority in the gesture of his white- gloved hand as he motions Gillis inside. GILLIS Look, maybe I'd better take my car -- MAX Wipe your feet! Automatically, Gillis wipes his feet on an enormous shabby cocoanut mat. MAX You are not dressed properly. GILLIS Dressed for what? THE WOMAN'S VOICE Max! Have him come up, Max! MAX (Gesturing) Up the stairs! GILLIS Suppose you listen just for a minute - MAX Madame is waiting. GILLIS For me? Okay. Gillis enters. A-31 INT. NORMA DESMOND'S ENTRANCE HALL It is grandiose and grim. The whole place is one of those abortions of silent-picture days, with bowling alleys in the cellar and a built-in pipe organ, and beams imported from Italy, with California termites at work on them. Portieres are drawn before all the windows, and only thin slits or sunlight find their way in to fight the few electric bulbs which are always burning. Gillis starts up the curve of the black marble staircase. It has a wrought-iron rail and a worn velvet rope along the wall. MAX (From below) If you need help with the coffin call me. The oddity of the situation has caught Gillis' imagination. He climbs the stairs with a kind of morbid fascination. At the top he stops, undecided, then turns to the right and is stopped by WOMAN'S VOICE This way! Gillis swings around. Norma Desmond stands down the corridor next to a doorway from which emerges a flickering light. She is a little woman. There is a curious style, a great sense of high voltage about her. She is dress- ed in black house pyjamas and black high-heeled pumps. Around her throat there is a leopard-pat- terned scarf, and wound around her head a turban of the same material. Her skin is very pale, and she is wearing dark glasses. NORMA In here. I put him on my massage table in front of the fire. He always liked fires and poking at them with a stick. Gillis enters the SHOT and she leads him into A-32 NORMA DESMOND'S BEDROOM It is a huge, gloomy room hung in white brocade which has beconle dirty over the years and even slightly torn in a few places. There's a great, unmade gilded bed in the shape of a swan, from which the gold had begun to peel. There is a disorder of clothes and negligees and faded photographs of old-time stars about. In an imitation baroque fireplace some logs are burn- ing. On the massage table before it lies a small form shrouded under a Spanish shawl. At each end on a baroque pedestal stands a three-branched cande- labrum, the candles lighted. NORMA I've made up my mind we'll bury him in the garden. Any city laws against that? GILLIS I wouldn't know. NORMA I don't care anyway. I want the coffin to be white. And I want it specially lined with satin. White, or deep pink. She picks up the shawl to make up her mind about the color. From under the shawl flops down a dead arm. Gillis stares and recoils a little. It is like a child's arm, only black and hairy. NORMA Maybe red. bright flaming red. Gay. Let's make it gay. Gillis edges closer and glances down. Under the shawl he sees the sad, bearded face of a dead chimpanzee. Norma drops back the shawl. NORMA How much will it be? I warn you - don't give me a fancy price just because I'm rich. GILLIS Lady. you've got the wrong man. For the first time. Norma really looks at him through her dark glasses. GILLIS I had some trouble with my car. Flat tire. I pulled into your garage till I could get a spare. I thought this was an empty house. NORMA It is not. Get out. GILLIS I'm sorry, and I'm sorry you lost your friend, and I don't think red is the right color. NORMA Get out. GILLIS Sure. Wait a minute -- haven't I seen you -- ? NORMA Or shall I call my servant? GILLIS I know your face. You're Norma Desmond. You used to be in pictures. You used to be big. NORMA I am big. It's the pictures that got small. GILLIS I knew there was something wrong with them. NORMA They're dead. They're finished. There was a time when this busi- ness had the eyes of the whole wide world. But that wasn't good enough. Oh, nol They wanted the ears of the world, too. So they opened their big mouths, and out came talk, talk, talk... GILLIS That's where the popcorn business comes in. You buy yourself a bag and plug up your ears. NORMA Look at them in the front offices -- the master minds! They took the idols and smashed them. The Fairbankses and the Chaplins and the Gilberts and the Valentinos. And who have they got now? Some nobodies -- a lot of pale little frogs croaking pish-poshl GILLIS Don't get sore at me. I'm not an executive. I'm just a writer. NORMA You are! Writing words, words! You've made a rope of words and strangled this businessl But there is a microphone right there to catch the last gurgles, and Technicolor to photograph the red, swollen tongue! GILLIS Ssh! You'll wake up that monkey. NORMA Get out! Gillis starts down the stairs. GILLIS Next time I'll bring my autograph album along, or maybe a hunk of cement and ask for your footprints. He is halfway down the staircase when he is stopped by NORMA Just a minute, you! GILLIS Yeah? NORMA You're a writer, you said. GILLIS Why? Norma starts down the stairs. NORMA Are you or aren't you? GILLIS I think that's what it says on my driver's license. NORMA And you have written pictures, haven't you? GILLIS Sure have. The last one I wrote was about cattle rustlers. Before they were through with it, the whole thing played on a torpedo boat. Norma has reached him at the bottom of the staircase. NORMA I want to ask you something. Come in here. She leads him into A-33 THE HUGE LIVING ROOM It is dark and damp and filled with black oak and red velvet furniture which looks like crappy props from the Mark of Zorro set. Along the main wall, a gigantic fireplace has been freezing for years. On the gold piano is a galaxy of photographs of Norma Desmond in her various roles. On one wall is a painting -- a California Gold Rush scene, Carthay Circle school. (We will learn later that it hides a motion picture screen.) One corner is filled with a large pipe organ, and as Norma and Gillis enter, there is a grizzly moaning sound. Gillis looks around. NORMA The wind gets in that blasted pipe organ. I ought to have it taken out. GILLIS Or teach it a better tune. Norma has led him to the card tables which stand side by side near a window. They are piled high with papers scrawled in a large, uncertain hand. NORMA How long is a movie script these days? I mean, how many pages? GILLIS Depends on what it is -- a Donald Duck or Joan or Arc. NORMA This is to be a very important picture. I have written it myself. Took me years. GILLIS (Looking at the piles of script) Looks like enough for six impor- tant pictures. NORMA It's the story or Salome. I think I'll have DeMille direct it. GILLIS Uh-huh. NORMA We've made a lot of pictures together. GILLIS And you'll play Salome? NORMA Who else ? GILLIS Only asking. I did't know you were planning a comeback. NORMA I hate that word. It is a return. A return to the millions of people who have never forgiven me for deserting the screen. GILLIS Fair enough. NORMA Salome -- what a woman! What a part! The Princess in love with a Holy man. She dances the Dance of the Seven Veils. He rejects her, so she demands his head on a golden tray, kissing his cold, dead lips. GILLIS They'll love it in Pomona. NORMA (Taking it straight) They will love it every place. (She reaches for a batch of pages from the heap) Read it. Read the scene just before she has him killed! GILLIS Right now? Never let another writer read your stuff. He may steal it. NORMA I am not afraid. Read it! NORMA (Cont'd) (Calling) Max! Max! (To Gillis) Sit down. Is there enough light? GILLIS I've got twenty-twenty vision. Max has entered. NORMA Bring something to drink. MAX Yes. Madame. He leaves. Norma turns to Gillis again. NORMA I said sit down. There is compulsion in her voice. Gillis looks at her GILLIS' VOICE and starts slowly Well. I had no pressing reading. engagement, and she'd men- tioned something to drink.. Max comes in, wheeling Sometimes it's interesting a wicker tea wagon on to see just how bad bad which are two bottles o writing can be. This prom- f champagne and two ised to go the limit. I red Venetian glasses, wondered what a handwriting a box of zwieback and expert would make of that a jar of caviar. Norma childish scrawl of hers. sits on her feet. deep Max wheeled in some champagne in a chair, a gold ring and some caviar. Later, I on her forefinger with found out that Max was the a clip which holds a only other person in that cigarette. She gets up grim Sunset castle, and I and forces on Gillis found out a few other things another batch of script, about him... As for her, she goes back to her chair. sat coiled up like a watch spring, her cigarette clamped in a curious holder... I could sense her eyes on me from behind those dark glasses, defying me not to like what I read, or maybe begging me in her own proud way to like it. It meant so much to her... A-34 SHOT OF THE GILLIS' VOICE CEILING It sure was a cozy set-up. That bundle of raw nerves,and PAN DOWN to the moan- Max, and a dead monkey upstair ing organ. PAN OVER and the wind wheezing through TO THE ENTRANCE DOOR. that organ once in a while. Max opens it, and a Later on, just for comedy solemn-faced man in relief, the real guy arrived undertaker's clothes with a baby coffin. It was brings in a small all done with great dignity. white coffin. (Thru He must have been a very these shots the room important chimp. The great has been growing grandson of King Kong, maybe. duskier.) DISSOLVE TO: A-35 GILLIS It got to be eleven. I was feeling a little sick at my reading. The lamp stomach, what with that sweet beside him is now champagne and that tripe I'd really paying its been reading -- that silly way in the dark room. hodgepodge of melodramatic A lot of the manu- plots. However, by then I'd script pages are started concocting a little piled on the floor plot of my own... around his feet. A half-empty champagne glass stands on the arm of his chair. THE CAMERA SLOWLY DRAWS BACK to include Norma Desmond sitting in the dusk, just as she was before. Gillis puts down a batch of script. There is a little pause. NORMA (Impatiently) Well? GILLIS This is fascinating. NORMA Of course it is. GILLIS Maybe it's a little long and maybe there are some repetitions... but you're not a professional writer. NORMA I wrote that with my heart. GILLIS Sure you did. That's what makes it great. What it needs is a little more dialogue. NORMA What for? I can say anything I want with my eyes. GILLIS It certainly could use a pair of shears and a blue pencil. NORMA I will not have it butchered. GILLIS Of course not. But it ought to be organized. Just an editing job. You can find somebody. NORMA Who? I'd have to have somebody I can trust. When were you born -- I mean, what sign of the zodiac? GILLIS I don't know. NORMA What month? GILLIS December twenty-first. NORMA Sagittarius. I like Sagittarians. You can trust them. GILLIS Thank you. NORMA I want you to do this work. GILLIS Me? I'm busy. Just finished one script. I'm due on another assignment. NORMA I don't care. GILLIS You know, I'm pretty expensive. I get five hundred a week. NORMA I wouldn't worry about money. I'll make it worth your while. GILLIS Maybe I'd better take the rest of the script home and read it - NORMA Oh no. I couldn't let it out of my house. You'll have to finish it here. GILLIS It's getting kind of late -- NORMA Are you married, Mr. -- ? GILLIS The name is Gillis. I'm single. NORMA Where do you live? GILLIS Hollywood. The Alto Nido Apart- ments. NORMA There's something wrong with your car, you said. GILLIS There sure is. NORMA You can stay here. GILLIS I'll come early tomorrow. Norma takes off her glasses. NORMA Nonsense. There's room over the garage. Max will take you there...Max! THE CAMERA MOVES GILLIS' VOICE TOWARD NORMA'S FACE, She sure could say a lot of right up to her things with those pale eyes of eyes. hers. They'd been her trade mark. They'd made her the Num- ber One Vamp of another era. I remember a rather florid des- cription in an old fan magazine which said: "Her eyes are like two moonlit waterholes, where strange animals come to drink." DISSOLVE TO: A-36 SMALL STAIRCASE, LEAD- GILLIS'VOICE ING TO ROOM OVER GARAGE I felt kind of pleased with the way I'd handled the sit- Max, an electric light uation. I'd dropped the hook, bulb in his hand, is and she'd snapped at it. Now leading Gillis up. my car would be safe down Gillis carries a batch below, while I did a patch- of the manuscript. up job on the script. And there should be plenty of money in it... Max pushes open a door at the top of the stairs. MAX (Opening the door) I made your bed this afternoon. GILLIS Thanks. (On second thought) How did you know I was going to stay, this afternoon? Max doesn't answer. He walks across to the bed, screws a bulb in the open socket above it. The light goes on, revealing: A-37 A GABLED BEDROOM There are dirty windows on two sides, and dingy wall- paper on the cracked plaster walls. For furniture there is a neatly made bed, a table and a few chairs which might have been discarded from the main house. MAX This room has not been used for a long time. GILLIS It will never make house Beautiful. I guess it's O.K. for one night. Max gives him an enigmatic look. MAX (Pointing) There is the bathroom. I put in soap and a toothbrush. GILLIS Thanks. (He starts taking off his coat) Say, she's quite a character, that Norma Desmond. MAX She was the greatest. You wouldn't know. You are too young. In one week she got seventeen thousand fan letters. Men would bribe her mani- curist to get clippings from her fingernails. There was a Maharajah who came all the way from Hyderabad to get one of her stockings. Later, he strangled himself with it. GILLIS I sure turned into an interesting driveway. MAX You did, sir. GILLIS' VOICE He goes out. Gillis I pegged him as slightly looks after him, hangs cuckoo, too. A stroke maybe. his coat over a chair, Come to think of it, the walks over to the win- whole place seemed to have dow, pulls down the been stricken with a kind of rickety Venetian blind. creeping paralysis, out of As he does so, he looks beat with the rest of the down at: world, crumbling apart in slow motion ... A-38 THE TENNIS COURT OF GILLIS' VOICE THE DESMOND HOUSE There was a tennis court, or (MOONLIGHT) rather the ghost of a tennis court, with faded markings The cement surface is and sagging net ... cracked in many places, and weeds are growing high. A-39 GILLIS - IN THE WINDOW He looks away from the court to: A-40 THE DESMOND SWIMMING POOL GILLIS' VOICE There is no water in And of course she had a pool. it, and hunks of Who didn't then? Mabel Norm- mosaic which lines its and and John Gilbert must enormous basin are have swum in it ten thousand broken away. midnights ago, and Vilma Banky and Rod La Roque. It was empty now....or was it? A-41 GILLIS - IN THE WINDOW He stares down, his stomach slowly turning. A-42 THE SWIMMING POOL At the bottom of the basin a great rat is eating a decaying or,ange. From the inlet pipe crawl two other rats, who join battle with the first rat over the orange. A-43 GILLIS -IN THE WINDOW He starts away, but some- GILLIS' VOICE thing attracts his atten- There was something tion. He turns back and else going on below: looks down again. the last rites for that hairy old chimp, performed with the A-44 THE LAWN BELOW utmost seriousness -- as if she were laying Norma Desmond and Max are to rest an only child. carrying the white coffin Was her life really towards a small grave as as empty as that? which has been dug in the dead turf. Norma carries one of the candelabra, all of its candles flickering in the wind. They reach the grave and lower the coffin into it. Then, Norma lighting his task with the candelabrum, Max takes a spade from the loose earth and starts filling in the grave. A-45 GILLIS - IN THE WINDOW He watches the scene be- GILLIS' VOICE low, then turns into the It was all very queer, room, goes to the door but queerer things to lock it. There is no were yet to come. key, and only a hole where the lock has been gouged out. Gillis moves a heavy overstuffed chair in front of the door, then walks towards the bed, throws himself on it, picking up some of the manuscript pages to read. DISSOLVE END OF SEQUENCE "A" SEQUENCE "B" DISSOLVE IN ON: B-1 LONG SHOT THE DESMOND HOUSE - (MORNING) The day is overcast. The SOUND: (Distant organ house is shrouded in low music - improvisations fog. on an odd, mournful theme - not too loud, continuing throughout B-2 THE TENNIS COURT, blurred the scene.) over with fog. B-3 THE EMPTY SWIMMING POOL Its dark outline even more melancholy under the misty blanket. B-4 THE ROOM OVER THE GARAGE Muted daylight seeps GILLIS' VOICE through the blinds. Gillis That night I'd had a lies on the bed, under a mixed-up dream. In it shabby quilt. The manu- was an organ grinder. script is beside him, some I couldn't see his of the pages scattered on face, but the organ the floor. He is just was all draped in opening his eyes. It takes black, and a chimp was him a moment to adjust him- dancing for pennies. self to the strange sur- When I opened my eyes, roundings. His eyes, wander- the music was still ing about the room. suddenly there... Where was stop, startled. He lifts I? himself on one elbow and stares at - B-5 THE DOOR The heavy chair he had set Oh yes, in that empty against it the night before room over her garage. has been pushed back. The Only it wasn't empty door is wide ajar. any more. Somebody had brought in all my belongings - my B-6 GILLIS books, my typewriter, my clothes... He jumps out of bed. He wears, shirt, trousers and socks. Suddenly he realizes that all his possessions have GILLIS' VOICE been brought in. In What was going on? the closet hang his shirts. His books and typewriter are neatly arranged on the table. His phonograph-radio combination is all installed. Gillis looks around startled, then sits down and starts putting on his moccasins hastily. DISSOLVE TO: B-7 A PAIR OF HANDS IN WHITE GLOVES, PLAYING THE ORGAN PULL BACK: They belong to Max von Mayerling. He is sitting erect, his bull neck taut as a wrestler's as he rights out somber chord after somber chord. He sits in a shaft of gray light coming from an open French window. Through the far archway, Gillis storms into the big room. GILLIS Hey, you -- Max -- whatever -your- name-is -- what are my things doing here? No answer. GILLIS I'm talking to you. My clothes and things are up in the room. MAX Naturally. I brought them myself. GILLIS (Furiously) Is that so! MAX Why are you so upset? Is there anything missing? GILLIS Who said you could? Who asked you to? Norma Desmond's shadow moves into the shaft of light. NORMA'S VOICE I did. Gillis looks around. On the couch by the fireplace reclines Norma Desmond, dressed in a negligee. She rises. NORMA I don't know why you should be so upset. Stop that playing, Max. (To Gillis again) It seemed like a good idea -- if we are to work together. GILLIS Look, I'm supposed to fix up your script. There's nothing in the deal about my staying here. NORMA You'll like it here. GILLIS Thanks for the invitation, but I have my own apartment. NORMA You can't work in an apartment where you owe three months' rent. GILLIS I'll take care of that. NORMA It's all taken care of. It's all paid for. GILLIS I'm used to paying my own bills. NORMA You proud boy, why didn't you tell me you were having difficulties. GILLIS Okay. We'll deduct it from my salary. NORMA Now, now, don't let's be small about such matters. We won't keep books. (To Max) Go on, unpack Mr. Gillis' things. GILLIS Unpack nothing. I didn't say I was staying. NORMA (Her glasses off again) Suppose you make up your mind. Do you want this job or don't you? DISSOLVE TO: B-8 BIG ROOM, NORMA DESMOND'S HOUSE - (DAY) GILLIS' VOICE Gillis sits at an impro- So I let him unpack my vised table, his typewriter things. I wanted the in front of him, working dough, and I wanted to hard at the manuscript. get out of there as Pencils, shears and a quickly as possible. paste-pot at hand. I thought if I really got going I could toss Facing him at some dis- it off in a couple or tance sits Norma,dressed weeks. But it wasn't in another version of her so simple, getting some favorite lounging pajamas, coherence into that wild, the cigaette contraption scrambled melodrama on her finger. She is she'd concocted. What autographing large photo- made it tougher was that graphs of herself and put- she was around all the ting them in envelopes. time -- hovering over me, afraid I'd do injury to that precious brain- child of hers. Gillis takes two or three pages from Norma's hand- written script, crosses them out and puts them to one side. Norma rises, crosses towards Gillis, looks over his shoulder. NORMA What's that? GILLIS Just a scene I cut out. NORMA What scene? GILLIS The one where you go to the slave market. You can cut right to the scene where John the Baptist - NORMA Cut away from me? GILLIS Honestly, it's a little old hat. They don't want that any more. NORMA They don't? Then why do they still write me fan letters every day. Why do they beg me for my photo- graphs? Because they want to see me, me, me! Norma Desmond. GILLIS (Resigned) Okay. He pulls the page from his typewriter. As he does so he glances over towards Norma. GILLIS' VOICE On the table in front I didn't argue with her. of her are the photo- You don't yell at a graphs which she is sign- sleepwalker-- he may fall ing. On the long table and break his neck.That's in the living room is a it -- she was still gallery of photographs sleepwalking along the in various frames -- all giddy heights of a lost Norma Desmond. On the career --plain crazy piano more photographs. when it came to that one Above the piano an oil subject: her celluloid portrait of her. On the self, the great Norma highboy beside him still Desmond. How could She more photographs. breathe in that house, so crowded with Norma DISSOLVE TO: Desmonds? More Norma Desmond and still more Norma Desmond. B-9 THE BIG ROOM - (NIGHT) GILLIS' VOICE Shooting towards the big It wasn't all work - of Gold Rush painting. Max, course. Two or three white gloves and all, times a week Max would steps into the shot, shoves haul up that enormous oil the painting up towards painting that had been the ceiling,revealing a presented to her by some motion picture screen. Nevada Chamber of Com- Max exits. merce, and we'd see a movie,right in her living room. B-1O NORMA AND GILLIS GILLIS' VOICE They sit on a couch,facing "So much nicer than going the screen. On a table in out," she'd say. The front of them are champagne, plain fact was that she cigarettes and coffee. was afraid of that world Above their heads are the outside. Afraid it typical openings for a pro- would remind her that jector. The lights go off. time had passed. From the opening above their heads shoots the wide beam of light. B-11 MAX, IN THE PROJECTION They were silent movies, BOOTH BEHIND THE ROOM and Max would run the projection machine, which The light of the machine was just as well -- it flickering over his face, kept him from giving us which is frozen, a somber an accompaniment on enigma. that wheezing organ. B-12 NORMA AND GILLIS She'd sit very close to watching the screen. me, and she'd smell of Gillis looks down and sees tuberoses, which is not that Norma's hand is clasp- my favorite perfume, not ing his ann tight. He by a long shot. Sometines doesn't like it much but as we watched, she'd c he can't do anything about lutch my arm or my hand it. However. when she for forgetting she was my a second lets go his arm employer becoming just a to pick up a glass of fan, excited about that champagne, he gently with- actress up there on the draws his arm, leans away screen....I guess I don't from her and crosses his have to tell you who the arms to discourage any star was. They were resumption of her approach. always her pictures -- Norma puts the glass down that's all she wanted doesn't find his arn, but to see. is not aware of any signifi- cance in his maneuver. They both watch the screen. B-13 THE OTHER END OF THE BIG ROOM. WITH THE SCREEN On it flickers a famous scene from one of Norma's old silent pictures. It is not to be a funny scene. It is old-fashioned, but shows her incredible beauty and the screen presence which made her the great star of her day. B-14 NORMA AND GILLIS ON THE COUCH NORMA Still wonderful, isn't it? And no dialogue. We didn't need dialogue. We had faces. There just aren't any faces like that any more. Well, maybe one -- Garbo. In a sudden flareup she jumps to her feet and stands in the flickering beam of light. NORMA Those idiot producers! Those imbeciles! Haven't they got any eyes? Have they forgotten what a star looks like? I'll show them. I'll be up there again. So help me! DISSOLVE TO: B-15 THE BIG ROOM - (NIGHT) It is apparently empty. GILLIS' VOICE The elaborate lamps Sometimes there'd be a make pools of light. little bridge game in the house, at a twentieth-of- THE CAMERA PULLS BACK a cent a point. I'd get AND PANS to reveal a half her winnings. Once card table around they ran up to seventy which sit Norma and cents, which was about three friends - three the only cash money I actors of her period. ever got. The others They sit erect and play around the table would with grim seriousness. be actor friends - dim figures you may still Beside Norma sits remember from the silent Gillis, kibitzing on a days. I used to think of game which bores him them as her Wax Works. extremely. An ashtray on the card table is full and Norma holds it out for Gillis to take away. He crosses the room to the fire- place. but his eyes fall on the entrance door and he stops. B-16 THE ENTRANCE HALL - (FROM GILLIS' POINT OF VIEW) Max stands in the open door. Outside are the two men who came to the apartment for Gillis' car. B-17 GILLIS He steps back so that he cannot be seen from the door. A second later Max appears, looking for him. MAX (Quietly) Some men are here. They asked for you. GILLIS I'm not here. MAX That's what I told them. GILLIS Good. MAX They found your car in the garage. They are going to tow it away. Gillis doesn't know what to do. From offstage comes: NORMA'S VOICE The ashtray, Joe dear! Can we have the ashtray? Gillis dumps the cigarette butts into the cold fire- place, crosses to the bridge table, puts the ashtray down, leans over and speaks into Norma's ear. GILLIS I want to talk to you for a minute. NORMA Not now, my dear. I'm playing three no trump. GILLIS They've come for my car. NORMA Please. Now I've forgotten how many spades are out. GILLIS I need some money right now. NORMA Can't you wait till I'm dummy? 3.22.49 GILLIS No. NORMA (Angry by now) Please! Gillis stands frustrated, hideously embarrassed by the stares of the waxworks. He turns away and hurries to the door. B-18 ENTRANCE DOOR TO THE HOUSE It is half open. Gillis comes into the shot and, taking cover, looks out. B-19 COURTYARD (FROM GILLIS' ANGLE) The men from the finance company are cranking up the car. Max stands watching silently. When they finish the cranking job, the men climb into the front seat of the truck. B-2O GILLIS - AT THE DOOR Over the shot the SOUND of the truck being started and the cars moving away. Gillis moves out into the courtyard and stands staring after the car. From the house comes Norma. NORMA Now what is it? Where's the fire? GILLIS I've lost my car. NORMA Oh...and I thought it was a matter of life and death. GILLIS It is to me. That's why I came to this house. That's why I took this job -- ghost writing! NORMA Now you're being silly. We don't need two cars. We have a car. And not one of thuse cheap new things made of chromium and spit. An Isotta-Fraschini. Have you ever heard of Isotta-Fraschinis? All hand-made. Cost me twenty-eight thousand dollars. THE CAMERA HAS PANNED over to the garage and FOCUSES on the dirty Isotta-Fraschini on its blocks. DISSOLVE TO: B-21 NORMA'S ISOTTA-FRASCHINI DRIVING IN THE HILLS ABOVE SUNSET (DAY) Max is at the wheel, GILLIS' VOICE dressed as usual except So Max got that old bus for a chauffeurfs cap. down off its blocks and polished it up. She'd take me for rides in the B-22 INSIDE THE CAR hills above Sunset. Gillis sits beside Norma, The whole thing was up- who is wearing a smart holstered in leopard tailleur and her eternal skin, and had one of sun glasses. Gillis those car phones, all wears his sport jacket- gold-plated. flannel trousers-moccasin combinatIon. He sits uncomfortably. Norma is studying him. NORMA That's a dreadful shirt you're wearing. GILLIS What's wrong with It? NORMA Nothing, if you work in a fill- ing station. And I'm getting rather bored with that sport jacket, and those same baggy pants. (She picks up the car phone) Max, what's a good men's shop in town? The very best... Well, go there ! GILLIS I don't need any clothes, and I certainly don't want you buy- ing them for -- NORMA Why begrudge me a little fun? I just want you to look nice, my stray little boy. By this time Max has made a U-turn. QUICK DISSOLVE TO: B-23 INT. MEN'S DEPARTMENT, AN ELEGANT WILSHIRE STORE Gillis stands in front of a full-length triple mirror, surrounded by a couple of salesmen and the tailor, who is busily working out alterations. Gillis wears a double-breasted gray flannel coat with chalk stripes. His trousers belong to another suit of glen plaid. Norma is running the show. NORMA There's nothing like gray flannel with a chalk stripe. (she points at the trousers) This one single-breasted, of course. (to another salesman) Now we need a topcoat. Let's see what you have in camel's hair. The salesman leaves. NORMA How about some evening clothes? GILLIS I don't need a tuxedo. NORMA Of course you do. A tuxedo and tails. GILLIS Tails. That's ridiculous. NORMA You'll need them for parties. You'll need them for New Year's Eve. (to a salesman) Where are your evening clothes? SALESMAN This way, Madame. He leads her off. The other salesman arrives with a selection of topcoats. SALESMAN Here are some camel hairs, but I'd like you just to feel this one. It's Vicuna. Of course, it's a little more expensive. GILLIS A camel's hair will do. SALESMAN (With an insulting inflection) As long as the lady is paying for it, why not take the Vicuna? DISSOLVE: END OF SEQUENCE "B" SEQUENCE "C" DISSOLVE IN: C-1 LONG SHOT DESMOND HOUSE A day in December. Rain. QUICK DISSOLVE TO: C-2 INT. ROOM OVER GARAGE Water is drizzling from GILLIS' VOICE two or three spots in the The last week in December ceiling into pans and the rains came -- a great bowls set to catch it, big package of rain. one bowl right on the Over-sized, like every- bed. The room is almost thing else in California. emptied of Gillis' be- longings by now. Max It came right through is carrying out a hand- the old roof of my room full of new suits on above the garage. She hangers. He has a had Max move me to the dressing gown over his main house. I didn't shoulder. Gillis holds much like the idea -- the a stack of shirts, his only time I could have typewriter, and some to myself was in that manuscript. He surveys room -- but it was better the room for the last than sleeping in a rain- time, to see whether coat and galoshes. he's forgotten any- thing. He has. He puts down the typewriter and picks up from under the bed a pair of very smart red leather bedroom slippers. He tucks them under his arm, picks up the typewriter and leaves. QUICK DISSOLVE TO: C-3 A BEDROOM IN TIiE MAIN HOUSE It is obviously a man's room -- heavy Spanish furniture -- one wall nothing but a closet with shelves and drawers for shirts and shoes. Max is hanging up the suits. Gillis throws the shirts on a big chair, tosses the slippers at the foot of the bed, places the typewriter and manuscript on a desk at the window. GILLIS Whose room was this? MAX It was the room of the husband. Or of the husbands, I should say. Madame has been married three times. Slightly embarrassed, Gillis picks up his toilet kit with razor, toothbrushes, soap, etc., and starts towards the bathroom, pausing en route at a rain- splattered window. GILLIS I guess this is the one you can see Catalina from. Only this isn't the day. He proceeds towards the half-opened door leading to the bathroom. Something strikes his attention and he stops. As in the door to the room above the garage, this lock, too, has been gouged out. GILLIS Hey, what's this with the door? There isn't any lock. MAX There are no locks anywhere in this house. He points to the entrance door of the room, and to another door. GILLIS How come? MAX The doctor suggested it. GILLIS What doctor? MAX Madame's doctor. She has moments of melancholy. There have been some suicide attempts. GILLIS Uh-huh? MAX We have to be very careful. No sleeping pills, no razor blades. We shut off the gas in her bed- room. GILLIS Why? Her career? She got enough out of it. She's not forgotten. She still gets those fan letters. MAX I wouldn't look too closely at the postmarks. GILLIS You send them. Is that it, Max? MAX I'd better press your evening clothes, sir. You have not for- gotten Madame's New Year's party. GILLIS No, I haven't. I suppose all the waxworks are coming? MAX I don't know, sir. Madame made the arrangements. Max leaves. Gillis comes out of the bathroom, picks up his shirts, goes over to a closet, opens it. As he does so one of the doors without a lock swings slightly open. Gillis looks through the half-open door and sees. C-4 NORMA DESMOND'S ROOM It is empty. The rainy GILLIS' VOICE day does nothing to There it was again - that help its gloom. room of hers, all satin and ruffles, and that bed like a gilded rowboat. The per- fect setting for a silent movie queen. Poor devil, still waving proudly to a parade which had long since passed her by. He pushes the door shut and walks back into the room. DISSOLVE TO: C-5 STAIRCASE OF DESMOND HOUSE (NIGHT) Gillis is coming down the GILLIS' VOICE stairs in his tailcoat It was at her New Year's adjusting the handkerchief party that I found out in his pocket. He obviously how she felt about me. feels a little uneasy in Maybe I'd been an idiot this outfit. From below not to have sensed it comes a tango of the Twen- was coming - that sad, ties. played by a small embarrassing revelation. orchestra. Gillis stops in the archway leading to the big room and looks around. C-6 THE BIG ROOM has been deco- rated for the occasion with laurel garlands. Dozens of candles in all the sconces and candelabra are ablaze. Their flickering flames are reflected in the waxed sur= face of the tile floor. There is a buffet, with buckets of champagne and caviar on ice. In one corner on a little platform banked with palms. a four-piece orchestra is playing. At the buffet are Max and Norma. She is drinking a glass of champagne. She is wearing a diamonte evening dress. very high style. with long black gloves and a headdress of paradise feathers. Her eyes fall on Gillis. She puts down the glass of champagne. picks up a gardenia boutonniere and moves toward him. NORMA Joe, you look absolutely divine. Turn around! GILLIS (Embarrassed} Please. NORMA Come on! Gillis makes a slow 36O-degree turn. NORMA Perfect. Wonderful shoulders. And I love that line. She indicates the V from his shoulders to his hips. GILLIS All padding. Don't let it fool you. NORMA Come here! She puts the gardenia on his lapel. GILLIS You know, to me dressing up was always just putting on my dark blue suit. NORMA I don't like those studs they've sent. I want you to have pearls. Nice big pearls. GILLIS Now, I'm not going to wear ear- rings, I can tell you that. NORMA Cute. Let's have some drinks. She leads him over to the buffet. GILLIS Shouldn't we wait for the others? NORMA (Pointing at the floor) Careful, it's slippery. I had it waxed. They reach the buffet. Max is ready with two glasses of champagne. Norma hands Gillis a glass. NORMA Here's to us. They drink. NORMA You know, this floor used to be wood but I had it changed. Valentino said there is nothing like tiles for a tango. She opens her arms. GILLIS Not on the same floor with Valentino! NORMA Just follow me. They start to tango. After a moment -- NORMA Don't bend back like that. GILLIS It's those feathers. They tickle. Norma pulls the paradise feathers from her hair and tosses them away. C-7 THE ORCHESTRA As they play the tango, the musicians eye the danc- ing couple, take in the situation, exchange glances and turn away with professional discretion. C-8 NORMA AND GILLIS, TANGOING Gillis glances at his wrist watch. GILLIS It's a quarter past ten. What time are they supposed to get here? NORMA Who? GILLIS The other guests? NORMA There are no other guests. We don't want to share this night with other people. This is for you and me. GILLIS I understand some rich guy bought up all the tickets for a perfor- mance at the Metropolitan and sat there listening to La Traviata, all by himself. He was afraid of catching cold. NORMA Hold me tighter. GILLIS Come midnight, how about blind- folding the orchestra and smash- ing champagne glasses on Max's head? NORMA You think this is all very funny. GILLIS A little. NORMA Is it funny that I'm in love with you? GILLIS What's that? NORMA I'm in love with you. Don't you know that? I've been in love with you all along. They dance on. Gillis is acutely embarrassed. THE CAMERA SLOWLY PULLS BACK, PANS past the faces of the musicians, who play on with a rather overe- mphasized lack of interest. Finally it winds up on Max, behind the buffet. He stands watching Gillis, a faint trace of pity in his eyes. DISSOLVE TO: C-9 NORMA'S FINGER, WITH THE CIGARETTE GADGET, as she GILLIS' VOICE inserts a cigarette. I'm sure a lot of you will laugh about this. Ridicu- lous situation, wasn't it? -- a woman almost twice my age ... It got to be about a quarter of eleven. I felt caught, like a cig- arette in the prongs of that contraption on her finger. PULL BACK TO: NORMA AND GILLIS sitting on a couch in front of the cavernous fireplace. Norma holds out her cigarette to Gillis, who lights it. NORMA. What a wonderful next year it's going to be. What fun we're going to have. I'II fill the pool for you. Or I'll open my house in Malibu, and you can have the whole ocean. Or I'll buy you a boat and we'll sail to Hawaii. GILLIS Stop it. You aren't going to buy me anything more. NORMA Don't be silly. (She reaches under a pillow of the couch and brings out a leather box) Here. I was going to give it to you at midniglht. Gillis opens the box. It contains a matched gold cigarette case and lighter. NORMA Read what's inside. Gillis snaps open the case. Engraved inside the cover is: TO JOE FROM NORMA, and two bars of music. GILLIS What are the notes? NORMA "Mad about the boy." GILLIS Norma, I can't take it. You've bought me enough. NORMA Shut up. I'm rich. I'm richer than all this new Hollywood trash. I've got a million dollars. GILLIS Keep it. NORMA I own three blocks downtown. I have oil in Bakersfield -- pumping, pumping, pumping. What's it for but to buy us anything we want. GILLIS Cut out that us business. He rises. NORMA What's the matter with you? GILLIS What right do you have to take me for granted? NORMA What right? Do you want me to tell you? GILLIS Has it ever occurred that I may have a life of my own? That there may be some girl I'm crazy about? NORMA Who? Some car hop, or a dress extra? GILLIS Why not? What I'm trying to say is that I'm all wrong for you. You want a Valentino -- somebody with polo ponies -- a big shot -- NORMA (Getting up slowly) What you're trying to say is that you don't want me to love you. Is that it? Gillis doesn't answer. Norma slaps his face and rushes from the room and upstairs. Gillis stands paralyzed, the slap burning his cheek. C-1O THE TOP OF THE STAIRCASE AND CORRIDOR Norma rushes up the last few steps, down the corridor and into her bedroom, banging the door. MOVE THE CAMERA toward the closed door, centering on the gouged-out lock. C-11 GILLIS, IN THE BIG ROOM He still stands motionless. He glances around fur- tively, to see if his humiliation has been observed. C-12 THE ORCHESTRA The musicians are playing away. They have turned their eyes away from Gillis rather too ostentatious- ly for comfort. C-13 GILLIS His eyes move over toward C-14 MAX He is subtler than the musicians. He appears very busy at the buffet, putting empty bottles and used glasses on a tray. He walks across the room with them. C-15 GILLIS He starts slowly out. As he does so his long gold key chain catches on a carved ornament of the sofa and holds him for a second of additional embarrass- ment. He yanks it loose and walks with as much nonchalance as he can muster to C-16 THE HALL Crossing towards the coat closet, Gillis throws a look upstairs. Then he pulls the Vicuna coat from its hangar and slips into it as he crosses to the entrance door. He opens the door on the darkness of the courtyard. C-17 EXT. DESMOND HOUSE (NIGHT - RAIN) Gillis shuts the door. GILLIS'VOICE He takes a few steps I didn't know where I was forward, then stands going. I just had to get for a while breathing out of there. I had to be deep. The rain is with people my own age. I balm to that cheek had to hear somebody laugh where the slap still a again. I thought of Artie burns. He walks for- Green. There was bound to ward with a great be a New Year's shindig sense of relief. going on in his apartment down on Las Palmas -- the hock shop set -- not a job C-18 DRIVEWAY LEADING TO in the room. but lots of fun on the cuff. Gillis walks to the street, which is dark and empty. He starts down Sunset in an Easterly direction. A car passes. He tries to thumb a ride, without success. However, the second car, a florist's delivery wagon, stops. Gillis jumps in and the car drives off. DISSOLVE TO: C-19 ARTIE GREEN'S APARTMENT It is the most modest one-room affair, jam packed with young people flowing over into the miniature bathroom and the microscopic kitchenette. The only drink being served is punch from a pressed-glass bowl -- but everybody is having a hell of a time. Most of the men are in slacks and sweaters, and only a few of the girls in something that vaguely suggests party dress. Abe Burroughs sits at a small, guest-festooned piano and sings Tokio Rose. By the door, a group of young men and girls respond to the song by sing1ng Rinso White or Dentyne Chewing Gum or something similar, in the manner of a Bach choral. Artie Green, a dark haired, pleasant-looking guy in his late twenties, is conducting with the ladle from the punch bowl. The door behind some of the singers is pushed open, jostling them out of their places. In comes Gillis, his hair and face wet, the collar of his Vicuna coat turned up. Artie stops conducting, but the commer- cial goes right on. ARTIE Well, what do you know ! Joe Gillis ! GILLIS Hi, Artie. ARTIE Where have you been keeping that gorgeous face of yours? GILLIS In a deep freeze. ARTIE I almost reported you to the Bureau of Missing Persons. (To the company) Fans, you all know Joe Gillis, the well-known screen writer, opium smuggler and Black Dahlia suspect. Gillis greets some of the kids by name as he and Artie push their way into the room. ARTIE Give me your coat. GILLIS Let it ride for a while. ARTIE You're going to stay, aren't you? GILLIS That was the general idea. ARTIE Come on. Artie starts peeling the coat off Gillis. Its texture takes his breath away. ARTIE What is this - mink? He has taken the coat. He looks at Gillis standing there in tails. ARTIE Judas E. Priest, who did you borrow that from? Adolphe Menjou? GILLIS Close, but no cigar. Gillis stands embarrassed While Artie rolls up the Vicuna coat and tucks it above the books on a book- shelf. ARTIE Say, you're not really smuggling opium these days, are you? GILLIS Where's the bar? The two make their way toward the punch bowl. It's a little like running the gauntlet for Gillis. There are whistles and 'stares of astonishlnent at his tails. When they reach the punch bowl, Artie picks up a half-filled glass and fills it. GILLIS Good party. ARTIE The greatest. They call me the Elsa Maxwell of the assistant directors. (To some guests who are dipping their empty cups into the punch bowl) Hey, easy on the punch bowl. Budget only calls for three drinks per extra. Fake the rest. GILLIS Listen, Artie, can I stick around here for a while? ARTIE Sure, this'll go on all night. GILLIS I mean, could you put me up for a couple of weeks? ARTIE It just so happens we have a vacancy on the couch. GILLIS I'll take it. ARTIE I'll have the bell-hop take care of your luggage. He runs his finger across the decollete back of a girl standing in a group next them. ARTIE Just register here. The girl turns around. She is Betty Schaefer. BETTY Hello, Mr. Gillis. ARTIE You know each other? Gillis looks at her a little puzzled. BETTY Let me help you. Betty Schaeter, Sheldrake's office. GILLIS Sure. Bases Loaded. ARTIE Wait a minute. This is the woman I love. What's going on? Who was loaded? GILLIS Don't worry. She's just a fan for my literary output. BETTY (to Artie) Hurt feelings department. GILLIS About that luggage. Where's the phone? ARTIE Over by the Rainbow Room. Gillis squeezes his way through groups of people to the telephone, which is next to an open door leading to the bathroom. The phone is busy. A girl sits listening to it, giggling wildly. Another girl beside her is laughing too. They are apparently sharing a conversation with some man on the other end of the wire. The telephone passes from hand to hand. Gillis watches impatiently, then GILLIS When youlre through with that thing, can I have it? The girl just nods, going on with her chattering. Gillis stands waiting, and Betty Schaefer comes up with his glass. BETTY You forgot this. GILLIS Thanks. BETTY I've been hoping to run into you. GILLIS What for? To recover that knife you stuck in my back? BETTY I felt a little guilty, so I got out some of your old stories. GILLIS Why, you sweet kid. BETTY There's one called....Window... something with a window. GILLIS Dark Windows. How did you like it? BETTY I didn't. GILLIS Thank you. BETTY Except for about six pages. You've got a flashback there ... There is too much racket for her. BETTY Is there someplace we can talk? GILLIS How about the Rainbow Room? They squeeze their way towards the bathroom, past Artie. ARTIE I said you could have my couch. I didn't say you could have my girl. BETTY This is shop talk. She and Gillis go through the open door into C-20 ARTIE'S BATHROOM It's a little less noisy, although there are some guests there, chatting and having fun. Betty and Gillis sit down on the edge of the tub. GILLIS Now if I got you correctly, there was a short stretch of my fiction you found worthy of notice. BETTY The flashback in the courtroom, when she tells about being a school teacher. GILLIS I had a teacher like that once. BETTY Maybe that's why it's good. It's true, it's moving. Now why don't you use that character... GILLIS Who wants true? Who wants moving? BETTY Drop that attitude. Here's some- thing really worth while. GILLIS Want me to start right now? Maybe there's some paper around. BETTY I'm serious. I've got a few ideas. GILLIS I've got some ideas myself. One of them being this is New Year's Eve. How about living it up a little? BETTY As for instance? GILLIS Well.... BETTY We could make some paper boats and have a regatta. Or should we just turn on the shower? GILLIS How about capturing the kitchen and barricading the door? BETTY Are you hungry? GILLIS Hungry? After twelve years in the Burmese jungle. I am starving, Lady Agatha -- starving for a white shoulder -- BETTY Phillip, you're mad! One of the girls who was on the phone comes to the door. GIRL You can have the phone now. GILLIS (Paying no attention) Thirsting for the coolness of your lips - BETTY No, Phillip, no. We must be strong. You're still wearing the uniform of the Coldstream Guards! Furthermore, you can have the phone now. GILLIS O.K. (He gets up, starts out, turns) I find I'm terribly afraid of losing you. BETTY You won't. (She takes the glass out of his hand) I'll get us a refill of this awful stuff. GILLIS You'll be waiting for me? BETTY With a wildly beating heart. GILLIS Life can be beautiful! He leaves. C-21 THE MAIN ROOM Gillis squeezes himself through some guests to the phone. He has to stand in a cramped position, holding the instrument close to him as he dials a number. GILLIS Max? This is Mr. Gillis. I want you to do me a favor. C-22 NORMA DESMOND HOUSE Max is at the phone, in the lower hall. MAX I am sorry, Mr. Gillis. I cannot talk now. C-23 GILLIS ON THE PHONE GILLIS Yes you can. I want you to get my old suitcase and I want you to throw in my old clothes -- the ones I came with, and my typewriter. I'll have somebody pick them up. C-24 MAX AT THE PHONE MAX I have no time to talk. The doctor is here. C-25 GILLIS ON THE PHONE GILLIS What doctor? What's going on? C-26 MAX AT THE PHONE MAX She got the razor from your room. She cut her wrists. Max hangs up, moves toward the staircase. C-27 GILLIS AT THE PHONE GILLIS Max ! Max ! He hangs up the dead receiver, stands numb with shock. Betty elbows her way up to him, carrying the two punch glasses filled again. BETTY I just got the recipe: take two packages of cough drops, dissolve in one gallon of lukewarm grape juice -- Gillis looks up at her. Without a word he pushes her aside so that she spills the drink. He makes his way through the guests to the Vicuna coat, pulls it from the shelf, some books tumbling with it, and rushes towards the door and out. Betty stands look- ing after him, completely bewildered. DISSOLVE TO: C-28 EXT. DESMOND HOUSE - (NIGHT, RAIN) The doctor's car is parked in the driveway. A taxi pulls up. Gillis, in his Vicuna coat now, jumps out, throws a couple of dollars to the rdriver and runs toward the house. C-28a DOORWAY, NORMA DESMOND HOUSE> Max is opening the door to let out the doctor, a professional looking man carrying a black bag. Gillis runs into the SHOT. GILLIS How is she? MAX She is upstairs. Gillis starts to push past Max. Max grabs his arm. MAX Be careful. Do not race up the stairs. The musicians must not know what has happened. Gillis goes into the house. C-29 ENRANCE HALL AND STAIRCASE Gillis crosses the hall and starts up the stairs. C-3O INT. NORMA DESMOND'S ROOM Only one alabaster lamp lights the big, cold room. On the bed lies Norma in her evening dress. She is white as a sheet. Her wrists are bandaged. Her eyes are wide open, staring at the ceiling. One of her shoes has halt slipped off her foot. The other is on. Gillis opens the door and stands there tor a second. Then he slowly moves to the toot of the bed. He takes the shoes from her feet and puts them on the floor. NORMA Go away. GILLIS What kind of a silly thing was that to do? NORMA To fall in love with you -- that was the idiotic thing. GILLIS It sure would have made attractive headlines: Great Star Kills Her- self for Unknown Writer. NORMA Great stars have great pride. She puts one bandaged forearm over her eyes, sobbing. Gillis walks slowly over to the mantelpiece, stands there for awhile. NORMA Go away. Go to that girl of yours. GILLIS Look, I was making that up because I thought the whole thing was a mistake. I didn't want to hurt you. You've been good to me. You're the only person in this stinking town that has been good to me. NORMA Why don't you just say thank you and go, go, go -- GILLIS Not until you promise to act like a sensible human being. NORMA I'll do it again, I'll do it again, I'll do it again! Gillis stands looking at her helplessly. C-31 LIVING ROOM, THE DESMOND HOUSE The candles burned down, the orchestra playing to the emptiness. The orchestra leader looks at his watch, rises, silences the orchestra, then starts them in on Auld Lang Syne. C-32 INT. NORMA'S ROOM Gillis still stands. Norma lies on the bed, arms over her eyes, sobbing. GILLIS Happy New Year. Norma continues to sob. Gillis goes to the bed, puts his arms on her shoulders and turns her around. GILLIS Happy New Year. Norma looks at him, tears in her eyes. Slowly she enfolds him in her bandaged arms. NORMA Happy New Year. darling. She kisses him. DISSOLVE END OF SEQUENCE "C" SEQUENCE "D" DISSOLVE IN ON: D-1 INT. HALLWAY, NORMA GILLIS' VOICE DESMOND'S HOUSE (DAY) Around the middle of May some incidents happened The telephone is heard which I think I should tell ringing. Max comes from you about. living room to the phone, picks it up. MAX Hello ... Yes? D-1a BETTY SCHAEFER, AT THE PHONE ON HER DESK IN THE READERS' DEPARTMENT BETTY Is this Crestview 5-1733? ... I'm sorry to bother you again, but I've confirmed the number. I must speak to Mr. Gillis. D-1b MAX, AT THE PHONE MAX He is not here. D-1c BETTY ON THE PHONE BETTY Where can I reach him? Maybe somebody else in the house could tell me. D-1d MAX ON THE PHONE MAX Nobody here can give you any information. You will please not call again. He hangs up. From off comes: NORMA'S VOICE Who was it, Max? What is it? D-1e PATIO, NORMA'S HOUSE It is a sunny day. The garden is in somewhat better shape. The old house looks less unkept. The pool is filled. Norma sits on a wicker chaise longue, her face shielded by an enormous straw hat, her eyes by dark glasses. Gillis, in bathing trunks, is on a rubber mattress in the pool. Max comes to the entrance door. MAX Nothing, Madame. Somebody Inqu- iring about a stray dog. We must have a number very similar to the pound. He starts to turn back. NORMA Wait a minute. I want you to get out the car. You're going to take the script over to Paramount and deliver it to Mr. De Mille in person. MAX Yes, Madame. He goes into the house. GILLIS (climbing out of the water) You're really going to send it to De Mille? NORMA This is the right day. She indicates a typewritten letter she is holding. NORMA (Cont'd) The chart from my astrologer. She read deMille's horoscope. She read mine. GILLIS Did she read the script? NORMA DeMille is Leo. I'm Scorpio. Mars has been transmitting Jupiter for weeks. Today is the day of greatest conjuction. Now turn around. Let me dry you. She puts the towel around his sholders and starts drying him. GILLIS I hope you realize, Norma, that scripts don't sell on astrologers' charts. NORMA I'm not just selling the script. I'm selling me. DeMille always said I was his greatest star. GILLIS When did he say it, Norma? NORMA So he said it quite a few years ago. So what? I never looked better in my life. Do you know why? Because I've never been as happy in my life. She kisses him. DISSOLVE TO: D-2 INT. THE ISOTTA, DRIVING DOWN SUNSET ABOUT 8:30 IN THE EVENING GILLIS' VOICE A few evenings later we Max is driving. In the were going to the house of tonneau sit Norma, in a one of the waxworks for chinchilla wrap, and some bridge. She'd taught Gillis in his tuxedo. me how to play bridge by Norma is rummaging then, just as she'd taught through her evening me some fancy tango steps, bag. She finds a and what wine to drink cigarette case, opens with what fish. it. It is empty. NORMA That idiot. He forgot to fill my cigarette case. GILLIS (Proffering his case) Have one of mine. NORMA They're awful. They make me cough. GILLIS (Pushing open the glass partition, to Max) Pull up at the drugstore, will you, Max. (To Norma) I'll get you some. NORMA You're a darling. She takes a dollar bill from her purse and gives it to him. D-3 EXT. SCHWAB'S DRUGSTORE The car drives up and Gillis hurries into the store. D-4 INT. SCHWAB'S DRUGSTORE Business is still rather lively. There are about a dozen shoppers, and the soda counter is half filled. Gillis enters and steps to the tobacco counter. GILLIS (To the salesgirl) Give me a pack of those Turkish cigarettes -- Melachrinos. The girl opens the glass showcase to locate the fancy brand. From OFF comes ARTIE'S VOICE Stick 'em up, Gillis, or I'll let you have it! Gillis turns. D-5 AT THE SODA FOUNTAIN Artie Green and Betty Schaefer sit having a sandwich and a milk shake. With his forefinger and a sound effect, Artie riddles Gillis' body. Gillis walks INTO THE SHOT. GILLIS Hello, Artie. Good evening, Miss Schaefer. BETTY (Excitedly) You don't know how glad I am to see youl ARTIE Walking out on the mob. What's the big idea? GILLIS I'm sorry about New Year's. Would you believe me if I said I had to be with a sick friend? ARTIE Someone in the formal set, no doubt, with a ten-carat kidney stone. BETTY Stop it, Artie, will you? (To Gillis) Where have you been keeping your- self? I've got the most wonderful news for you. GILLIS I haven't been keeping myself at all. Not lately. BETTY I called your agent. I called the Screen Writers Guild. Finally your old apartment gave me some Crestview number. There was always somebody with an accent growling at me. You were not there. You were not to be spoken to. They never heard of you. GILLIS Is that so? What's the wonderful news? BETTY Sheldrake likes that angle about the teacher. GILLIS What teacher? BETTY Dark Windows. I got him all hopped up about it. GILLIS You did? BETTY He thinks it could be made into something. GILLIS Into what? A lampshade? BETTY Into something for Barbara Stan- wyck. They have a commitment with Barbara Stanwyck. ARTIE Unless you'd rather have Sarah Bernhardt. BETTY This is on the level. Sheldrake really went for it. GILLIS O.K. Where's the cash? BETTY Where's the story? I bluffed it out with a few notions of my own. It's really just a springboard. It needs work. GILLIS I was afraid of that. BETTY I've got twenty pages of notes. I've got a pretty good character for the man. ARTIE Could you write in plenty of back- ground action, so they'll need an extra assistant director? BETTY Shut up, Artie. (To Gillis) Now if we could sit down for two weeks and get a story. GILLIS Sorry, Miss Schaefer, but I've given up writing on spec. BETTY I tell you this is half sold. GILLIS As a matter of fact. I've given up writing altogether. Max has appeared in the door. MAX Mr. Gillis, if you please. GILLIS Right with you. Max leaves. ARTIE The accent! I've got it: this guy is in the pay of a foreign government. Get those studs. Get those cuff-links. GILLIS I've got to run along. Thanks any- way for your interest in my career. BETTY It's not your career -- it's mine. I kind of hoped to get in on this deal. I don't want to be a reader all my life. I want to write. GILLIS Sorry if I crossed you up. BETTY You sure have. GILLIS So long. He leaves. ARTIE (Patting her hand) Babe, it's like that producer says: In life, you've got to take the bitter with the sour. D-6 THE ISOTTA, PARKED OUTSIDE Gillis comes from Schwab's, gets into the car. Max takes off. NORMA What on earth, darling? It took you hours. GILLIS I ran into some people I knew. NORMA Where are my cigarettes? GILLIS Where are your...? He realizes he's forgotten them, takes the dollar and hands it back to her. GILLIS Norma, you're smoking too much. DISSOLVE TO: D-7 LIVING ROOM, NORMA DESMOND'S HOUSE (EARLY AFTERNOON) Start on a tiny GILLIS' VOICE parasol being Whenever she suspected I twirled...Norma was getting bored, she peeks out from one would put on a live show side of the parasol, for me: the Norma Desmond a bandanna tied Follies. Her first number around her head with was always the Mack Sennett a rabbit's-ear bow. Bathing Beauty. She bats her eyes, winks roguishly. THE CAMERA PULLS BACK to reveal that Norma's black pyjama trousers are rolled up over her knees and her black stockings rolled down below them. The whole effect approximates a Mack Sennett bathing costume pretty effectively. She points at a leather pour. NORMA This is a rock. She climbs on it, pantomimes timidity, an attempted dive, then jumps off. Gillis lolls on a couch, watching the performance, very bored. NORMA I can still see myself in the line: Bebe Daniels, Marie Prevost, Mabel Normand ... Mabel was always stepping on my feet ...What's the matter with you, darling? Why are you so glum? GILLIS (Lighting a cigarette with a match) Nothing is the matter. I'm having a great time. Show me some more. NORMA (Taking the match) All right. Give me this. I need it for a moustache. Now close your eyes. She runs out of the GILLIS' VOICE picture. Gillis has Something was the matter, closed his eyes. all right. I was thinking THE CAMERA MOVES to about that girl of Artie's, his face. that Miss Schaefer. She was so like all us writers when we first hit Holly- wood -- itching with am- bition, panting to get your names up there: Screenplay by. Original Story by. Hmph! Audiences don't know somebody sits down and writes a picture. They think the actors make it up as they go along. NORMA'S VOICE Open your eyes. Gillis opens his eyes. Norma has equipped herselr with a derby hat, a cane, and blacked in a small moustache. She goes into a little Chaplin routine. While she is doing it, the telephone rings. After a moment Max comes to the living room door. MAX Madame is wanted on the telephone. NORMA You know better than to interrupt me. MAX Paramount is calling. NORMA Who? MAX Paramount studios. NORMA (To Gillis) Now, now do you belive me? I told you deMille would jump at it. MAX It is not Mr. deMille in person. It is someone by the name or Gordon Cole. He says it's very important. NORMA Certainly it's important. It's important enough for Mr. deMille to call me personally. The idea of having an assistant call me! MAX I myself was surprised at Mr. de Mille's manners. NORMA Say that I'm busy, and hang up. MAX Very good, Madam. He bows and exits. NORMA How do you like that? We've made twelve pictures together. His greatest successes. GILLIS Maybe deMille is shooting. NORMA I know that trick! He wants to belittle me. He's trying to get my price down. I've waited twenty years for this call. Now Mr. deMille can wait till I'm good and ready. DISSOLVE TO: D-8 NORMA, IN THE TONNEAU OF THE LIMOUSINE, DRIVING DOWN MELROSE She is in full makeup, GILLIS' VOICE with a veil, a daring About three days later she hat, a suit so stunning was good and ready. In- only she would venture credible as it may seem, to wear it. THE CAMERA there had been some more PULLS BACK. Beside her of those calls from sits Gillis in the glen Paramount. So she put on plaid suit. Max is about half a pound of driving. makeup, fixed it up with a veil, and set forth to see deMille in person. Norma is examining her face in the mirror of her vanity. Max, while driving, sees her in the rear view mirror. MAX If you will pardon me, Madame. The shadow over the left eye is not quite balanced. NORMA Thank you, Max. With a handkerchief, she corrects it. D-9 MAIN GATE, EXT. PARAMOUNT STUDIO The car drives down Bronson and stops smack in front of the iron gate. A young policeman is talking to an extra; an old policeman sits reading a newspaper. Max sounds the horn impatiently. YOUNG POLICEMAN Hold that noise! MAX To see Mr. de Mille. Open the gate. YOUNG POLICEMAN Mr. deMille is shooting. You got an appointment? MAX No appointment is necessary. I am bringing Norma Desmond. YOUNG POLICEMAN Norma Who? Norma has rolled down the window on her side. She calls to the old policeman. NORMA Jonesy! Come here, Jonesy! OLD POLICEMAN Yeah? (He comes forward slowly) Why, if it isn't Miss Desmond! How have you been, Miss Desmond? NORMA Fine, Jonesy. Now open that gate. OLD POLICEMAN Sure, Miss Desmond. (To the young policeman} Come on, Mac. YOUNG POLICEMAN They can't drive on the lot without a pass. OLD POLICEMAN Miss Desmond can. Come on. They fling open the gate. OLD POLICEMAN (As the car drives through) Stage eighteen, Miss Desmond. NORMA Thank you, Jonesy. And teach your friend some manners. Tell him without me he wouldn't have any job, because without me there wouldn't be any Paramount Studio. (To Max) Go on. They drive through the gates. The old policeman goes to wall phone beside the gate, dials a number. OLD POLICEMAN (Into phone) Norma Desmond coming in to see Mr. deMille. D-10 STAGE 18 A scene from SAMPSON AND DELILAH is being rehearsed in the background. The usual turbulent activity surrounds it: extras. makeup men, grips, assistants, etc., etc. In the dim foreground a stage hand is answering a stand telephone. He puts down the phone and moves (CAMERA WITH HIM) to a second assistant. STAGE HAND Norma Desmond is coming to see Mr. deMille. The second assistant walks (CAMERA WITH HIM) to the first assistant. 2nd ASSISTANT Norma Desmond coming in to see Mr. deMille. The first assistant (CAMERA WITH HIM) hurries to the set. Sitting with his back toward us is C.B. himself. He is rehearsing a scene with Hedy Lamarr. 1ST ASSISTANT Norma Desmond is coming in to see you, Mr. deMille. C. B. turns his head. DEMILLE Norma Desmond? lst ASSISTANT She must be a million years old. DEMILLE I hate to think where that puts me. I could be her father. 1ST ASSISTANT I'm terribly sorry, Mr. de Mille. By this time de Mille is on his feet. DEMILLE It must be about that appalling script of hers. What can I say to her? What can I say? 1ST ASSISTANT I can tell her you're all tied up in the projection room. I can give her the brush ... DEMILLE Listen, thirty million fans have given her the brush. Isn't that enough? 1ST ASSISTANT I didn't mean to -- DEMILLE Of course you didn't. You didn't know Norma Desmond as a plucky little girl of seventeen, with more courage and wit and heart than ever came together in one youngster. 1ST ASSISTANT I hear she was a terror to work with. DEMILLE She got to be. A dozen press agents working overtime can do terrible things to the human spirit. (to the set) Hold everything. He leaves, accompanied by his entourage. D-11 EXT. STAGE 18 Norma's limousine drives up. Max dismounts and opens the door. NORMA (taking Gillis's hand) Don't you want to come along, darling? GILLIS I don't think so. It's your script. It's your show. Good luck. NORMA Thank you, darling. She presses his hand against her cheek, descends from the car and walks toward - D-12 THE DOOR OF STAGE 18 The first assistant is holding it open. In the door- way stands Mr. deMille. Seeing Norma, he stretches out his arms. DE MILLE Hello, young fellow. NORMA Hello, Mr. deMille. She has reached him. They embrace. NORMA Last time I saw you was someplace very gay. I remember waving to you. I was dancing on a table. DE MILLE Lots of people were. Lindbergh had just landed in Paris. Come on in. He leads her into D-13 STAGE 18 During the ensuing dialogue, Mr. deMille walks Norma towards the set. DE MILLE Norma, I want to apologize for not calling you. NORMA You'd better. I'm very angry. DE MILLE I'm pretty busy, as you can see... NORMA That's no excuse. You read the script, didn't you? DE MILLE Yes, I did. NORMA Then you could have picked up the phone yourself instead of leaving it to one of your assistants. DE MILLE What assistant? NORMA Don't play innocent. Somebody named Gordon Cole. DE MILLE Gordon Cole? NORMA And if you hadn't been pretty darned interested in that script, he wouldn't have tried to get me on the phone ten times. DE MILLE Gordon Cole... Look, Norma, I'm in the middle of a rehearsal. (Indicating his own chair) Make yourself comfortable. He walks onto the set, accompanied by his assistants. DE MILLE (Sotto voce, to his first assistant) Get me Gordon Cole on the phone. Meanwhile, Norma starts to sit, sees the name MISS LAMARR on the chair and with a look of distaste changes and sits on the one marked C.B. DE MILLE. From somewhere comes A VOICE Hey, Miss Desmond! Miss Desmond! She looks around her. VOICE Up here! Norma looks up at the scaffolding. On the scaffold stands one of the electricians, next to his light. ELECTRICIAN It's met It's Hog-eyel Norma waves at him. NORMA Hello. Hog-eye points his light at her. HOG-EYE Let's get a look at you. The beam of the lamp moves toward Norma. It hits her. She sits bathed in light. A couple of old costume extras recognize her. EXTRAS Say, it's Norma! Norma Desmond! They rush over and start wringing her hand. Into the shot comes a middle-aged hairdresser. HAIRDRESSER Hello, Miss Desmond. It's Bessie. Some elderly electricians and stagehands move in. D-14 ANOTHER PART OF THE STAGE The first assistant brings the portable phone to deMille. DeMille lifts the receiver. DE MILLE Hello. D-15 GORDON COLE'S OFFICE IN THE PROPERTY DEPARTMENT, GORDON COLE ON THE PHONE. COLE Prop Department. Gordon Cole speaking. D-16 DE MILLE ON THE PHONE DE MILLE Cole, this is C. B. deMille. Have you been calling Norma Desmond?... What's it about? D-17 GORDON COLE, ON THE PHONE COLE It's that car of hers -- an old Isotta-Fraschini. Her chauffeur drove it on the lot the other day. It looks just right for the Crosby picture. We want to rent it for a couple of weeks. D-18 DE MILLE ON THE PHONE DE MILLE (Troubled) Oh. Well, thank you. He hangs up, walks back towards Norma. (CAMERA WITH HIM). Norma stills sits in the shaft of light, surrounded by about a dozen people who have come up to pay court. DeMille gestures up to Hog-eye and the light shifts away. The people about Norma disperse slowly with various ad-libs. DE MILLE Well, Norma ... (He sits down next to her) I got hold of Gordon Cole. Norma hasn't heard a word. NORMA Did you see them? Did you see how they came? DE MILLE You know, crazy things happen in this business. I hope you haven't lost your sense of humor ... Suddenly he realizes that she is crying. She takes the handkerchief from his pocket and puts it over her eyes. DEMILLE What's the matter, Norma? NORMA Nothing. I just didn't realize what it would be like to come back to the old studio. I had no idea how I'd missed it. DEMILLE We've missed you too, dear. NORMA We'll be working again, won't we, Chief? We'll make our greatest picture. DEMILLE That's what I want to talk to you about. NORMA It's a good script, isn't it? DEMILLE It's got a lot of good things. Of course, it would be an expensive picture... NORMA I don't care about the money. I just want to work again. You don't know what it means to know that you want me. DEMILLE Nothing would thrill me more -- if it were possible. NORMA But remember, darling -- I don't work before ten in the morning, and never after 4:30 in the afternoon. The first assistant comes up. 1ST ASSISTANT We're ready with the shot, Mr. deMille. DEMILLE You'll pardon me, Norma? Why don't you just sit and watch? (He steps onto the set) O.K. Here we go. 1ST ASSISTANT Roll 'em. DEMILLE Action! The scene starts. D-19 THE ISOTTA, PARKED OUTSIDE STAGE 18 Max stands talking to Gillis, who is seated in the car. MAX (Pointing to the row of offices in the building opposite) You see those offices there, Mr. Gillis? They used to be her dressing room, The whole row. GILLIS That didn't leave much for Wallace Reid. MAX He had a great big bungalow on wheels. I had the upstairs. See where it says 'Readers' Department'? I remember my walls were covered with black patent leather... The words "Readers' Department" have registered on Gillis' mind. He gets out of the car. GILLIS I'll be with you in a minute. He crosses the street towards the green staircase leading to the second floor. Meanwhile, two prop men walking down the street come into the SHOT. 1ST PROP MAN Hey, that's the comic car Cole was talking about! (To Max) Do you mind if we look inside? MAX Go away. Go away. D-2O CUBICLE IN THE READERS' DEPARTMENT Behind the desk sits Betty, typing the synopsis of a novel, a half-eaten apple marking her place. The door behind her opens and Gillis enters. GILLIS Just so you don't think I'm a complete swine -- if there's anything in Dark Windows you can use, take it. It's all yours. BETTY Well, for heaven's sake! She moves the book and the apple aside and points at the free space on the desk. BETTY Have a chair. Gillis sits on the desk. GILLIS I mean it. It's no good to me anyway. Help yourself. BETTY Why should you do that? GILLIS If you get a hundred thousand for it, you buy me a box of chocolate creams. If you get an Oscar, I get the left foot. BETTY You know, I'd take you up on that in a minute. I'm just not good enough to do it all by myself. GILLIS What about all those ideas you had? BETTY See if they make sense. To begin with, I think you should throw out all that psychological stuff -- exploring a killer's sick mind. GILLIS Psychopaths sell like hotcakes. BETTY This story is about teachers -- their threadbare lives, their struggles. Here are people doing the most important job in the world, and they have to wprry about getting enough money to re-sole their shoes. To me it can be as exciting as any chase, any gunplay. GILLIS Check. BETTY Now I see her teaching day classes while he teaches night school. The first time they meet ... From below comes the SOUND of the Isotta's horn. GILLIS Look, if you don't mind, I haven't got time to listen to the whole plot ... BETTY I'll make it short. GILLIS Sorry. It's your baby now. BETTY I'm not good enough to write it alone. We'll have to do it together. GILLIS I'm all tied up. I can't. BETTY Couldn't we work in the evenings? Six o'clock in the morning? This next month I'm completely at your disposal. Artie is out of town. GILLIS What has Artie to do with it. BETTY We're engaged. GILLIS Good for you. You've got yourself the best guy in town. BETTY I think so. They're on location in Arizona, shooting a Western. I'm free every evening, every week- end. If you want, we could work at your place. GILLIS It's just impossible. BETTY Nobody can be that busy. There is another honk: from down below. GILLIS Look, Betty, It can't be done. It's out. BETTY You're tough, all right. GILLIS You're on your own. Stop being chicken-hearted and write that story. BETTY Honest to goodness, I hate you. GILLIS (Turning 1n the open door) And don't make it too dreary. How about this for a situation: she teaches daytimes. He teaches at night. Right? They don't even know each other, but they share the same room. It's cheaper that way. As a matter of fact, they sleep in the same bed -- in shifts, of oourse. BETTY Are you kidding? Because I think it's good. GILLIS So do I. BETTY Came on back. Let me show you where it fits in. She reaches in a drawer for her notes on Dark Windows. GILLIS (At the door) So long. Betty picks up the apple and is about to throw it after him. BETTY Oh, you -- GILLIS And here's a title: AN APPLE FOR THE TEACHER. He ducks out quiokly, slamming the door behind him. Betty looks after him, then angrlly hurls the apple into the wastebasket. D-21 STAIRCASE OUTSIDE READERS' DEPARTMENT Max is rush1ng up the stairs toward the descending Gillis. GILLIS What's the matter, Max? MAX I just found out why all those tele- phone calls. It is not Miss Desmond they want. It is the car they want to rent. GILLIS What? Max has seen something off. MAX Ssh... With his head he indicates D-22 ENTRANCE TO STAGE 18 The first assistant has opened the door. DeMille is showing Norma out. DE MILLE Goodbye, young fellow. We'll see what we can do. NORMA (embracing him) I'm not worried. Everything will be fine. The old team together. Nothing can stop us. She turns and walks out of the shot. De Mille stands for a second watching her, then turns to his assistant. DE MILLE Get Gordon Cole. Tell him to forget about her car. He can find another old car. I'll buy him five old cars, if necessary. 1ST ASSISTANT Yes, Mr. De Mille. They turn back into Stage 18. D-23 THE ISOTTA Gillis seated in the rear. Max is helping Norma in and putting the robe over her. GILLIS (Apprehensively) How did it go? NORMA It couldn't have gone better. It's practically set. Of course, he has to finish this picture first, but mine will be his next. There is an exchange of looks between Max and Gillis. GILLIS He must be quite a guy. NORMA He'a a shrewd old fox. He can smell box office. Only I'm going to outfox him a litt1e. This isn't going to be C. B. deMille's Salome. It's going to be Norma Desmond's Salome, a Norma Desmond Production, starring Norma Desmond...Home, Max. MAX Yes, Miss Desmond. As he says the words, he and Gillis exchange a glance in the rear view mirror. SLOW DISSOLVE: END OF SEQUENCE "D" SEQUENCE "E" DISSOLVE IN ON: E-1 CLOSEUP OF NORMA'S FACE GILLIS' VOICE Absolutely no makeup. A After that, an army of hand with a strong small beauty experts invaded flashlight comes into the her house on Sunset picture. The beam of the Boulevard. She went flashlight travels over the through a merciless face, exploring it merci- series of treatments, lessly. While the light is massages, sweat cabinets, still on it, two pairs of mud baths, ice compres- creamed hands come into the ses, electric devices. shot and start to massage it. She lived on vegetable juices and went to bed DISSOLVE TO: at nine. She was deter- mined to be ready -- ready for those cameras E-2 A SHORT MONTAGE of various that would never turn. beauty treatments applied to Norma. DISSOLVE TO: E-3 NORMA BEFORE THE MIRROR IN HER BEDROOM It is nine o'clock in the evening. She is in night gown and negligee and has put triangular patches on the saddle of her nose and at the outer corner of each eye. She is rubbing lotion on her hands. She gets up and crosses to the door of Gillis' room and opens it a crack. NORMA Joe darling, are you there? E-4 GILLIS' ROOM It is dark except for a lamp over the chaise longue. Gillis lies on it, fully clothed, reading a book. GILLIS Yes, Norma. Through the slit in the door there is a suggestion of Norma. NORMA Don't turn around. Keep your eyes on the book. GILLIS Yes, Norma. Norma pushes the door open and comes in. NORMA I just came to say good night. I don't want you to see me -- I'm not very attractive. GILLIS Good night. NORMA I've lost half a pound since Tuesday. GILLIS Good. NORMA I was a little worried about the line of my throat. This woman has done wonders with it. GILLIS Good. NORMA You'd better get to bed yourself. GILLIS I think I'll read a little. NORMA You went out last night, didn't you, Joe? GILLIS Why do you say that? NORMA I just happen to know it. I had a nightmare and I screamed for you. You weren't here. Where were you? GILLIS I went for a walk. NORMA No you didn't. You took the car. GILLIS All right, I drove to the beach. Norma, you don't want me to feel I'm locked up in this house? NORMA Of course not, Joe. It's just that I don't want to be left alone. Not now, while I'm under this terrible strain. My nerves are being torn apart. All I ask is for you to be a little patient and a little kind. GILLIS I haven't done anything, Norma. NORMA Of course you haven't. I wouldn't let you. She bends and kisses the top of his head. NORMA Good night, my darling. She goes into her room, shutting the door behind her. Gillis puts his book down and looks at her door. E-5 THE DOOR TO NORMA'S ROOM The light can be seen through the gouged-out keyhole. It goes out. DISSOLVE TO: E-6 UPPER LANDING STAIRWAY AND HALL BELOW (NIGHT) GILLIS' VOICE Gillis, with his coat on by Yes, I was playing hooky now, comes cautiously to the upper railing and looks every evening along in down into the lighted hall below. there. It made me think I Max is just extinguishing of when I was twelve and the lights. Max exits in, the direction of the liv- used to sneak out on the ing room. folks to see a gangster After a moment Gillis starts silently down the stairs. picture. This time it wasn't to see a picture, E-7 LIVING ROOM it was to try and write (Lighted only by the last flicker of a fire on the one. That story of mine hearth). Max is putting a fire screen in front of Betty Schaerer had dug the fire. He hears some steps and the creak or the up kept going through main door being opened. He looks out and sees my head like a dozen locomotives... E-7a THE MAIN DOOR Gillis, in the moonlit porch, is closing the main door behind him. E-8 LIVING ROOM Max looks after Gillis, his face enigmatic as ever. DISSOLVE TO: E-9 GARAGE AND DRIVEWAY (MOONLIGHT) Gillis comes into the shot, gets into the Isotta, drives it out or the garage and down the driveway to Sunset, as quietly as possible. DISSOLVE TO: E-10 READERS' OFFICE BUILDING PARAMOUNT (NIGHT) Start on a LONG SHOT. THE GILLIS' VOICE BOOM MOVES FORWARD to the only So we'd started two lights. They are the door working on it, the and window of Betty Schaefer's two of us. Nights, cubicle. Betty sits at the when the studio was desk, typing. Gillis, his deserted, up in her coat off, his shirt-sleeves little cubby-hole rolled up, j.s pacing the floor, of an office. discussing the construction of a sentence. The discussion at a stalemate, Betty suggests some coffee. Gillis agrees. From the electric plate on the shelf beside her, Betty takes a glass coffee machine. Gillis seats himself in her chair and starts typing. Betty opens the door and comes out on the balcony to fill the coffee machine from the water cooler stand- ing beside the door. BETTY I got the funniest letter from Artie. It's rained every day since they got to Arizona. They re-wrote the whole picture for rain and shot half of it. Now the sun is out. Nobody knows when they'll get back. She moves back into the room. GILLIS Good. BETTY What's good about it? I miss him something fierce. GILLIS I mean this is good dialogue along in here. It'll play. BETTY It will? GILLIS Sure. Especially with lots of music underneath, drowning it out. BETTY Don't you sometimes hate yourself? GILLIS Constantly. No, in all serious- ness, it's really good. It's fun writing again. I'm happy here, honest I am. He resumes typing. Betty puts the water on. She picks up a pack of cigarettes on the desk, finds it's empty and throws it away, sees Gillis' open gold cigarette case and lighter on the table by the couch. Betty reaches for a cigarette. The inscription en- graved inside the case catches her eye. It reads: MAD ABOUT THE BOY -- Norma BETTY Who's Norma? GILLIS Who's who? BETTY I'm sorry. I don't usually read private cigarette cases. GILLIS Oh, that. It's from a friend of mine. A middle-aged lady, very foolish and very generous. BETTY I'll say. This is solid gold. GILLIS I gave her some advice on an idiotic script. BETTY It's that old familiar story, you help a timid little soul across a crowded street. She turns out to be a multimillionaire and leaves you all her money. GILLIS That's the trouble with you readers. You know all the plots. Now suppose you proof-read page ten while the water boils. DISSILVE TO: E-11 AN EMPTY STREET AT THE GILLIS' VOICE PARAMOUNT STUDIO (NIGHT) Sometimes when we got stuck we'd make a Gillis and Betty are walking litte tour of the down it. From a stage where drowsing lot, not talk- they are erecting a new set ing much, just wandering comes a great shaft of light. down alleys between the They stop at an apple-vending sound stages, or through machine in the foreground,buy the sets they were get- themselves a couple of apples ting ready for the next and walk on. day's shooting. As a matter of fact, it was DISSOLVE TO: on one of those walks when she first told me about her nose ... E-12 PARAMOUNT'S NEW YORK STREET (NIGHT) Betty and Gillis are walking down it, THE CAMERA AHEAD OF THEM. BETTY Look at this street. All card- board, all hollow, all phoney. All done with mirrors. I like it better than any street in the world. Maybe because I used to play here when I was a kid. GILLIS What were you -- a child actress? BETTY I was born just two blocks from this studio. Right on Lemon Grove Avenue. Father was head elec- trician here till he died. Mother still works in Wardrobe. GILLIS Second generation, huh? BETTY Third. Grandma did stunt work for Pearl White. I come from a picture family. Naturally they took it for granted I was to become a great star. So I had ten years of dramatic lessons, diction, dancing. Then the studio made a test. Well, they didn't like my nose -- it slanted this way a little. I went to a doctor and had it fixed. They made more tests, and they were crazy about my nose -- only they didn't like my acting. GILLIS (Examining her nose by the flame of his lighter) Nice job. BETTY Should be. It cost three hundred dollars. GILLIS Saddest thing I ever heard. BETTY Not at all. It taught me a little sense. I got me a job in the mail room, worked up to the Stenographic. Now I'm a reader... GILLIS Come clean, Betty. At night you weep for those lost closeups, those gala openings... BETTY Not once. What's wrong with being on the other side of the cameras? It's really more fun. GILLIS Three cheers for Betty Schaefer! I will now kiss that nose of yours. BETTY If you please. Gillis kisses her nose. As he stands there, his face close to hers - GILLIS May I say you smell real special. BETTY It must be my new shampoo. GILLIS That's no shampoo. It'smore like a pile of freehly laundred hand- kerchiefs, like a brand new auto- mobile. How old are you anyway? BETTY Twenty-two. GILLIS That's it -- there's nothing like being twenty-two. Now may I suggest that if we're ever to finish this story you keep at least two feet away from me ... Now back to the typewriter. They start walking in the direction of the office. DISSOLVE TO: E-13 THE GARAGE Gillis gets out. From the seat next him he takes a batch of script, folds it and puts it in his pocket. He suddenly becomes aware that he is watched, turns. Max stands in the moonlight, evidently waiting for him. GILLIS What is it, Max? Want to wash the car, or are you doing a little spying in your off hours? MAX You must be very careful as you cross the patio. Madame may be watching. GILLIS How about my going up the kitchen stairs and undressing in the dark. Will that do it? MAX I'm not inquiring where Mr. Gillis goes every night... GILLIS Why don't you? I'm writing a script and I'm dying to finish it, no matter what. MAX It's just that I'm very worried about Madame. GILLIS Sure you are. And we're not help- ing her any, feeding her lies and more lies. Getting herself ready for a pioture ... What happens when she finds out? MAX She never will. That is my job. It has been for a long time. You must understand I discovered her when she was eighteen. I made her a star. I cannot let her be destroyed. GILLIS You made her a star? MAX I directed all her early pictures. There were three young directors who showed promise in those days: D.W. Grirrith, C.B. deMille, and Max von Mayerling. GILLIS And she's turned you into a servant. MAX It was I who asked to come back, humiliating as it may seem. I could have gone on witn my career, only I found everything unendur- able arter she divorced me. You see, I was her rirst husband. DISSOLVE TO: E-14 NORMA DESMOND'S BEDROOM One lamp lit. Norma, in a white negligee, with the patches on her face, is pacing up and down -- a small, tormented, pitiable woman. Finally she opens the door to: E-15 GILLIS' ROOM (MOONLIGHT) Gillis lies in bed asleep, Norma in the doorway. NORMA You're here, Joe ... When did you come home? Where were you? Is it a woman? I know it's a woman ... Who is she? Oh Joe, why can't I ask you? I must know, I must! Her eyes fall on Gillis' coat, which hangs over a chair. In a pocket is part of the script. Norma takes it out, looks at it. She can't see it in the moonlight. She hurries with it into: E-16 NORMA'S BEDROOM Carrying the script Norma goes to the lamp and looks at it. On the first page she sees something which confirms all her suspicionso It reads: UNTITLED LOVE STORY by Joseph C. Gilliss and Betty Schaefer DISSOLVE: E-17 BETTY'S CUBICLE (NIGHT) Betty is typing. Gillis sits on the couch, proof- reading a scene. Betty stops typing and Gillis becomes aware of her eyes fixed on him. GILLIS Hey, what's the matter... Betty, wake up! (He whistles and catches her attention) Why are you staring at me like that? BETTY Was I? I'm sorry. GILLIS What's wrong with you tonight? What is it, Betty? BETTY Something came up. I don't want to talk about it. GILLIS Why not? BETTY I just don't. GILLIS What is it you've heard. Come on, let's have it. Betty gets up. GILLIS Is it about me? Betty doesn't answer, walks out on E-18 THE BALCONY She leans against a post, crying. Gillis comes out after her. GILLIS Betty, there's no use running out on it. Let's face it, what- ever it is. BETTY It's nothing. I got a telegram from Artie. GILLIS From Artie. What's wrong? BETTY He wants me to come on to Arizona. He says it only oosts two dollars to get married there. It would kind of save us a honeymoon. GILLIS Why don't you? We can finish the script by Thursday. Betty stands crying silently. GILLIS Stop crying. You're getting married. That's what you've always wanted. BETTY I don't want it now. GILLIS Why not? Don't you love Artie? BETTY Of course I love him. I always will. I'm just not in love with him any more. GILLIS What happened? BETTY You did. There is a moment's pause before he takes her in his arms. THE CAMERA MOVES AWAY. DISSOLVE TO: E-19 HALL AND STAIRCASE GILLIS' VOICE DESMOND HOME- (NIGHT) It wasn' t until I got back to that peculiar Gillis enters, closes prison of mine that I the door as quietly as started facing the facts. he can, and goes up There it was -- Betty the stairs. Schaefer's future right in the palm of my hand. E-20 GILLIS' ROOM Betty Schaefer engaged to Artie Green, as nice He enters and turns on the a guy as ever lived. light. He sinks down on And she was in love with the chaise longue,thinking. me. Me ! She was a fool His eyes wander to the not to sense that there door of Norma's room. was something phony in Through the gouged-out key- my set-up. And I was a hole he sees the light. heel not to have told her. But you just can't say those things to somebody you're crazy about. Maybe I'd never have to. Maybe I could get away with it, get away from Norma. Maybe I could wipe the whole nasty mess right out of my life... From Norma's room comes the sound of a telephone being dialled. Gillis enters the shot and stands listening. NORMA'S VOICE Is this Gladstone 0858? E-21 NORMA'S BEDROOM Norma lies in bed, dialing a number. She has the beauty patches at the corners of her eyes and over her nose. NORMA Can I speak to Miss Betty Schaefer? She must be home by now. E-22 A BEDROOM IN BETTY'S FLAT Connie, a girl of Betty's age with whom she shares the flat, is on the phone. Betty, in a dressing- gown, comes from the bathroom, toothbrush in hand. CONNIE (Hand over mouthpiece) Betty, here's that weird-sounding woman again. BETTY What is this anyway? (Taking the phone) This is Betty Schaefer. E-23 NORMA AT IHE PHONE NORMA Miss Schaefer, you must forgive me for calling you so late, but I really feel it's my duty. It's about Mr. Gillis. You do know Mr. Gillis? ...Exactly how much do you know about him? Do you know where he lives? Do you know how he lives? Do you know what he lives on? E-24 BETTY AT THE PHONE BETTY Who are you? What do you want? What business is it of yours anyway? E-25 NORMA ON THE PHONE NORMA Miss Schaefer, I'm trying to do you a favor. I'm trying to spare you a great deal of misery. Of course you may be too young to even suspect there are men of his sort... NORMA (Cont'd) I don't know what he's told you, but he does not live with relatives, nor with friends, in the usual sense of the word. Ask him ... Ask him again. During the latter part of her call, the doors from Gillis' room have been pushed open and Gillis has walked towards her. Suddenly Norma senses his pre- sence and turns around. The telephone freezes in her hand. She tries to hang it up. Very calmly Gillis takes the receiver from her hand. GILLIS (Into phone) That's right, Betty, ask me again. This is Joe. E-26 BETTY ON THE PHONE BETTY Joe, where are you? What's this all about? E-27 GILLIS ON THE PHONE Norma beside him. GILLIS Or maybe it would be a better idea if you came over and saw it for yourself. The address is 10086 . He hangs up. Norma looks up at him as he crosses to the other end of the room and stands staring at her. The silence becomes unbearable. NORMA Don't hate me, Joe. I did it because I need you. I need you as I never needed you. Look at me. Look at my hands, look at my face, look under my eyes. How can I go back to work if I'm wasting away under this torment? You don't know what I've been through these last weeks. I got myself a revolver. You don't believe me, but I did, I did! I stood in front of that mirror, only I couldn't make myself. It wouldn't be NORMA (Cont'd) fair to all those people who are waiting to see me back on the screen. I can't disappoint them. Only, if I'm to work, I need sleep, I need quiet, I need you! Don't just stand there hating me! Shout at me, strike me! But don't hate me, Joe. Don't you hear me, Joe? GILLIS Yes, I hear you. And I wish you'd keep still so I can hear the doorbell when she rings it. E-28 BETTY AND CONNIE, DRIVING IN A SMALL COUPE DOWN (NIGHT) E-29 INT. COUPE Connie is looking at the house numbers. CONNIE Here's ten thousand seventy-nine, Betty. It must be over there. Betty turns the car into the driveway of Norma's place, stops at the entrance steps. Betty gets out. CONNIE Betty, let me come along with you. Please. BETTY No, I'll be all right. She shuts the door of the car and goes up the steps. E-30 NORMA'S BEDROOM Norma lies on the bed. Gillis sits in a far corner of the room, motionless. NORMA (In a whimpering monotone) I love you, Joe. I love you, Joe. I love you, Joe. I love you, Joe. There is the sound of footsteps below and the ringing of a doorbell. Gillis rises. NORMA What are you going to do, Joe? Without a word, he leaves the room. Norma raises herself on the bed, reaching for a black negligee lying at the foot of it. As she does so, she dis- lodges her pillow a little, revealing a revolver hidden beneath it. E-31 DOWNSTAIRS HALL, THE DESMOND HOUSE (DARK) Max crosses the hall, putting on his alpaca jacket. He turns on the lights. Outside stands Betty. From the staircase comes - GILLIS' VOICE It's all right, Max. I'll take it. MAX Yes, sir. He stands back as Gillis opens the door. GILLIS Hello, Betty. BETTY (On the threshold) I don't know why I'm so scared, Joe. Is it something awful? GILLIS Come on in, Betty, Betty enters. As he leads her into the living room, Gillis puts his arm around her shoulders. GILLIS Ever been in one of these old Hollywood palazzos? That's from when they were making eighteen thou- sand a week, and no taxes. Careful of these tiles, they're slippery. Valentino used to dance here. BETTY This is where you live? GILLIS You bet. BETTY Whose house is it? They have reached E-32 THE LIVING ROOM Gillis leads Betty in. GILLIS Hers. BETTY Whose? GILLIS Just look around. There's a lot of her spread about. If you don't remember the face, you must have heard the name of Norma Desmond. BETTY That was Norma Desmond on the phone? GILLIS Want something to drink? There's always champagne on ice, and plenty of caviar. BETTY Why did she call me? GILLIS Jealous. Ever see so much junk? She had the ceiling brought from Portugal. Look at this. He pulls the rope, showing the projection screen under the picture. GILLIS Her own movie theatre. BETTY I didn't come here to see a house. What about Norma Desmond? GILLIS I'm trying to tell you. This is an enormous place. Eight master bedrooms. A sunken tub in every bathroom. There's a bowling alley in the cellar. It's lonely here, so she got herself a companion. A very simple set-up: An older woman who is well-to-do. A younger man who is not doing too well ... Can you figure it out yourself? BETTY No. GILLIS All right. I'll give you a few more clues. BETTY No, no! I haven't heard any of this. I never got those telephone calls. I've never been in this house ... Get your things together. Let's get out of here. GILLIS All my things? All the eighteen suits, all the custom-made shoes and the eighteen dozen shirts, and the cuff-links and the platinum key- chains, and the cigarette cases? BETTY Come on, Joe. GILLIS Come on where? Back to a one-room apartment that I can't pay for? Back to a story that may sell and very possibly will not? BETTY If you love me, Joe. GILLIS Look, sweetie -- be practical. l've got a good thing here. A long-term contract with no options. I like it that way. Maybe it's not very admirable. Well, you and Artie can be admirable. BETTY Joe, I can't look at you any more. GILLIS Nobody asked you to. Betty turns from him, to hide the fact that she is crying. GILLIS All right, baby. This way out. He leads her in the direction of the door. E-33 UPPER LANDING, DESMOND HOUSE Sitting crouched behind the balustrade is Norma, peering down into E-34 THE LOWER HALL Betty and Gillis have reached the entrance door. Gillis opens it. GILLIS Good luck to you, Betty. You can finish that story on the way to Arizona. When you and Artie get back, if the two of you ever feel like a swim, here's the pool ... He switches on the light. E-35 THE PATIO The lights go on in the pool, which shines brilliant- ly in the dark garden. E-36 BETTY She doesn't even look. Her eyes filled with tears, she runs down the entrance porch toward her car. E-37 THE ENTRANCE HALL Gillis looks after her, closes the door. From the upper landing comes the sound of soft sobbing. He looks up. E-38 NORMA, ON THE UPPER LANDING Gillis ascends the stairs. NORMA Thank you, Joe -- thank you, Joe. She tries to take his hand to kiss it as he passes. He doesn't stop. Norma catches his coat. Gillis moves right on into his room. Norma lies on the floor looking after him. She crawls toward a con- sole, pulls herself up by it, starts towards Gillis' door, passes a mirror, realizes how she looks, moves back to the mirror and takes the patches off her face and does a hasty job of removing the cream with her handkerchief, readjusts her expression to a poor travesty of a smile and goes to the door of Gillis' room. NORMA May I come in? I've stopped cry- ing. I'm all right again. Joe, tell me you're not cross -- tell me everything is just as it was, Joe. She opens the door. E-39 GILLIS' ROOM In the foreground, open on the bed, is a half-packed suitcase, Gillis just putting some of his old shirts in. Norma stands staring, speechless, for a second. Gillis moves out of the shot towards the closets. NORMA What are you doing, Joe? What are you doing? You're not leaving me? GILLIS Yes, I am, Norma. NORMA No, you're not. (Calling) Max! Max! GILLIS Max is a good idea. He can help with my luggage. (He gestures in the direction of the closet) Thanks for letting me wear the handsome wardrobe. And thanks for the use of all the trinkets. He takes the cigarette case and throws it on the chaise longue. Then he throws the lighter, the wrist watch, the platinum key-chain and the tie clip. GILLIS (Indicating the bureau) The rest of the jewelry is in the top drawer. NORMA It's yours, Joe. I gave it to you. GILLIS And I'd take it in a second, Norma -- only it's a little too dressy for sitting behind the copy desk in Dayton, Ohio. NORMA These are nothing. You can have anything you want if you'll only stay. What is it you want -- money? GILLIS Norma, you'd be throwing it away. I don't qualify for the job, not any more. NORMA You can't do this! Max! Max! ... I can't face life without you, and I'm not afraid to die, you know. GILLIS That's between you and yourself, Norma. NORMA You think I made that up about the gun... She rushes into her room. Gillis closes the suitcase calmly, notices that he is still wearing some cuff- links Norma gave him, takes them off. Norma reappears in the door, carrying the revolver. NORMA See, you didn't believe me!.. Now I suppose you don't think I have the courage! GILLIS Oh. sure -- if it would make a good scene. NORMA You don't care. do you? But hundreds of thousands of people will carel GILLIS Wake up, Norma. You'd be killing yourself to an empty house. The audience left twenty years ago. Now face it. During the preceding. Max has entered. He stands listening, paralyzed. NORMA That's a lie! They still want me! GILLIS No, they don't. NORMA What about the studio? What about De Mille? GILLIS He was trying to spare your feelings. The studio wanted to rent your car. NORMA Wanted what? GILLIS De Mille didn't have the heart to tell you. None of us has had the heart. NORMA That's a lie! They want me, they want me! I get letters every day! GILLIS You tell her, Max. Come on, do her that favor. Tell her there isn't going to be any picture -- there aren't any fan letters, except the ones you write yourself. NORMA That isn't true! Max? MAX Madame is the greatest star of them all... I will take Mr. Gillis' bags. He leaves. NORMA You heard him. I'm a star! GILLIS Norma, grow up. You're a woman of fifty. There's nothing tragic about being fifty - not unless you try to be twenty-five. NORMA I'm the greatest star of them all. GILLIS Goodbye. Norma. NORMA No one leaves a star. That makes one a star. Gillis picks up the typewriter and leaves. NORMA You're not leaving me! E-40 STAIRCASE Gillis descending with the typewriter. NORMA'S VOICE Joe! ...Joe! There is the SOUND OF A SHOT. The glass of the front door is shattered. Gillis at the door opens it and walks out, without looking back. Down the staircase rushes Norma. a disordered wild- ness in the way she moves. NORMA You're not leaving me! She hurries after Gillis. E-41 PATIO (NIGHT) Dark except for lights from the house and the luminousness of the lit pool. Gillis is crossing the patio towards the garage. He is carrying the typewriter. He doesn't accelerate his step, although he has heard the shot. Behind him Norma comes from the lighted house. NORMA You're not leaving me! She shoots twice in rapid succession. Gillis drops the typewriter. The shots have swung him around. He is now facing Norma. She shoots him. This shot hits him in the belly. He doubles up, instinctively backs away from her, plummets into the lit pool. Up the stone steps from the garage rushes Max. He sees the situation, hurries towards Norma, who stands exultant in the strange light from the pool. NORMA Stars are ageless, aren't they? DISSOLVE TO: E-42 THE PATIO Dawn is breaking. At the edge of the pool stand policemen, detectives and police photographers. Motorcycle policemen are holding off the mob which is trying to storm the house. A lietuenant from the Homicide Bureau leaves the crowd around the pool and goes into E-43 THE LOWER HALL, DESMOND HOUSE It is filled with a pandemonium of police officers, newspaper people, etc. who are kept from the upper floor by two policemen at the head of the stairs. The lieutenant from the Homicide Bureau goes through the crowd to the telephone at the foot of the stairs, picks up the phone and dials. LIEUTENANT Coroner's office? ... I want to speak to the Coroner ... Who's on this phone? E-44 THE WHITE TELEPHONE IN NORMA'S BEDROOM Standing talking into it is Hedda Hopper. MISS HOPPER I am! Now get off, this is more important ... Times City Desk? Hedda Hopper speaking. I'm talking from the bedroom of Norma Desmond. Don't bother with a rewrite man, take this direct. Ready? -- As day breaks over the murder house, Norma Desmond, famed star of yesteryear, is in a state of complete mental shock ... THE CAMERA PANS TO ANOTHER PART OF THE BEDROOM, where Norma sits at a mirror, staring at herself blankly. Firing questions at her are the Captain of the Holmby Hills Division and the L.A. Homicide Squad. Max stands by faithfully. HOLMBY HILLS CAPTAIN You do not deny having killed this man, Miss Desmond? HEAD OF HOMICIDE Did you intend to kill him? Just answer me that. HOLMBY HILLS CAPTAIN Was it a sudden quarrel? Had there been any trouble between you before? HEAD OF HOMICIDE If it was a quarrel, how come you had the gun right there? HOLMBY HILLS CAPTAIN This guy -- where did you meet him for the first time? Where did he come from? Who is he? HEAD OF HOMICIDE Did he have a wife? Did he had a girl friend? Did you know them? HOLMBY HILLS CAPTAIN Had he been trying to blackmail you? E-45 PATIO - (DAWN) GILLIS' VOICE The body of Gillis Well, this is where you came. being fished from Here's that pool again,the one the pool, put on a I always wanted. They must have stretcher, covered photographed me a hundred times. with an army blanket.Then they got a couple of prun- Two men from the ing hooks from the garden and Coroner's office fished me out ever so gently. carry it towards Funny how gentle people get with the Coroner's you once you're dead. They hearse, CAMERA beached me, like a harpooned PANNING with them. baby whale, and started to check the damage, just for the record ... By this time the whole joint was jumping -- cops,reporters, neighbors, passersby -- as much hoopdedoo as we get in Los Angeles when they open a Super Market. Even the newsreel guys came roaring in. Here was an item everybody could have some fun with, the heartless so-and- so's. What would they do to her? Even if she got away with it in court- crime of passion - tempo- rary insanity - those headlines would kill her: Forgotten Star a Slayer--Aging Actress-- Yesterday's Glamour Queen... E-46 NORMA'S BEDROOM The interrogators are still firing questions at Norma who sits lifeless, staring at herself. Max watches. HEAD OF HOMICIDE Did the deceased ever threaten you? Were you in fear of bodily injury? HOLMBY HILLS CAPTAIN Did you hate him? Had you ever thought of doing something like this before? HEAD OF HOMICIDE Was theft involved? Did you catch him trying to steal something, or find he had stolen something? A police lieutenant has entered, goes to the Head of Homicide. LIEUTENANT The newsreel guys have arrived with the cameras. HEAD OF HOMICIDE Tell them to go fly a kite. This is no time for cameras. A word has pierced the mists that surround Norma. NORMA Cameras? ...What is it, Max? MAX The cameras have arrived, Madame. NORMA They have? Thank you, Max. Tell Mr. DeMille I will be on the set at once. Max flashes a look at the Head of Homicide. HEAD OF HOMICIDE What is this? MAX Please ... HOLMBY HILLS CAPTAIN (sotto voce, to Head of Homicide) Well, it's one way to get her down stairs. HEAD OF HOMICIDE Okay. And let's have the car right outside. 7-1 NORMA You will pardon me, gentlemen. I have to get ready for my scene. She takes a comb and runs it through her hair, then starts applying some wild makeup. E-47 STAIRCASE AND LOWER HALL Max makes his way down the stairs through the crowd of newsmen to the newsreel cameras, which are being set up in the hall below. MAX Is everything set up, gentlemen? Are the lights ready? From the stairway comes a murnur. They look up. Norma has emerged from the bedroom and comes to the head of the stairs. There are golden spangles in her hair and in her hand she carries a golden scarf. The police clear a path for her to descend. Press cameras flash at her every step. Max stands at the cameras. MAX Is everything set up, gentlemen? CAMERAMAN Just about. The portable lights flare up and illuminate the staircase. MAX Are the lights ready? 2ND CAMERA MAN All set. MAX Quiet, everybody! Lights! Are you ready, Norma? NORMA (From the top of the stairs) What is the scene? Where am I? MAX This is the staircase of the palace. NORMA Oh, yes, yes. They're below, waiting for the Princess ... I'm ready. MAX All right. (To cameramen) Camera! (To Norma) Action! Norma arranges the golden GILLIS' VOICE scarf ebout her and proudy So they were grinding starts to descend the stair- after all, those cam- case. The cameras grind. eras. Life, which can Everyone watches in awe. be strangely merciful, had taken pity on Norma Desmond. The dream she had clung to so des- perately had enfolded her... At the foot of the stairs Norma stops, moved. NORMA I can't go on with the scene. I'm too happy. Do you mind, Mr. DeMille, if I say a few words? Thank you. I just want to tell you how happy I am to be back in the studio making a picture again. You don't know how much I've missed all of you. And I promise you I'll never desert you again, because after "Salome" we'll make another picture, and another and another. You see, this is my life. It always will be. There's nothing else - just us and the cameras and those wonderful people out there in the dark... All right, Mr. DeMille, I'm ready for my closeup. FADE OUT. THE END | 1 | 5.3% |
| STRANGERS ON A TRAIN by Raymond Chandler and Czenzi Ormonde FINAL DRAFT October 18, 1950 Converted to PDF by SCREENTALK FOR EDUCATIONAL PURPOSES ONLY www.screentalk.org FADE IN: EXT. UNION STATION, WASHINGTON, D.C. DAY LONG SHOT THE CAPITOL DOME IN THE B.G. AND THE AUTOMOBILE ENTRANCE TO THE STATION IN THE F.G. LOW CAMERA Activity of cars and taxis arriving and discharging passengers with luggage, busy redcaps, etcetera. We FOCUS on a taxi pulling up and stopping, The driver hands out modest looking luggage, including a bunch of tennis rackets in cases to a redcap. CAMERA PANS DOWN as the passenger gets out of the taxi so that we see only his shoes and the lower part of his trousers. He is wearing dark colored brogues and a conservative suit apparently. The feet move toward, the entrance to the station and out of scene. Immediately a chauffeur-driven limousine drives up and an expensive place of airplane luggage is handed out of this, and the passenger alighting from the back is seen to be wearing black and white sport shoes which, as before, are all we see of him. The sport shoes start off in the wake of the brogues. INT. STATION LOBBY CAMERA FOLLOWS the sport shoes and the brogues across the lobby into a passenger tunnel. There is the usual activity of passengers walking to and from, a loud-speaker announcing trains, etc. EXT. PASSENGER TUNNEL As the brogues and the sport shoes emerge to the train platform, CAMERA PANS them over to the steps of the train. INT. TRAIN The brogues and the sport shoes pass separately down the aisle, the sport shoes turning in at a compartment door and the brogues continuing toward the parlor car. DISSOLVE TO: INT. PARLOR CAR (PROCESS) The brogues come to rest before a chair as the owner sits down. A moment later the sport shoes come to rest. before in adjoining chair. Converted to PDF by www.screentalk.org 2. The legs belonging to the sport shoes stretch out, and one of the shoes touches one of the brogues. MAN'S VOICE (over scene) Oh, excuse Me! CAMERA PULLS BACK AND UP to SHOW two young men seated in two parlor car chairs. BRUN0 ANTHONY, the wearer of the sport shoes, is about twenty-five. He wears his expensive clothes with the tweedy nonchalance of a young man who has always had the best. The wearer of the brogues is a fine looking but, at the moment, a somewhat troubled young man. This is GUY HAINES. He, too, is in his middle twenties and is well dressed because he can now afford to be. He nods politely, acknowledging Bruno's apology, then turns away with the gesture implying he wants privacy. BRUNO (smiling with sudden recognition) I beg your pardon, but aren't you Guy Haines. Guy nods with a polite half smile. Being a well known tournament tennis player, he has had this sort of experience before. BRUNO (snapping his finger) Sure! I saw you blast Faraday right off the court in South Orange last season. What a backhand! Made the semi-finals, didn't you? Guy acknowledges this with a modest nod and turns to his magazine rolled up in is fist. BRUNO (with open admiration) I certainly admire people who do things. (smiling and introducing himself) I'm Bruno Anthony. Bruno. See Guy looks up. Bruno indicates his gold tie pin which bears his name in cut- out letters. Guy looks at it with the faintest expression of disdain. I suppose you think it's corny. But my mother gave it to me so of course I wear it to please her. Converted to PDF by www.screentalk.org 3. GUY (patiently)(a faint smile) How do you do. BRUNO (with an apologetic grin) I don't usually talk so much. Go Ahead and read. GUY (wryly) Thanks. Guy tries to read but is uneasily aware of Bruno's open appraisal. BRUNO It must be pretty exciting to be so important. GUY (fidgeting slightly) A tennis player isn't so important. BRUNO People who do things are important. I never seem to do anything. Not knowing how to answer this, Guy looks a little embarrassed. BRUNO (still insistent on being friendly) I suppose you're going to Southampton -- for the doubles. GUY (politely) You are a tennis fan. Bruno is inordinately pleased by this small tribute. BRUNO Wish I could see you play. But I've got to be back in Washington tomorrow. I live in Arlington, you know. He has taken out a cigarette case. Holds it out to Guy. Converted to PDF by www.screentalk.org 4. BRUNO Cigarette? GUY Not now, thanks. I don't smoke much. BRUNO I smoke too much. He fumbles for a match. Guy brings out a lighter and hands it to Bruno. BRUNO Thanks. (he stares at the lighter, impressed) Elegant. CLOSE SHOT OF THE LIGHTER Showing that it has the insignia of crossed rackets embossed on it, and underneath is engraved the inscription: "To G from A". BRUNO'S VOICE (reading) To G from A. Bet I can guess who A is. WIDER SHOT Guy reacts sharply. GUY (coldly) Yes? BRUNO Anne Burton. Sometimes I turn the sport page and look at the society news. And the pictures. She's very beautiful, Senator Burton's daughter. GUY You're quite a reader, Mr. Anthony. BRUNO Yes, I am. Ask me anything, from today's stock reports to Li'l Abner, and I got the answer. (MORE) Converted to PDF by www.screentalk.org 5. BRUNO (CONT'D) Even news about people I don't know. Like who'd like to marry whom when his wife gets her divorce. GUY (sharply) Perhaps you read too much. BRUNO (contritely) There I go again. Too friendly. I meet someone I' like and open my yap too wide. I'm sorry... At the appeal on Bruno's face, Guy slowly relents. GUY That's all right. Forget it. I guess I'm pretty jumpy. Bruno smiles with and signals a waiter. BRUNO There's a new cure for that. (to waiter) Scotch and plain water. A pair. Double. (to Guy with a chuckle) Only kind of doubles I play. GUY You'll have to drink both of them. BRUNO (grinning) And I can do it. (moving in) When's the wedding? GUY What? BRUNO The wedding. You and Anne Burton. (a gesture of explanation) It was in the papers. GUY It shouldn't have been. Unless they've legalized bigamy overnight. Converted to PDF by www.screentalk.org 6. BRUNO I have a theory about that. I'd like to tell you about it some time. But right now I suppose divorce Is still the simplest operation. The waiter has brought the drinks. Bruno slips the lighter into hip pocket to free his hands for the bills which he gives to the waiter, waving away the change. He offers a glass to Guy. Guy takes it. GUY (as if he needs it) I guess I will. BRUNO (happily) This is wonderful -- having your company all the way to New York. GUY (forced to explain) As a matter of fact, I'm not going direct. I'm stopping off. At Metcalf. BRUNO Metcalf? What would anybody want to go there for? GUY It's my home town. BRUNO Oh, I get it! A little talk with your wife to about the divorce! I suppose she was the girl next door. Held her hand in high school and before you knew it -- hooked! (proud of his perspicacity) Am I right? GUY (laconically) Close enough. BRUNO (raises his glass) Well, here's luck, Guy. Drink up -- then we'll have some lunch sent to my compartment. Converted to PDF by www.screentalk.org 7. GUY Thanks very much. But I think I'll go to the dining car. (he hails a waiter who is passing through with a food-laden tray) Do you know if there are any vacant seats in the dining car now? WAITER Not for about twenty minutes I'm afraid, Sir. BRUNO (pleased) See? You'll have to lunch with me. (motions the waiter back) Say, waiter, bring me some lamb chops and French fries and chocolate ice cream, Compartment D, Car 121. (turns to Guy) What'll you have, Guy? GUY Thanks just the same, but I really don't think -- BRUNO Oh, go on and order. The waiter is hovering impatiently. Guy gives in out of embarrassment. GUY Well, I'll Just have a hamburger and a cup of coffee. BRUNO (delighted, lifts his glass in another toast) To the next Mrs. Haines. Guy nods curtly. DISSOLVE TO: Converted to PDF by www.screentalk.org 8. INT. BRUNO'S COMPARTMENT ON TRAIN (PROCESS) Bruno and Guy are finishing lunch. Bruno has been drinking and his eyes are bright and feverish. An almost empty liquor bottle is near a couple of detective novels covered with gaudily Illustrated dust jackets. Bruno has in unlighted cigarette in his mouth. Guy's lighter is on the table. Bruno snaps it a couple of times, as though fascinated, lights his cigarette and puts the lighter on the table again. BRUNO Sure, I went to college. Three of them. Every time they kicked me out my father threw me back in. (bitterly) He finally gave up. He thinks I'm awfully small fry, not worth the bait. (wistfully) You my friend, Guy? GUY Sure. I'm your friend, Bruno. BRUNO (a little woozy) No, you're not, nobody thinks I'm anything special. Only my mother. (empties the bottle into his glass) My father hates me. Guy smiles this off as nonsense. GUY You must be imagining things. BRUNO (hitting the bottom of the bottle for the last drop) And I hate him. He thinks I ought to catch the eight-five bus every morning, punch a timeclock and work my way up selling paint or something. Him -- with all his money! GUY (amused by Bruno) Well, what do you want to do? BRUNO You mean before or after I kill him? Converted to PDF by www.screentalk.org 9. GUY (chuckling) Before, of course. BRUNO (leaning forward eagerly) I want to do everything. I got a theory you're supposed to do everything before you die. Have you ever driven a car, blindfolded, at a hundred and fifty miles an hour? GUY Not lately. BRUNO I did. I flew in a jet plans too. (his hand traces a swift streak through the air, and he adds sound effects) Zzzzzzzp! Man, that's a thrill! Almost blow the sawdust out of my head. I'm going to make a reservation on the first rocket to the moon... GUY (amused and curious) What are you trying prove? BRUNO I'm not like you, Guy. You're lucky. You're smart. Marrying the boss's daughter is a nice short cut to a career, isn't it? GUY (quickly) Marrying the senator's daughter has nothing to do with it. Can't a fellow look past a tennis not without being a goldbricker? BRUNO Take it easy, boy. I'm your friend, remember? I'd do anything for you. GUY (humoring Bruno) Sure, Bruno, sure. (MORE) Converted to PDF by www.screentalk.org 10. GUY (CONT'D) (glancing at his watch) We'll be pulling in soon. I've got to change trains. BRUNO What'd you say her name was -- your wife's? GUY Miriam. BRUNO That's it. Miriam Joyce Haines. Played around a lot, I suppose? GUY Let's not talk about it any more. BRUNO (almost hopefully) Maybe she'll make more trouble for you. GUY I don't think so. BRUNO You mean you got enough on her to get your divorce no matter what? GUY Let's change subject, Bruno, can't we? BRUNO Okay, Guy. Want me to tell you one of my ideas for murdering my father? GUY (indicating the detective novels) You've been reading too many of these. BRUNO (going right on) You want to hear about the busted light socket in the bathroom, or the carbon monoxide in the garage? GUY No. I may be old fashioned, but I thought murder was against the law. Converted to PDF by www.screentalk.org 11. BRUNO But not against the law of nature. My theory is that everybody is a potential murderer. Didn't you ever want to kill somebody? Say one of those useless fellows Miriam was running around with? GUY You can't go around killing people just because you think they're useless. BRUNO Oh, what's a life or two? Some people are bitter off dead, Guy. Take your -- wife and my father, for instance. It reminds me of a wonderful idea had once. I used to put myself to sleep at night -- figuring it out. Now, let's say you want to get rid of your wife. GUY Why? BRUNO Let's say she refuses to give you a divorce -- (raises a finger and stops Guy's protest) Let's say. You'd be afraid to kill her because you'd get caught. And what would trip you up? Motive. Now here's the plan... GUY I'm afraid I haven't time to listen. BRUNO (ignoring the remark) It's so simple, too. A couple of fellows meet accidentally, like you and me. No connection between them at all. Never saw each other before. Each of them has somebody he'd like to get rid of, but he can't murder the person he wants to get rid of. He'll get caught. So they swap murders. GUY Swap murders? Converted to PDF by www.screentalk.org 12. BRUNO Each fellow does the other fellow's murder. Then there is nothing to connect them. The one who had the motive isn't there. Each fellow murders a total stranger. Like you do my murder and I do yours. GUY (with relief) We're coming into my station. BRUNO For example, your wife, my father. Criss-cross. GUY (sharply) What? BRUNO (with a smile) We do talk the same language -- don't we, Guy? GUY (preparing to leave) Sure, we talk the same language. Thanks for the lunch. BRUNO (beaming) I'm glad you enjoyed it. I thought the lamb chops were a little overdone myself. He holds out his hand. Guy is in a hurry but he shakes hands. GUY Nice meeting you, Bruno. BRUNO (detaining him at the door) You think my theory is okay, Guy? You like it? GUY Sure, sure, Bruno. They're all okay. (he salutes a quick goodbye and hurries away) Converted to PDF by www.screentalk.org 13. Left alone, Bruno picks up Guy's lighter from the table, starts to call Guy back to hand It to him.Then he looks closer at the insignia of crossed tennis rackets. BRUNO (smiling) Criss-cross. DISSOLVE TO: A WIDE VIEW OF THE TOWN OF METCALF METCALF RAILROAD STATION as the train comes in. THE TRAIN STATION PLATFORM MED. SHOT As Guy gets off the with his suitcase and tennis rackets. A baggage man with baggage truck is passing. GUY Hi, Bill. BAGGAGE MAN (smiling) Guy Haines! Good to too you, boy. You be sure to win at Southampton tomorrow, hear me? I've got two dollars on your nose. GUY (indicating his suitcase and rackets) Then park these in a lucky spot for a few hours, will you? BAGGAGE MAN Sure thing. He loads them onto a truck. DISSOLVE TO: INT. METCALF STREET LONG SHOT Guy is walking up the main street. Converted to PDF by www.screentalk.org 14. EXT. MUSIC SHOP Typical music shop of a small town, with plate glass windows and displays of radios, records, sheet music, etc. Activity of a couple of customers and salespeople inside. Guy comes along the street and goes into the shop. INT. MUSIC SHOP As Guy enters. There are the usual counters and shelves, pianos and radios on display, and the sound of a piano being tuned in the back of the store. MIRIAM is finishing with a customer at a counter. MR. HARGREAVES, the manager, is busy at the shelves. Another girl clerk is serving a customer. In one of the glass cubicles where records are tried out, a customer is playing symphonic music; in a second glass cubicle another customer is listening to a record of popular music. A third cubicle is empty. Activity of the street is seen through the plate glass front. Guy walks straight to Miriam, just as she is finishing with her woman customer, handing over a small package. MIRIAM (taking money from customer) Even change. Thank you, Madam. (she looks up at Guy as the woman moves off) Well -- hello, Guy. GUY You're looking well, Miriam. Miriam's face is pretty because it is still young. She is self-centered and inclined to be vindictive. She wears harlequin glasses with myopic lenses which tend to make her eyes look small. MIRIAM So are you. You've got a nice tan, playing tennis with all your rich friends. GUY (ignoring the remark) What time do we meet your lawyer? MIRIAM (sly little smile) What's your hurry? Converted to PDF by www.screentalk.org 15. GUY My hurry? That's funny, coming from you! You're the one who's in a hurry, aren't you? MIRIAM (coyly) When you wouldn't give me the divorce right away, I sort of hoped it was because you were a little bit jealous. GUY (biting) I got over being jealous, a long time ago Miriam. Miriam's eyes slide toward the other girl clerk who has moved closer, within listening range. MIRIAM (indicating empty glass cubicle) Let's talk in there. Guy follows Miriam across to the empty room. Miriam has brought her purse along. They enter. INT. CUBICLE Once inside, the sounds of the music playing from other parts of the shop are heard but very faintly. The piano tuning still goes on, but less stridently. Miriam and Guy are cooped together in the close quarters. MIRIAM (intimately) Now this is cosier. Sort of like old times, isn't it, Guy? GUY (coldly) Oh, skip it, Miriam. It's pretty late to start flirting with a discarded husband. Especially when you're going to have another man's baby. MIRIAM Do you know, I think you're handsomer than ever? Converted to PDF by www.screentalk.org 16. GUY Let's see your lawyer and get this over with. MIRIAM Did you bring the money, Guy? Lawyers are expensive. GUY (taking money from his wallet) Here it is. MIRIAM (taking the money greedily) If I'd known what all that tennis nonsense of yours was going to lead to, I wouldn't have run out on you. GUY What are you trying to say, Miriam? Come out with it. MIRIAM (tucking the bills away) I'm not getting a divorce. GUY (tense and angry) Why, you little doublecrosser. I didn't want this divorce, you did. That's what you've been harping about for the past year. MIRIAM It's a woman's privilege to change her mind... Now I can shop for some pretty clothes. I wouldn't want you to be ashamed of me in Washington when we go to all those dinners and swanky parties. GUY And what do you mean by that? MIRIAM (Coyly) Don't look so mad, Guy. You always smile when your picture is being taken for the papers. (MORE) Converted to PDF by www.screentalk.org 17. MIRIAM (CONT'D) Especially when you have Anne Burton hanging on your arm. GUY Let's not talk about Anne Burton. MIRIAM So, it's really serious between you two? Well, you can throw your dreams about her into the ashcan. Guy, I'm coming to Washington. GUY What for? MIRIAM To have my baby and be with you. GUY Why me? It's not my baby. MIRIAM But people don't know that, Guy, do they? It would make a pretty story, wouldn't it -- the senator's daughter involved with a married man who's about to become a father. GUY (furiously) You black conniving little liar! A few people in the shop look around as Guy's voice rises above the sound of the record playing. MIRIAM Keep your voice down. GUY What happened? Did he run out on you? MIRIAM No man runs out on me. Not even you. GUY You're a liar and a cheat, Miriam. You've wanted to get rid of me long enough and now I'll go you one better -- I never want to see or hear of you again. Converted to PDF by www.screentalk.org 18. MIRIAM (demurely) I could be very pathetic as the deserted little mother in a courtroom, Guy. Think it over. Who would believe you? Guy seizes her angrily and in so doing, knocks the tone arm across the record with a loud screech. From outside we can see heads turn. Mr. Hargreaves, the manager, is very disturbed. MED. SHOT THROUGH GLASS PARTITION FROM HARGREAVES' VIEWPOINT We see Guy gripping Miriam's arms and apparently addressing her in a threatening manner, although we do not hear his words. The smile has faded from Miriam's face and something like cringing fear has taken its place. She is drawn and tense and seems to cower beneath Guy's rage. Mr. Hargreaves moves forward and opens Guy's tirade. GUY ...That's what should happen to people like you. And if I... HARGREAVES (interrupts) Break it up, folks. This isn't the place for a family quarrel. GUY (his eyes blazing) Sorry. I'm leaving. He starts to exit from the booth. Miriam grabs his arm and screams at him: MIRIAM (yelling like a fishwife) You heard what I said, Guy Haines. You can't throw me away like an old shoe. I'm coming to Washington to have my baby. Tell that to the senate! Guy strides out of the store, the manager and a few customers turning around in surprise. Converted to PDF by www.screentalk.org 19. The two customers in other booths, seeing the quarrel, open their doors simultaneously and Miriam's tirade is climaxed by a cacophony of noise, a big symphony, loud hot music, and the apparently unaware piano tuner. EXT. MAIN STREET METCALF SHOOTING TOWARDS STATION Guy is striding along angrily. He comes to the same intersection and the same cop. The officer makes a friendly gesture, is if he'd like to talk awhile, but Guy strides past him without noticing. EXT. METCALF STATION (PROCESS) Guy comes into the scene, crosses to a row of public telephone booths, enters one. Inside the telephone booth, he dumps some loose change on the shelf, sticks a nickel in the telephone, speaks into it. GUY Long distance. (a pause) I want Washington, D. C. The number is Republic 0800. Person to person. Miss Anne Burton. Another pause, very long. Guy is very restless. He digs a cigarette out of his pocket and sticks it in his mouth, then looks through his pockets for his lighter, doesn't find it. He looks puzzled, but about that time the operator speaks to him. GUY (continuing) Right. Guy picks coins up off the shelf and drops them into the telephone, then waits. He shifts the receiver and fumbles in his other jacket pocket, then turns to the phone. GUY (tautly, into phone) Anne, -- Anne darling. Yes, I'm in Metcalf -- (gets a grip on himself) No, everything didn't go smoothly. She doesn't want a divorce, not now.... Converted to PDF by www.screentalk.org 20. INT. BURTON LIVING ROOM ANNE BURTON is a beautiful, high-spirited and well-bred young woman. The smile on her face his faded to anxiety as she listens over the telephone which is on the desk. ANNE (after a pause then with unpleasant realization) Another man's child! But she can't do that to you, Guy -- it's unbelievable -- it's, it's evil! (she listens, then calmly) Yes, I know how you must feel. (pause) But you sound so savage. BACK TO GUY IN TELEPHONE BOOTH GUY (furiously) Sure I sound savage. I feel savage. I'd like to break her neck! (a pause, then raising his voice) I said I'd like to break her foul, poisonous, useless little neck! (the connection is bad and he strains to hear) What's that? Meantime the noise of a through train has been HEARD, and the horn on a streamliner locomotive. It has come up very fast, it is now almost to the station. Guy rises his voice and yells into the telephone. His voice fights the roar of the train: GUY I SAID I COULD STRANGLE HER! The expression on his face is frenzied and suggesting that he means exactly what he is saying. DISSOLVE TO: Converted to PDF by www.screentalk.org 21. INT. ANTHONY LIVING ROOM The scene opens on a CLOSEUP OF A MAN'S HANDS. One of them is semi-flexed and turning slowly, The other is receiving the final touches of a manicure. CAMERA PULLS BACK to reveal that these are Bruno's hands, and that, he is studying them moodily, CAMERA PULLS BACK FARTHER to reveal his mother, MRS. ANTHONY, sitting opposite him at a little table in the Anthony living room. She is working with scissors, file and nail buffer. Mrs. Anthony is a gentle, once pretty woman, whose pastel exterior harbors a tigress-like determination to protect her son, Bruno is in his robe and is unshaven. There is evidence of long established wealth in the heavy dark appointments of this room. MRS. ANTHONY Since you insisted on a manicure, dear, I do wish you'd keep your hands quiet. You're so restless lately. BRUNO (almost dreamily as he admires the free hand) I like them to look just right. Mrs. Anthony looks up, notices his moody expression. MRS. ANTHONY Did I file them too short? BRUNO No, Ma. They look fine. Thanks. MRS. ANTHONY Then what's the matter? BRUNO I'm all right, Ma. Don't worry about me. MRS. ANTHONY You look so Pale, dear. Are you out of vitamins? BRUNO I bought a bottle of them yesterday. A whole fifth. Converted to PDF by www.screentalk.org 22. MRS. ANTHONY (anxiously) But you have that 'look'. I can always tell. You haven't got into any more mischief, Bruno? He denies this with a slow, solemn shake of his head. MRS. ANTHONY I do hope you've forgotten about that silly little plan of yours? BRUNO (sharply) Which one? MRS. ANTHONY (smiling) About blowing up the White House? BRUNO (his eyes dancing) I was only kidding, Ma. Besides, what would the president say? MRS. ANTHONY (laughing gaily) You're a naughty boy, Bruno. But you can always make me laugh. (she rises) Now get shaved, dear, before your father gets home. Bruno's fist crashes down on the little table, upsetting it, as he gets to his feet. BRUNO I'm sick and tired of bowing and scraping to the king. MRS. ANTHONY (placating him) Now, now, Let's not lose control. Come see my painting, dear -- (she leads him toward an easel) I do wish you'd take up painting. It's such a soothing pastime. They look at the painting. Converted to PDF by www.screentalk.org 23. INSERT The painting is a horrible mess. Out of the violence of the pattern a man's face can be discerned, wild-eyed and distorted. We hear laughter from Bruno. BACK TO SCENE Bruno's roar of laughter puzzles Mrs. Anthony, but she is pleased to hear his good humor. He puts an arm around her. BRUNO You're wonderful, Ma! It's the old boy, all right. That's father! MRS. ANTHONY (bewildered) It is? I was trying to paint Saint Francis. At this moment there is the sound of the front door opening. Then immediately the telephone bell rings in the hall. Bruno is instantly alert, as if he had been expecting a call. He goes toward the door to the hall, as the butler enters. BUTLER (to Bruno) They are ready with your call to Southampton, Sir. Bruno's father MR. ANTHONY, purposefully enters the living room. He an impeccably dressed business man with an uncompromising eye. His entrance momentarily blocks Bruno's exit. MRS. ANTHONY (to her husband) How nice that you're early, Charles. I'll tell cook.... Bruno now exits into the hall, passing his father without speaking. MR. ANTHONY Just a minute, Eunice. (calls after Bruno) Bruno! Come here! I want to talk to you and your mother. Converted to PDF by www.screentalk.org 24. INT. HALL CLOSE SHOT BRUNO as he approaches the telephone. BRUNO (calls back to his father) Sorry father. Long distance. (he picks up the telephone) Hello... CAMERA MOVES IN TO A BIG HEAD CLOSEUP OF BRUNO at the telephone as the Voices of his mother and father can be heard from the other room. MR. ANTHONY'S VOICE Now it's hit and run driving! And you knew about it all the time! BRUNO (eagerly into phone) Guy? (pause) Bruno, Bruno Anthony. MR. ANTHONY'S VOICE You're going to protect him once too often. After all we do have a responsibility to society. Bruno gives a look in his father's direction, before he speaks into the telephone in a low voice. BRUNO I just wanted to ask how you made out with Miriam. INT. LOCKER ROOM OF TENNIS CLUB CLOSE SHOT GUY AT TELEPHONE GUY (puzzled) What? (listens) Metcalf? Who'd you say you were? Converted to PDF by www.screentalk.org 25. CLOSEUP BRUNO BRUNO (sotto voce) Bruno, Guy. Bruno Anthony. Don't you remember? On the train. The voices of Mr. and Mrs. Anthony can still be heard in dispute as Bruno listens at phone: MRS. ANTHONY I never permit it! Bruno gives a significant look in direction of the living room as he speaks into the phone. BRUNO (softly) Are you getting your divorce? MR. ANTHONY'S VOICE I tell you he should be sent somewhere for treatment before it's too late. BRUNO (into phone, with satisfaction) So she double-crossed you! Are you going to see her again? The phone clicks in Bruno's ear. He looks hurt for an instant, then replaces the receiver. Bruno listens to his father off scene and his expression becomes more enigmatic. MR. ANTHONY'S VOICE I tell you, Eunice, I'm going to have that boy put away if it's the last thing I do! Bruno looks off in direction of his farther's voice with an expression which says, "Crow while you can, you haven't much time." He reaches into his pocket, brings out Guy's cigarette lighter and as he flicks it on and off. DISSOLVE TO: EXT. METCALF STATION LONG SHOT DAY This is the same shot we saw when Guy arrived in Metcalf. We see the station and one of the main streets beyond the station. Converted to PDF by www.screentalk.org 26. LONG SHOT A NEARER VIEW We see the train come around the curve. Again this is just the same angle that we used for Guy. It comes to a stop in the foreground and we see Bruno alight onto the platform. He looks about him for a moment and then strolls away in the direction of the town. He approaches the row of telephone booths. EXT. STATION CLOSE SHOT We see Bruno enter the small booth and start to glance through the telephone directory. INSERT TELEPHONE DIRECTORY Bruno's finger runs down the names until it stops at: Joyce, Miriam Haines. 2420 Metcalf Avenue. A RESIDENTIAL STREET IN METCALF LONG SHOT It is now much later. It is beginning to get dark, and the street lights are on. In the far distance we see a local bus approaching. MED. SHOT SHOOTING DOWN onto a small seat by a bus stop, we see Bruno with an open newspaper in front of him. It is held up as he reads it. CLOSEUP Bruno is glancing over the top of the paper. LONG SHOT From his viewpoint we see a typical frame house. The upper windows are lit as are the lower ones as well. A woman is sitting in a rocker on the front porch. This is MRS. JOYCE, Miriam's mother. She has white hair. A woman comes along the street and pauses as she gets to Mrs. Joyce. Converted to PDF by www.screentalk.org 27. WOMAN (calls out as she passes) Hello Mrs. Joyce. Warm, ain't it? MRS. JOYCE That it is. WOMAN I've been reading where your son-in- law's been coming right along at tennis. MRS. JOYCE (sourly) We don't have any interest in tennis any more. The neighbor passes on. CLOSE UP Bruno, still glancing over the top of his paper. LONG SHOT Again from Bruno's viewpoint, we see Miriam's house. At this moment the front door swings open, emitting a long streak ot bright light. We see the silhouette of a woman emerge, followed by two other men. They're laughing and joking. Suddenly they look up the street. At this very moment the bus pulls up in front of Bruno's view, cutting off the sight of his quarry. The bus comes to a stop. CLOSE SHOT Bruno rises in alarm and moves around toward the end of the bus so that he shall not lose sight of the girl coming out of the house. SEMI-LONG SHOT From his viewpoint, the girl, whom we now see is Miriam, is running followed by the two young men. They are calling for the bus not to go - shouting, "Hi - stop!" Mrs. Joyce calls from the porch: MRS. JOYCE Don't you stay out too late, Miriam. Converted to PDF by www.screentalk.org 28. MIRIAM (calling back) Goodnight, Mother. See you later. CLOSE UP Bruno watches Miriam. MED. SHOT Miriam comes nearer and nearer to Bruno. With her two companions she brushes past him and jumps onto the bus. THE CAMERA PANS BRUNO AFTER THEM. EXT. AMUSEMENT PARK LONG SHOT We see the bus pull up outside the Amusement Park, and the various passengers alight. These include Miriam nd her companions, and Bruno. LONG SHOT NEARER VIEW OF THE AMUSEMENT PARK We see the usual midway with its various concessions on each side: in the distance the Ferris wheel, Merry-go-rounds, etc., and beyond that a lake. In the foreground we see people filling in and out. DISSOLVE TO: MED. LONG SHOT A GROUP BY A FROZEN CUSTARD STAND This group comprises Miriam and her two boy-friends. They lick their way out of the crowd and debate between themselves where to go next. CLOSE SHOT Miriam's eye catches the attention of something off screen. SEMI-LONG SHOT From her viewpoint we see Bruno standing and casually watching her. Other people pass around and in front of him, so that he is the only immobile figure. Converted to PDF by www.screentalk.org 29. SEMI-CLOSEUP Miriam, with a kind of coy consciousness, turns away with the others and they go on to some other concession. MED. SHOT As Bruno starts to advance in the direction of Miriam he is momentarily held up by a small boy in cowboy uniform carrying a gun and a balloon. The small boy points the gun at Bruno. SEMI-CLOSE UP The small boy pointing the gun fires it twice with a couple of 'bangs!' He then starts to move off. SEMI-CLOSE UP Bruno moves on past the boy. He casually touches the balloon with his cigarette end -- it goes off with a 'pop'. CLOSE UP The small boy turns and looks with dismay at his pricked balloon, wondering what happened. SEMI-CLOSE UP Bruno moves on, pleased with himself, returning his attention to Miriam who is somewhere ahead of him. MEDIUM SHOT Miriam and her two boy-friends by the sledge hammer concession where the aim is to swing the hammer hard enough down onto its target to ring the bell and register the 100 mark. Miriam is in the foreground of the shot. The first boy steps up to try his hand. As he swings, Miriam turns and glances about her, obviously looking for Bruno. LONG SHOT FROM MIRIAM VIEWPOINT The crowds milling, but no sign of Bruno. Converted to PDF by www.screentalk.org 30. MEDIUM SHOT The first boy having failed to ring the bell, the second stops up and slams the hammer down. CLOSE SHOT The register shooting up only to the hallway mark. CLOSE SHOT MIRIAM She looks a little disdainful and again glances around for Bruno. Looking first to her left where she sees nothing, she then looks to her right, and as she does THE CAMERA PANS to show Bruno standing right it her shoulder. Miriam gives a little start. Bruno smiles at her. With a smirk he walks over and after paying his fee, goes to take up the hammer. CLOSE UP MIRIAM She watches Bruno. CLOSE SHOT Bruno looks down at his hands. INSERT Bruno's two strong hands - as he holds them palms tilted upward and fingers curled in. CLOSE UP Bruno, as he smiles faintly, glancing across at Miriam. CLOSE UP MIRIAM She gives a faint smile in return. CLOSE SHOT With a studied movement, Bruno picks up the handle of the hammer and swings. Converted to PDF by www.screentalk.org 31. CLOSE SHOT The register shoots up to the 100 mark and rings the bell. MEDIUM SHOT Bruno drops the hammer and glances around at Miriam again. Her two boy-friends are calling for her from a little distance. BOY'S VOICE Come On, Miriam. Come On! CLOSE SHOT MIRIAM She turns away and is lost in the crowd. MEDIUM SHOT OVER BRUNO'S SHOULDER AT MERRY-GO ROUND IN BACKGROUND Bruno turns to follow Miriam, his manner casual. As he takes a few steps, WE PAN ACROSS with him until, over his shoulder, we see a merry-go-round in the background. Miriam and the two boys are aboard and climbing onto horses. As Bruno goes toward the merry-go-round, the CAMERA MOVES UP A LITTLE with him. The merry-go-round starts to move slowly round as Bruno hops on. MEDIUM SHOT ON MERRY-GO-ROUND Bruno begins to look around for Miriam, who is apparently on the other side of the merry-go-round. He starts to thread his way through the horses which are beginning to move up and down. CAMERA FOLLOWING HIM. He passes one or two of the oncoming heads before he reaches Miriam. She is on an outside mount which is high in the air when she sees Bruno facing her. Her laughter dies for a moment and she smiles at him coyly. Bruno passes her and gets on the horse directly behind her, Miriam glancing at him as her horse comes down. MEDIUM SHOT BRUNO ON HORSE With horse's head in foreground, as it is coming toward us. Converted to PDF by www.screentalk.org 32. SIDE VIEW MIRIAM Miriam on her horse, moving from left to right. Miriam, holding the reins, glances back with a gay laugh. SIDE VIEW BRUNO Bruno on his horse, as though he is chasing Miriam. He is a little more open now in his laughter. GROUP SHOT MIRIAM AND TWO BOYS Miriam and her boy friends begin to sing the song being played on the calliope. CLOSE UP MIRIAM As she starts to sing, she glances back. CLOSE UP BRUNO He is starting to join in the singing. MEDIUM SHOT The horses of the merry-go-round are filling the screen as they whizz by, and again we get the picture of Bruno chasing Miriam as they rush past the CAMERA, the music and tempo at a high speed. LAP DISSOLVE TO: EXTERIOR OF BOAT LANDING ON SHORE OF ARTIFICIAL LAKE Across the water may be seen a small wooded island. Between this and the boat landing there is an artificially constructed "Tunnel of Love". We see Miriam and her companions approach the boat concession and CAMERA FOLLOWS THEM onto the little landing stage. CAMERA MOVES UP SLOWLY over the boy's shoulders until we get MIRIAM IN CLOSE UP. She glances back. Her expression changes to a coy smile of satisfaction as she sees: Converted to PDF by www.screentalk.org 33. MEDIUM SHOT (FROM MIRIAM'S VIEWPOINT) Bruno is approaching the pay box. MEDIUM SHOT Miriam and her companions are escorted to a small boat with electric motor. Once they are seated the boat chugs away from the landing stage and off into the darkness. Bruno steps into the foreground and gets into the next boat which floats alongside. He, too, moves away into the darkness. ENTRANCE TO THE TUNNEL As Miriam's boat passes through, she gives another little glance over shoulder before her boat disappears into the darkness of the tunnel. After a brief moment Bruno's boat comes into the picture, and it, too, goes into the tunnel. INSIDE THE TUNNEL We see the silhouettes of the occupants of Miriam's boat on the wall of the tunnel, lit dimly from the light coming from the tunnel exit. The silhouette of Bruno in his boat, lit by the tunnel entrance, gradually approaches the other three. When the silhouettes are almost touching, we -- CUT TO: EXIT OF THE TUNNEL It is empty. There is a sudden piercing scream from inside, followed after a second or two by protestations and giggling as Miriam's boat emerges into the light. She is pushing one of the boys away from her. MIRIAM (squealing) George, stop it, I tell you! Their boat moves out of the picture, toward the island. Presently Bruno's boat comes smilingly following and he, too, moves on out of the picture. Converted to PDF by www.screentalk.org 34. MEDIUM SHOT ISLAND The group of Miriam and her companions are scrambling out of their boat and moving onto the island, one of the boys trying the boat on the shore. They disappear into the Woods of the island. Again Bruno's boat comes into the picture. He steps out, lift the prow of the boat a little onto the shore. LONG SHOT ISLAND We see the amusement park lighted beyond the lake. Silhouetted in the foreground, the trees and foliage of the island. Nearby we see the silhouetted figures of Miriam and her companions move across the scene, right to left. Miriam is pushing George away from her. MIRIAM (protesting perfunctorily) George, no! She backs away from him and the boys go on picture. Miriam goes in another direction, around, the bushes. George obviously misses her, for we hear his voice call out: GEORGE'S VOICE Miriam! Miriam backs out of the bushes until the back of her head is in CLOSEUP in the foreground of the shot. Suddenly she hears steps in back of her and turns her head toward CAMERA. Her face changes as she recognizes someone offscene. MIRIAM Oh! She gives a coy smile of recognition. CAMERA PULLS BACK to reveal the mad and shoulders of Bruno between Miriam and the camera. His hand holds Guy's lighter which he flicks on as he raises it above Miriam's face. 0f Bruno, we see only the back of his head and shoulders. BRUNO Is your name Miriam? MIRIAM (with surprise) Why yes. How did you -- Converted to PDF by www.screentalk.org 35. We see Bruno's gloved hands dart quickly to Miriam's throat. The lighter falls down out of picture, and as Bruno's hands grip her throat, his head moves slightly to blot out Miriam's face. His head moves a bit farther until Miriam's face is nearly uncovered at the other side of the screen, and we see her glasses fall off. CLOSE SHOT Miriam's glasses hit the ground. The shadows of their struggling figures over the shot. CLOSE UP The screen is filled with one of the lenses of the glasses. They are of the diminishing type. Against the moonlit sky we see reflected, the elongated struggling figures, as though we were shooting up at them. Suddenly one of the figures falls forward. CLOSE UP Miriam's head drops into the picture by the glasses. Bruno's hand comes into the picture and picks up the glasses. One of the lenses has been broken by Miriam's fall. As we see Bruno's sport shoes move away, the CAMERA MOVES PAST MIRIAM'S HEAD until it comes to Guy's lighter pressed into the earth. CLOSE UP BRUNO Bruno glances back over his shoulder. He looks down and goes back one or two steps. CLOSE UP BRUNO'S HAND Bruno's hands retrieve the lighter from the ground. LONG SHOT ISLAND We see a full view of the island again, with the amusement park beyond. The faint noise of the calliope continues in the distance. Bruno has been lost to view. Converted to PDF by www.screentalk.org 36. Miriam's companions are still searching for her. We hear their faint voices in the distance. VOICES Miriam! Miriam! Where are you? MEDIUM SHOT Bruno comes to the shore where his boat is moored. He gets in and is quickly chugging away. He moves calmly, matter-of- fact and not furtively. LONG SHOT LAKE Bruno's boat throbbing its way across toward the landing stage. MEDIUM SHOT LANDING STAGE There are two boats unloading. Bruno's boat is approaching. We hear a loud call from the island. Someone has found Miriam. VOICES Hey, here she is! What's the matter with her? Has she fainted? More shouts from the island cause the people at the landing stage to look back. The boatman's attention is also attracted. Suddenly, as Bruno is getting out of boat, there is a loud scream from the island. VOICE (crying out) She is dead! OTHER VOICE (from island) Help! Help! Bruno by this time has stopped onto the landing stage, and in company with the other people, is looking back as if to see what's wrong on the island. Then he moves away, starting off of the landing stage. The boatman turns and glances at Bruno, but quickly returns his attention to the disturbance across on the island. He hurries forward and with a couple of men passengers jumps into one of the boats. He calls to his assistant as he gets into the boat. Converted to PDF by www.screentalk.org 37. BOATMAN Got a cop! The assistant runs off out of the pictures MEDIUM SHOT BRUNO As Bruno calmly threads his way along the midway, we hear above the noise of the various concessions, a shrill police whistle in the distance. Presently a couple of policemen comes running from direction of the main entrance and past Bruno. He glances at them over his shoulder, then strolls on toward the main entrance to the park. ENTRANCE TO AMUSEMENT PARK EXTERIOR As Bruno comes out through the turnstile, he stands for a moment on the street. At this moment a man hesitates at the curbstone. He is blind and tapping the sidewalk with his white cane. He takes one step into the roadway, then hesitates. Bruno steps forward and takes the blind man's arm. CAMERA PULLS BACK as Bruno escorts the blind man across the road. With a sweeping gesture he holds back a couple of cars to lot them pass. Once on the other side of the road, the blind man utters his thanks. BLIND MAN Thanks. He goes off. Bruno looks back toward the park, then glances down at his wristwatch. INSERT BRUNO'S WRISTWATCH The time is 9.30. LAP DISSOLVE TO: INT. OBSERVATION CAR OF A TRAIN NIGHT Through the rear window we see the tracks rushing away from us. Seated in the foreground are Guy Haines and a rather professorial type opposite him, a bespectacled man around forty-five or fifty who is extremely drunk. Converted to PDF by www.screentalk.org 38. MEDIUM SHOT GUY He is reading an evening newspaper. CLOSE SHOT The feet opposite Guy stretch out and touch Guy's feet. CLOSEUP GUY He lowers his paper and looks across. MED. SHOT The drunk opposite Guy looks down at his feet and then up to Guy resentfully as though Guy had kicked him. He eyes Guy up and down, then suddenly, without warning, bursts into song, to the tune of the Barber Shop Chord. COLLINS There was a man, now please take note. There was a man who had a goat. He loved that goat, Indeed he did. He loved that goat, just like a kid. (He stops singing abruptly and addresses Guy) What is your opinion? GUY (amused) You'll never make the Metropolitan. COLLINS (fuzzily -- pumping Guy's hand) Name's Collins. On sabbatical - Delaware Tech. Glad to meet you. I jus' gave a speech in New York. On integration. In the differential calculus a function is given and its differential is obtained. Understand? GUY (solemnly) Sure, I understand. Converted to PDF by www.screentalk.org 39. COLLINS (resentfully) Y'do? Again he bursts into loud song. LAP DISSOLVE TO: LONG SHOT WASHINGTON EXTERIOR ABOUT 1 A.M. MOONLIGHT A solitary taxi is seen driving past the Capitol Building. LAP DISSOLVE TO: The taxi comes to a side street and stops outside a small apartment house. MED. SHOT Guy gets out of the taxi with his rackets and bag, pays the driver and goes up the steps to the front door of his apartment. CLOSE SHOT As Guy is about to enter the front door and we see his name posted on a small card as one of the several tenants, he hears a soft call from across the street. VOICE (softly) Guy! Guy turns his head and looks across the street. MED. LONG SHOT (FROM GUY'S VIEWPOINT) We see a small space between two houses across the street. Out of the darkness the voice repeats. VOICE Over here, Guy. MED. SHOT GUY He turns, and with a slightly bewildered and wary expression, goes out of the picture to cross the street. Converted to PDF by www.screentalk.org 40. MED. SHOT Guy reaches the other side of the street and still puzzled and cautious, approaches the dark alleyway. MED. SHOT After a moment a figure steps out of the darkness. It is Bruno. He steps back into the darkness again as Guy comes up to him. TWO SHOT Guy frowning in puzzlement as he looks at Bruno. BRUNO (cheerfully) Hello, Guy. GUY (recognizes Bruno -- not pleased) What are you doing here? At this time of night? BRUNO (a little sadly) You don't seem very pleased to see me, Guy. Guy stands without answering. BRUNO (pleased again) I brought you a little present. GUY What do you mean? Bruno's hand comes out of his pocket and he hands Miriam's glasses to Guy. INSERT Guy's hands taking Miriam's glasses from Bruno. One of the lenses is broken. Converted to PDF by www.screentalk.org 41. TWO SHOT As Guy takes the glasses he looks at Bruno in bewilderment. GUY What's this all about? BRUNO Recognize them? CLOSEUP GUY He looks down at the glasses, mystified. He looks up again to Bruno. CLOSEUP BRUNO BRUNO It was very quick, Guy. She wasn't hurt in any way. It was all over in no time. CLOSEUP GUY He is horrified. He looks swiftly down at the glasses in his hand, then back to Bruno. BRUNO'S VOICE (bragging) I know you'd be surprised. Nothing for us to worry about. Nobody saw me, only Miriam. TWO SHOT Guy can hardly believe what he is hearing. BRUNO I was very careful. Even when I dropped your lighter there, I went right back to it up. If It'd been found, it would have ruined our whole scheme, wouldn't it? GUY Are you trying to tell me you've -- Why, you maniac! Converted to PDF by www.screentalk.org 42. BRUNO (looks at Guy with astonishment) But, Guy, you wanted it! We planned it on the train together, remember? Guy suddenly starts to go. Bruno grabs his arm. BRUNO Where are you going? GUY Where do you think I'm going? I'm going to call the police, of course. BRUNO But you can't, Guy. We'd both be arrested for murder. Guy turns back slowly and faces him. GUY We'd both be arrested for murder? BRUNO You're is much in it as I am. We planned it together. Criss-cross. I do your murder -- GUY (suddenly angry) You crazy fool! You think you can get away with that? BRUNO (a little hurt) Oh, come now, Guy. Why should I go to Metcalf and kill a total stranger, unless it was part of the plan and you were in on it? You're the one that benefits, Guy. You're a free man. I didn't even know the girl. Guy makes a move to leave, but Bruno holds on tight. GUY Let me go, Bruno. I had nothing to do with this and the police will believe me. Converted to PDF by www.screentalk.org 43. BRUNO (concerned) If you go to the police now, you'll just be turning yourself in as in accessory. You see, you have the motive. At this moment both turn at a sound across the street. LONG SHOT (FROM THEIR VIEWPOINT) We hear the sound of a telephone ringing in Guy's apartment. The top of one of his windows is open. BRUNO What is it? GUY My telephone. BRUNO (amused) Someone has some news for you, Guy. Guy still stares across the street. LONG SHOT (FROM HIS VIEWPOINT) We see a police car pull up outside Guy's apartment. TWO SHOT Bruno pulls Guy back further into the shadows. Guy instinctively flattens himself against the wall. He looks across the street again. LONG SHOT (FROM HIS VIEWPOINT) We see the two policemen go into his apartment building. TWO SHOT Guy is still flattened against the wall to keep out of light. BRUNO Tell them you know about it already, Guy. Converted to PDF by www.screentalk.org 44. CLOSEUP GUY He looks across at the police, then down at himself with some surprise and disgust, then over at Bruno, suddenly conscious he is behaving like a criminal and that Bruno is responsible for his predicament. GUY (muttering) You've got me acting, like a criminal, you crazy fool! Bruno for a moment looks menacingly at Guy. BRUNO Don't you call me that. Bruno's flare of anger dies. They both look again across the street. LONG SHOT (FROM THEIR VIEWPOINT) The two policemen come out of the house, get into their car and drive off. Guy's telephone is still ringing. TWO SHOT BRUNO You must be tired, Guy. I know I am. I've sure had a strenuous evening. Guy looks at him, almost numb. BRUNO Now look, Guy, about my father. I have the plans made. Two plans. A plan of the grounds and a plan of the house. I have in old Luger I bought at a pawn shop in San Francisco. My father -- Guy turns and starts to move in across the street. TWO SHOT Bruno follows Guy and we FOLLOW them across the street. CAMERA ON THEIR BACKS. Guy strides ahead to the house. Converted to PDF by www.screentalk.org 45. BRUNO Wait a minute, Guy. To have to talk. We have to arrange things. Guy turns at the door to his apartment building. GUY (furiously) Get away before I give you what you gave Miriam. BRUNO (sadly) You're not yourself, Guy. You're tired. When you think things over, you'll see I'm right. Tomorrow -- Guy opens his door, turns on Bruno. GUY (with finality) I don't know you. I never saw you before. I never want to see you again. He goes in and slams the door in Bruno's face. BRUNO (to the closed door) But we have to -- He realizes there is no use in trying to talk to Guy any further. He turns and faces the CAMERA IN CLOSE UP as he moves away, looking sad almost to the point of tears. INT. GUY'S APARTMENT Guy is standing at the telephone which is still ringing. He has Miriam's glasses in his hand. He looks down at them for a moment, then picks up the receiver. He hesitates, then speaks into the phone. GUY (hoarsely, into phone) Yes? (Pause) Yes, Anne. I'm sorry, darling. I just got in. (pause) Of course I'm all right. (MORE) Converted to PDF by www.screentalk.org 46. GUY (CONT'D) (forcing his voice to sound normal) But you sound upset. Is anything wrong? (Pause) All right. I'll come over. Right away. He hangs up but keeps his hand on the telephone, deliberating. He starts to dial, then suddenly hangs up and starts out. DISSOLVE TO: EXT. A RESIDENTIAL STREET, WASHINGTON LONG SHOT NIGHT A taxi drives up and stops in front of a handsome residence. It is the Burton home. Guy gets out of the taxi and goes up the steps. MED. SHOT OVER GUY'S SHOULDER His figure tense, he rings the bell. After a moment's wait, the door is opened from inside and Anne Burton stands in the lighted hallway. She looks at Guy with an anxious, taut expression, searches his face hastily, then as he takes a step inside she is suddenly in his arms. They embrace with wordless fervor. GUY (holding her close) Anne darling, you're trembling. Anne draws back and looks into his face as if searching for an answer to some question in her mind. ANNE Guy -- (her fingers gently touch his face) I wonder if you know how much I love you. Guy takes her hand from hIs face, caresses it with his lips. GUY (forcing a smile) Brazen woman. I'm the one to say that. Converted to PDF by www.screentalk.org 47. ANNE (tensely) But I wanted you to know, before... (forcing herself to be calm) Before we go into the living room. Father wants to see you. CLOSEUP GUY He looks apprehensively in direction of the living room, conscious of what the news is to be, but covering up. LONG SHOT LIVING ROOM FROM GUY'S VIEWPOINT SENATOR BURTON and BARBARA BURTON are seated near a desk on the farthest side of the room. Senator Burton is a distinguished fifty, a man with great pride in tradition, his family and his career. Barbara, Anne's younger sister, is a lively seventeen who loves excitement, says exactly what she thinks and rarely thinks before she says it. Superficially, in height and figure, she resembles Miriam. She also weirs glasses. By her gestures we gather she is speaking urgently, but softly, to her father, who lifts a weary hand to quiet her as she looks toward Guy in the hallway, Barbara keeps quiet and also looks toward Guy. They both wait for him to enter. CLOSEUP GUY He steels himself for the long walk across the hall and the living room. CLOSEUP ANNE Watching Guy closely. MED. SHOT As Guy starts to make the long trek across the living room, with Anne behind him -- GUY (stiffly) Good evening, sir. Hello, Babs. Converted to PDF by www.screentalk.org 48. Barbara has been squirming in her seat, then as if jet propelled she catapults out of it and runs to Guy, giving him a big hug and a smack on the cheek. BARBARA Something awful has happened, Guy. SENATOR (firmly) Sit down, Barbara. Subdued, she sits down. But Guy remains standing. SENATOR (finding it difficult to begin) There seems to be no way of diplomatically breaking tragic news. I'm sorry, Guy, to be the one to tell you. It concerns your wife. She's been murdered. Guy stares woodenly at the Senator, is if hypnotized. BARBARA The police have been using everything but radar to locate you. SENATOR You're to call Headquarters at Metcalf. The full impact of what has happened hits Guy once more. GUY Miriam...murdered. ANNE (with inner tension) She was...strangled. Slowly Guy's eyes meet hers. They are remembering what he said on the phone: "I could strangle her." He sinks into a chair. The Senator is quite distressed. During the following scene Barbara quietly goes about the business of pouring drinks and serving them. She knows everyone's preference. Converted to PDF by www.screentalk.org 49. SENATOR (wrylt, to Guy) It happened on an island in an amusement park. It was sort of a lovers lane, I believe. A rather sordid atmosphere. BARBARA (quickly, to Guy) Miriam went there with two boys. They were the ones who found her. So they're not suspects. But you probably will be. SENATOR Young lady, we can't overlook the fact that murder is at our doorsteps. But I forbid you to drag it into the living room! BARBARA (wide-eyed) Let's not fool ourselves. The police will say Guy wanted Miriam out of the way so he could marry Anne. In a crime of this sort the police first go after the husband, and Guy had every motive. SENATOR (aghast) Motive? GUY (quietly) She's right. Whichever way you look at it...I'm in a spot. SENATOR (disconcerted but whistling in dark) Oh come now, my boy. I'm sure you have nothing to worry about. BARBARA (flatly) If he hasn't an alibi for nine-thirty tonight he has plenty to worry about. Converted to PDF by www.screentalk.org 50. ANNE (who hasn't taken anxious eyes off Guy) You can tell them where you were, can't you, Guy? GUY (wearily) At nine-thirty I was on the train from New York to Washington. SENATOR (relieved) There you are. BARBARA Who saw you? Did you speak to anyone? You'll need a Witness, you know. GUY (as if it didn't matter) Yes, I spoke to someone. SENATOR (hopefully) Anyone you know? GUY No. His name was Collins. He is a professor. SENATOR (brightening) Harvard. GUY University of Virginia. The Senator's expression says: "Well, that's not too bad." CLOSEUP ANNE Her face shows her relief that Guy can account for his time. ANNE Then everything's's all right. Converted to PDF by www.screentalk.org 51. BACK TO SCENE BARBARA Not quite. Detectives play a game called Motive, Motive, Who'd got the Motive. ANNE (near the breaking point) I'm sick of hearing that word! BARBARA He'll still have to answer questions. SENATOR Routine. Pure routine. GUY I'm afraid there'll be a lot of reporters at your front door in the morning. BARBARA Daddy doesn't mind a little scandal. He's a senator. ANNE (answering Guy's look) It can't be helped, darling. It is not your fault. It's not as though anyone can say you had something to do with it. GUY Someone might say it...I'd do anything to keep you all out of this mess. SENATOR Profit by my experience, Guy. Never lose any sleep over accusations. (an afterthought) Unless they can be proved, of course. We'll help all we can. Dreadful business, dreadful. That poor unfortunate girl. BARBARA (flatly) She was a tramp. Converted to PDF by www.screentalk.org 52. SENATOR (pontificially) She was a human being. let me remind you that even the most unworthy of us has the right to life and the pursuit of happiness. BARBARA (unimpressed) From what I hear, she pursued it in all directions. SENATOR Barbara! ANNE Father, it's getting terribly late, and Guy looks so tired... SENATOR (quickly) Of course, of course. Back to bed, Barbara. BARBARA (ignoring this - to Anne and Guy) Well, you two. Nothing stands in your way now. You can be married right away. Think of it -- you're free! CLOSE TWO ANNE AND GUY look at one another with a growing realization of what Miriam's death actually means to their happiness -- they are free. BACK TO SCENE The Senator firmly urges Barbara to the door. SENATOR (to Barbara) One doesn't always have to say what one thinks! BARBARA (sweetly) Father, I'm not a politician. Converted to PDF by www.screentalk.org 53. The Senator gives her a gentle but firm push out of sight. SENATOR You won't forget that call, Guy? Captain Turley. GUY Yes sir. Goodnight. Barbara pokes her head quickly around the door. BARBARA I still think it would be wonderful to have a man love you so much he'd kill for you. (she ducks out) TWO SHOT Left alone, Guy and Anne embrace. Anne's nervous tension comes to the surface in a flood of relief. ANNE I told myself over and over I was being silly, but there was one horrible moment tonight when the news came through. I kept remembering what you shouted telephone from Metcalf. GUY That I could strang... Anne quickly puts her fingers over his mouth. ANNE Don't even say it. Forget you ever said it. Even more terrifying than the murder itself, Guy, was the awful thought that if you had anything to do with it we'd be separated, -perhaps forever. I'd never see you again. I couldn't bear it. DISSOLVE TO: LONG SHOT MAIN STREET OF METCALF DAY with its customary mid-afternoon activity. LAP DISSOLVE TO: Converted to PDF by www.screentalk.org 54. EXT. METCALF POLICE HEADQUARTERS DAY A knot of people are hanging around the entrance, including a few newspaper photographers. There is a rush of interest when a taxi pulls up and Guy steps out of it. Guy pushes his way through the people. Two or three bulbs flash. There is a murmur from the crowd and we hear Guy's name. He passes into the entrance. INT. CORRIDOR OF POLICE HEADQUARTERS Guy comes into the corridor from the street and approaches two policemen who are standing nearby. GUY Captain Turley's office? One of the policemen gestures to a door at the right. Guy crosses and enters. INT. RECEPTION ROOM OUTSIDE CAPTAIN TURLEY'S OFFICE At one side of the room is a young police sergeant seated at a typewriter. A group of people are seated in chairs lined against the opposite wall. Guy enters, crosses to the sergeant at the desk. GUY Captain Turley is expecting me. Guy Haines. SERGEANT Just a moment, Mr. Haines. He rises and goes into an adjoining room. CLOSEUP GUY He now has time to take stock of the waiting people. He catches his breath when he sees: CLOSEUP MRS. JOYCE Miriam's mother, dressed all in black, is seated in one of the chairs. She has been staring at the floor, but brings her eyes up slowly to glare at Guy with a look of burning hatred. Converted to PDF by www.screentalk.org 55. MRS. JOYCE (a fierce whisper) You'll pay for this! CLOSEUP MR. HARGREAVES Mr. Hargreaves from the music shop looks across at Guy, attempts in awkward nod but is very embarrassed. CLOSEUP GUY Guy nods in returns. MED. SHOT The two boys who were with Miriam at the amusement park. They look at Guy with interest. MED. SHOT GUY He looks about him uncomfortably, then turns suddenly as he sees: MED. SHOT Seated behind Guy, apart from the others who are waiting, is Professor Collins, Guy's drunken companion on the train of the night before. The professor is completely sober now, dignified and erect. He has removed his glasses to polish them and does not react to Guy's presence. CLOSEUP GUY Guy starts with a smile of recognition to say, "How do you do?" but at that moment he hears the door open and his name called: SERGEANT'S VOICE Will you come in, please, Mr. Haines? MED. SHOT Guy breaks away from his uncompleted greeting to the professor and goes through the door to Captain Turley's office, followed by the eyes of the waiting people. Converted to PDF by www.screentalk.org 56. INT. CAPTAIN TURLEY'S OFFICE CAPTAIN TURLEY is conscientious, methodical and always polite. He puts aside photographs and records and rises from behind his desk as Guy comes in. A detective lieutenant, CAMPBELL, is attending a coffee maker. Their expressions are grave by contrast with Guy's confident attitude after seeing the professor in the waiting room. CAPTAIN TURLEY Good of you to be so prompt, Mr. Haines. This is Lieutenant Campbell. (the two nod to each other) Won't you sit down? GUY Thank you, sir. (he sits) CAPTAIN TURLEY I know you're a busy man, so we won't detain you any longer than necessary...Now you already been good enough to tell us where you were last evening, and we've managed to locate the gentleman you spoke with on the train. Turley signals to Campbell to call the professor in. GUY (brightening) Yes. I saw him outside. CAMPBELL (at open door) Will you come in please, professor? CLOSEUP GUY He looks up eagerly. MED. SHOT Professor Collins comes in and sits in a chair opposite Guy. TURLEY Professor Collins, this is Mr. Haines. He was with you on the train last night. Converted to PDF by www.screentalk.org 57. The professor studies Guy for a moment, then awkwardly turns to Turley. COLLINS I'm terribly sorry, but I really don't remember meeting this gentleman. CLOSEUP GUY Surprised. His confident expression fades. CLOSEUP PROFESSOR COLLINS He turns from the captain to Guy. COLLINS (apologetically) Unfortunately, I remember very little about the journey from New York...You see, there had been a little celebration -- MED. SHOT GROUP Guy interrupts with a slight note of impatience. GUY But we were sitting opposite each other in the observation car! You were singing a song about a goat -- COLLINS (incredulously) A goat? GUY (urgently) And calculus. You were going over a speech you'd made. Turley and Campbell are watching closely. COLLINS I was? I'm sorry, Mr. Halnes. (shakes his head) I certainly must have celebrated! I can't remember you at all. Converted to PDF by www.screentalk.org 58. CLOSEUP GUY Momentarily Guy is frustrated, then he turns quietly to Turley. GUY (calmly, logically) Captain, is it so important whether or not Professor Collins remember me? Surely, the important thing is that I've been able to name a man who was on the train with me. You've been able to find him. Isn't that proof of where I was at nine-thirty last night? Guy asks this question with a look of near triumph that he has clearly established his alibi. DISSOLVE TO: INT. BURTON LIVING ROOM EVENING The Burtons are having coffee. Barbara has been glancing through a new murder mystery with a lurid cover. As Guy enters, Anne rises to greet him. ANNE Hello, darling. Have you had your dinner? GUY On the train. ANNE You weren't in Metcalf all this time? We expected you hours ago. BARBARA (flatly) I didn't. They sometimes throw a suspect in the can and keep him there all night. SENATOR (after a disapproving glance at Barbara) Sit down, Guy. Sit down. Give him some coffee, Anne. (MORE) Converted to PDF by www.screentalk.org 59. SENATOR (CONT'D) (back to Guy) You had no trouble with the police of course, once they verified your alibi? GUY (morosely) When an alibi is full of bourbon, sir, it can't stand up. BARBARA You mean the professor was boiled? GUY Completely. He didn't remember me. ANNE But, you knew he was on the train! Wasn't that enough to prove you were on it, too? GUY Apparently not at the right time. They suggested I could have caught the train at Baltimore after Miriam was murdered. They had it all worked out -- (taps his head) in their timetables. ANNE (growing indignant and increasingly nervous) That's ridiculous. They're acting as if you were guilty. BARBARA (somewhat subdued and trying to be comforting) Everything will be all right, Anne. The police were just being thorough -- (she's unsure of herself, and defers to the senator) Weren't they, daddy? SENATOR I certainly hope so. (to Guy) What is your next step? Converted to PDF by www.screentalk.org 60. GUY (wryly) Whatever it is, the police will know it. They gave me a present -- come take a look. He crosses to the window, lifts the curtain slightly, then turns back to the others. GUY (continuing) My guardian angel. The group move to look out the window, the senator with reluctance. LONG SHOT EXT. STREET FROM THEIR VIEWPOINT Through the window we see the figure of a man across the street. He is lighting a cigarette and strolling up and down. BACK TO GROUP BARBARA (impressed) You're being tailed! GUY (turning to them) That's Leslie Hennessy. He works sixteen hours a day. Somebody else takes over for the next eight. (drops the curtain, turns back into room) As a matter of fact, Hennessy's a very nice fellow. BARBARA Shouldn't we ask him in for Coffee -- or something? Nobody bothers to answer her. The Senator is disturbed, but confident of his own prestige as he goes back to his coffee. SENATOR I'll have him called off immediately of course. Converted to PDF by www.screentalk.org 61. GUY (calmly) I'm afraid where I go, Hennessy goes. Even to the Senate. SENATOR (Pausing with his cup hallway to his mouth) Is he likely to -- picket my office? GUY Very likely. The Senator's cup is suddenly back on its saucer and he is on his feet, pacing nervously. SENATOR I would suggest, Guy, for your own peace of mind, of course, that you work here at the house for a few days. (a pause) It would be less embarrassing for you. Guy has been looking at Anne and is concerned at the worry on her face. He nods in assent to the Senator's suggestion, but puts his hand over Anne's. GUY (hopelessly) Then what about practicing? Perhaps I'd better forget Forest Hills? SENATOR My dear boy, wouldn't it look rather -- awkward -- if you suddenly canceled all your plans. ANNE He's right, Guy. You mustn't do anything that would look suspicious. You've got to carry on as though nothing has happened. BARBARA (pointing out the window) Escorted by Mr. Hennessy. The are crestfallen again. RANDALL, the manservant, has entered with the telephone. Converted to PDF by www.screentalk.org 62. RANDALL A call for you, Mr. Haines. They say it is urgent. The phone is plugged in to a connection and Guy crosses the room and picks up the receiver. The Burtons watch him. GUY Hello -- INT. TELEPHONE BOOTH BIG HEAD CLOSEUP OF BRUNO His face wears the most affable expression. BRUNO Hello, Guy. I tried your apartment, but -- (pause) Why, Guy, this is Bruno! INT. BURTON LIVING ROOM Guy hangs up the telephone quickly. He looks at the others, awkwardly tries to explain: GUY Must be some mistake. It wasn't for me. His embarrassment grows as Anne looks at him with a puzzled expression. FADE OUT. FADE IN EXT. WASHINGTON STREET APPROACHING JEFFERSONS MEMORIAL DAY Guy and HENNESSY are walking along the street together, CAMERA MOVING WITH THEM. Their relationship is most friendly. Guy carries a briefcase. Hennessy is an amiable but not gullible young man in his early thirties. He knows his job, is well groomed, well educated, and well liked. GUY Well, I suppose I was pretty lucky to be seeded fifth, really. Converted to PDF by www.screentalk.org 63. HENNESSY I've never seen the Forest Hillss tournament before. I'm looking forward to it. GUY (wryly) Do you mean we'll be going there together, Hennessy? HENNESSY Oh, don't worry. This thing will be cleared up by that time. (changes the subject) Ever thought of turning professional, Guy? GUY I won't have to do that. When I'm through with tennis. I'll be going into politics, I hope. HENNESSY (aghast) Politics! It's a good thing for you I don't report that to the chief. He turns to light a cigarette. As he does, Guy gives a barely perceptible start at what he sees offscene. LONG SHOT JEFFERSON MEMORIAL FROM GUYS VIEWPOINT The tiny figure of a man is standing at the base of the tall white column. The figure lifts in arm and waves. Instinct tells us that this is Bruno. Hennessy is still mumbling his opinion of politics. HENNESSY'S VOICE If he knew you were getting into that rat-race -- TWO SHOT GUY AND HENNESSY Guy turns his back on Bruno's figure and looks frantically toward to street, wanting to get away. HENNESSY -- He'd put ten men on your trail. He says -- Converted to PDF by www.screentalk.org 64. GUY (interrupts) Let's take this cab. It's getting late. He hails a taxi which is cruising by, and they start to get in. Guy directs the driver. GUY Pentagon Building, please. HENNESSY Oh, no, not there! I always get lost. INT. TAXI CLOSE SHOT Guy turns and looks out of the window. LONG SHOT JEFFERSON MEMORIAL from Guy's viewpoint, shot through the cab window. Again we see the solitary figure of Bruno looking after Guy and beginning to recede with the background as the cab starts off. DISSOLVE TO: INT. GUY'S APARTMENT NIGHT As Guy comes in from outside, there is a note on the floor that has been pushed under the door. Guy picks it up, stares at it for a minute before he opens it. He takes out a handwritten note and reads it with an expression of disgust. INSERT NOTE (IN GUY'S HANDS) IT READS: Dear Guy: We have to meet and make plans. Call me at Arlington ----. Time's getting short. Bruno The handwriting is sprawling and erratic, embellished with conceited flourishes. Converted to PDF by www.screentalk.org 65. MEDIUM SHOT Guy looks off for a moment with set face, then tearing the note into shreds, crosses to a small desk, lights a match and holds it to the fragments, letting them burn and fall into an ash tray. Guy looks off for a moment with set face, then tearing the note into shreds, crosses to a small desk, lights a match and holds it to the fragments, letting them burn and fall into an ash tray. DISSOLVE TO: LONG SHOT EXT. MELLON GALLERY LATE AFTERNOON CAMERA is in a low setup, to take in the sign across the doorway which identifies the gallery. Hennessy stands in the foreground in front of the building, on duty. LAP DISSOLVE TO: INT. MELLON GALLERY Guy and Anne are walking slowly through a more or less deserted room of the gallery. Their manner is relaxed and intimate. ANNE Well, we'd better be getting back. GUY We've actually been alone for an hour. Seems almost indecent. You like? ANNE (softly) I like. GUY I was beginning to feel like a goldfish. ANNE So was I. When we build our house, darling, we won't even have glass windows. No doorbells, no newspapers, no telephone -- Converted to PDF by www.screentalk.org 66. GUY No Hennessy. ANNE (suddenly serious) How long can it go on? GUY I don't know. I suppose until they find out who did it. ANNE We'll be happier then, won't we? GUY I suppose so. Anne looks it him, surprised at his lack of enthusiasm. They walk on out of the picture. A figure steps out from behind a pillar in the main hall of the gallery, near the spot from which they have disappeared. It is Bruno. He calls. BRUNO (softly) Guy! Anne stops and looks back. Guy knows who it is and would not turn but that he is forced to by Anne's action. He takes a few steps towards Bruno. CLOSEUP Anne watches Guy approach this stranger. She looks downward at Bruno's tie pin. CLOSEUP Bruno's tie pin, bearing his name, gleams in the light. CLOSEUP Anne reads the name on the tie pin. TWO SHOT Guy comes up to Bruno, steps in front of him. Converted to PDF by www.screentalk.org 67. GUY (muttering harshly) Will you stop pestering me! BRUNO But Guy, you haven't called me. My father's leaving for Florida the end of this week -- GUY (interrupts) You crazy fool! There's a detective outside. He'll see us together! BRUNO (brushing this off) Oh, they can't have anything on you. (looking past Guy) Isn't that Anne Burton? Slight improvement over Miriam -- eh, Guy? GUY Stay away from me, I tell you! He leaves Bruno abruptly to rejoin Anne. Bruno looks after him, a little hurt. TWO SHOT Guy rejoins Anne and they start to walk away. ANNE Who was it, Guy? GUY (unnerved) I never saw him before. Just some tennis fan. Anne looks at him a little oddly. He seems unduly concerned about a casual stranger. CLOSEUP ANNE Her face is troubled. FADE OUT. Converted to PDF by www.screentalk.org 68. FADE IN INT. MORTON STUDY MED. SHOT Guy and a secretary have set up office in the Morton study. As the scene opens the secretary is handing Guy a large envelope. SECRETARY Here's a special delivery, Mr. Haines. It's marked personal. As Guy is opening the envelope, Barbara speaks to him from atop a library ladder. She is getting a book from one of the top shelves of a bookcase, which is next to a window. BARBARA Are you getting in any practice today, Guy? GUY (as he takes out a large folded sheet of paper and glances at it, mystified) Yes, if I can get a court at the club. As Guy's hands unfold the paper and hold it for moment, we see that it is a diagrammed plan of the grounds and the Interior of the Anthony house. There are dotted lines along the upper hall, with an arrow which points to one room and where Bruno has indicated in his handwriting, "My father's room." Over this we hear the voices of Barbara and the secretary: SECRETARY'S VOICE Barbara, who are you waving at? BARBARA'S VOICE Mr. Hennessy. I think it is a shame Daddy won't let us have him in the house to sit down. Have you met him yet, Louise? SECRETARY'S VOICE No. BARBARA'S VOICE He is awfully cute. Converted to PDF by www.screentalk.org 69. MED. SHOT Guy frowns, quickly folds the paper up and stuffs it into his pocket. He looks off abstractedly. CLOSEUP SECRETARY She looks at Guy sympathetically. SECRETARY Is anything wrong, Mr. Haines? CLOSEUP GUY Her voice breaks his reverie. He answers her with a forced smile. GUY No, thank you, Louise. FADE OUT. FADE IN TENNIS COURT AT WASHINGTON COUNTRY CLUB There are twenty or thirty people sitting in the bleacher seats opposite the umpire's chair. A game of mixed doubles is in progress. MED. SHOT AT THE ENTRANCE TO THE COURT Guy appears, carrying his racquets. His partner for the forthcoming game, and one or two other players, are close by. CLOSER SHOT Guy looks about him. Several people are looking at him awkwardly or avoiding his eyes. He moves self-consciously away, and the CAMERA PANS HIM around the court to the umpire's chair. MED. SHOT A couple of women players whisper something about Guy as he goes past them. Converted to PDF by www.screentalk.org 70. FIRST WOMAN I didn't think he'd show up after what happened. SECOND WOMAN And miss all the publicity? MED. SHOT As Guy stands at the umpire's chair, the umpire glances down and gives him a rather embarrassed greeting. CLOSEUP GUY He looks across at the watching crowd. MED. SHOT FROM GUY'S VIEWPOINT The heads of the people in the bleachers move from side to side, to follow the play on the court. One head is not moving. It is staring at Guy. It is Bruno. At this moment, we hear the umpire calling, "Game, set and match" to the winning mixed doubles pair. CLOSEUP GUY His expression becomes set. LONG SHOT The mixed doubles couples complete their handshaking at the net and move off the court. We see Guy move up to the base line while the other player takes his position for the preliminary knock-up. MED. SHOT As Guy casually knocks the ball across the net, he glances again toward Bruno. MED. SHOT FROM GUY'S VIEWPOINT Bruno is making his way out of the small stand. Converted to PDF by www.screentalk.org 71. CLOSEUP GUY Perplexed and apprehensive as to what Bruno may be up to. He hears his opponent's voice. PLAYER'S VOICE Ready, Guy? Guy shakes off his abstraction and poises himself to receive the ball. LAP DISSOLVE TO: MED. SHOT PASSAGEWAY LEADING TO TERRACE We see Guy coming alone, having fInIshed his game. He is carrying his rackets, wears a towel around his neck, etcetera. He walks into foreground, into CLOSEUP, and suddenly stops short at what he sees: MED. SHOT FROM GUY'S VIEWPOINT The group at the table comprising Bruno, Anne and the two French people. Bruno is preening himself as the others laugh uproariously, obviously at something Bruno has said. Anne catches sight of Guy and smiles at him. CLOSE SHOT GUY CAMERA MOVES WITH HIM as he comes forward toward the table. MED. SHOT GROUP AT TABLE As Guy comes into the scene. He stands staring. ANNE Guy, darling -- this is Mr. Antony -- a friend of Monsieur and Madame Darville... (to Bruno) Guy Haines. CLOSEUP GUY He gives a weak acknowledgment in Bruno's direction, realizing that Bruno has wormed his way into the group and that he must accept the introduction. Converted to PDF by www.screentalk.org 72. MEDIUM SHOT Bruno half rises, smiles affably at Guy, reaches out his hand. Guy is forced to shake hands with him BRUNO I've been a fan of yours for a long time, Mr. Haines. In fact, I follow everything you do. MME. DARVILLE Mr. Antony has been telling us such charming stories... Very funny. CLOSEUP GUY He gives another weak little smile. MED. SHOT In response to the Frenchwoman's attentive and eager expression, Bruno leans forward on the table and starts saying something more in extremely fluent French. CLOSEUP ANNE She is staring at Bruno with a new expression. CLOSEUP FROM ANNE'S VIEWPOINT Bruno's coat has spread open a bit, and his tie pin bearing the name "Bruno" is resting on the edge of the table. CLOSEUP ANNE She becomes aware that this is the man she has seen call to Guy in the art museum, that they have met before. Her eyes turn a little in Guy's direction, though she does not look at him. CLOSEUP GUY He is still watching Bruno talk to the French couple. Guy is unaware of Anne's looks. Suddenly his attention is arrested by the sound of Barbara's voice calling him. Converted to PDF by www.screentalk.org 73. BARBARA'S VOICE Guy! He turn his head and CAMERA PANS him to Barbara, who is standing a few steps from the table beckoning to him. BARBARA (Sotto voce) I've just been talking to your shadow. (very impressed) Guy, did you know Mr. Hennessy helped crack that axe murder I was reading about? You know, the one where the body was cut up and hidden in the butcher shop? He was locked in the ice box with the left leg for six hours! GUY He pulls those yarns right out of his hat, Babs. CLOSEUP GUY He gives a sharp look back toward Bruno. There is more laughter coming from the French couple at the table. CLOSE SHOT GROUP AT TABLE FROM GUY'S VIEWPOINT Bruno is occupied with his French joke, but Anne is looking at Guy strangely. TWO SHOT GUY AND BARBARA Guy turns back to Barbara. Barbara looks with interest toward Bruno. BARBARA Who's the nice looking Frenchman with the Darvilles? GUY He's not French. His name's Antony. Barbara steps toward the table. MED. SHOT AT TABLE as Barbara joins the group. Converted to PDF by www.screentalk.org 74. BARBARA How do you do, Madame Darville. Monsieur. They looks up. CLOSEUP BRUNO Bruno stops in the middle of some French to stare at Barbara. Her voice continues. BARBARA'S VOICE How are you? FRENCH COUPLES' VOICES Delightful to see you. How sweet you look, Miss Barbara. CLOSE SHOT BARBARA FROM BRUNO VIEWPOINT BARBARA I hope you aren't forgetting our little party on Thursday, Madame. From Bruno's viewpoint, as Barbara speaks,CAMERA MOVES IN CLOSER until to faintest impression of the merry-go-round fills the screen with the effect of whirling around Barbara's head. Her glasses seem to glint until her eyes are obliterated by the glare. MED. SHOT THE GROUP MME. DARVILLE We are planning on it? M. DARVILLE But of course. All talk dies out as all eyes turn to Bruno, who is staring at Barbara. Except Anne's, who is saying quietly to Bruno: ANNE This is my sister Barbara. Barbara, this is Mr. Antony. CLOSEUP BRUNO He does not acknowledge the introduction immediately. He is still staring at Barbara. Then he nods abstractedly. Converted to PDF by www.screentalk.org 75. CLOSEUP ANNE She is looking at Bruno, wondering what mystery lies behind this strange individual and why he and Guy have disclaimed any previous acquaintance. FADE OUT. FADE IN INT. GUY'S APARTMENT NIGHT CLOSEUP A LUGER PISTOL HELD IN GUY'S HANDS CAMERA PULLS BACK TO SHOW Guy staring down at it. He is partially dressed for an evening party, in black bow tie but without his jacket. He leans forward to take up a letter from among brown paper wrappings on the table. INSERT: LETTER Dear Guy -- Just two more days left. We must get together for final details. The note, in Bruno's handwriting, is unsigned. CLOSEUP GUY He stares down at the note. At this moment there is a knock at the door. MED. SHOT Guy hastily gather together the gun, the note and the wrappings and puts them in a dresser drawer. He crosses to the door and opens it. Hennessy enters, carrying a topcoat. GUY Hiya, Hennessy. Won't keep you out late tonight. (getting into his dinner jacket) With Forest Hills coming up tomorrow, I've got to get some sleep. Converted to PDF by www.screentalk.org 76. HENNESSY (helping himself to a cigarette) That's too bad. Hammond takes over in a couple of hours. I'd like to see him earn his salary. Guy turns to the dresser drawer in which he has put the note and the gun, maneuvering his body between the dresser and Hennessy's view. He takes out a handkerchief, closes the drawer, sticks the handkerchief in his pocket, speaking as he does so. GUY Doesn't that bloodhound over relax? He sticks so close he's beginning to grow on me -- like a fungus. HENNESSY (mildly) He thinks you're a very suspicious character. He doesn't trust anybody! Not even himself. Guy is eager to get out of the room, and Hennessy is maddeningly slow in his movements. GUY Come on. (indicating at Hennessy overcoat) Don't forget your sleeping bag. HENNESSY (taking his time) Yeah, If I have to wait too long on the sidewalk my feet get cold. And if I sit too long on those stone steps, my -- Guy has the door open and eases Hennessy toward the haIl. GUY (quickly) Don't worry. Since you told Barbara Burton about the icebox, you're her favorite charity. She'll send the butler out with something to defrost you. HENNESSY (grinning) Cute kid. Converted to PDF by www.screentalk.org 77. He's gone, and with a last glance at the dresser, Guy goes out and closes the door. LAP DISSOLVE TO: EXT. BURTON HOUSE LONG SHOT NIGHT The street outside the Burton house is lined with cars and limousines. Various guests are arriving. MED. SHOT On the opposite side of the street we see Hennessy, now wearing his topcoat. He looks bored as he glances across the street to the house. LAP DISSOLVE TO: INT. BURTON HOUSE BIG HEAD CLOSEUP OF ANNE Her face is troubled. CAMERA BEGINS TO PULL BACK. We see now that the reception is in progress and that Anne stands beside her father to greet the arriving guests. CAMERA PULLS BACK FURTHER to show us a full view of a very crowded Washington gathering Many white ties and tails and decollete in evidence. Many accents. Even some foreign languages are being spoken. Music and chatter in the b.g. CLOSE SHOT Anne and the Senator are still greeting new arrivals. Anne's manner is somewhat preoccupied. She glances around as she speaks, as though looking for someone. ANNE (to new arrival) Thank you so much, Mr. Lindsay. We'll look forward to it. PANNING SHOT FROM ANNE'S VIEWPOINT THE CAMERA PASSES various groups of guests in conversation including Guy and Barbara who are together. From this distance we cannot hear what they are saying. CAMERA CONTINUES TO the front door. It opens to admit a new arrival. It is Bruno. He wears white tie and tails, looking very elegant. Converted to PDF by www.screentalk.org 78. We see Guy excuse himself from Barbara, cross to Bruno and speak to him angrily, obviously asking, "What are you doing here?" Bruno, however, greets Guy with a smile then turns from him, unperturbed and bland. He sees Anne and moves toward her, smiling. CLOSEUP ANNE As Bruno comes in her direction, Anne's expression shows her mystification and concern about Bruno's presence and about Guy's attitude toward him. MED. SHOT Bruno comes up to Anne and the Senator. He gives a slight bow to the Senator; then puts his hand out to Anne. BRUNO Good evening, Miss Burton. The Senator looks inquiringly. Anne makes the introduction. ANNE This is Mr. Antony, father. SENATOR How do you do, sir. BRUNO I'd like to talk to you sometime, Senator, about my idea of harnessing the life force. It will make atomic power look like the horse and buggy. (the Senator and Anne are beginning to look at him in amazement) I'm already developing my faculty for seeing millions of miles. And, Senator, can you imagine being able to smell a flower on the planet of Mars? I'd like to lunch with you some day soon and tell you more about it. Interrupted by new arrivals, Bruno moves away out of the picture, with a charming smile to Anne. The Senator greets the new guests with open mouth and simply shakes their hands while glancing off in direction of the departing Bruno. Converted to PDF by www.screentalk.org 79. DOWAGER (to Senator) So nice to see you, my dear Senator. SENATOR Ah yes, indeed -- I beg your pardon? She realizes he hasn't heard a word she's said and haughtily moves on. The Senator turns to Anne. SENATOR (still looking after Bruno) I don't remember inviting that young man. Who is he? ANNE A friend of the Darvilles. SENATOR He has an unusual personality. Provocative. CLOSEUP ANNE She looks off in Bruno's direction extremely disturbed at this new aspect of the mysterious stranger. CLOSEUP GUY He is watching Bruno. MED. SHOT Guy sees Bruno join a group of several ladies who are seated on a settee and a couple of older men who are standing by. A waiter comes along with a tray of drinks. Bruno takes one. CLOSEUP BARBARA She comes from the same direction that Guy came. She stops short as she sees: MED. SHOT FROM BARBARA'S VIEWPOINT Bruno is now heartily joining in conversation with one of the elderly gentlemen. Converted to PDF by www.screentalk.org 80. CLOSE SHOT BRUNO AND GROUP Bruno talking to an elderly, dignified gentlemen. BRUNO But tell me, Judge, after you've sentenced a man to the chair, isn't it difficult to go and eat your dinner after that? JUDGE Young man, when a murderer is caught, he must be tried. When he is convicted, he must be sentenced. When he is sentenced to death, he must be executed. BRUNO Quite impersonal, isn't it, sir? JUDGE So it is. Besides, it doesn't happen every day. At this moment, Anne comes into the scene. She hesitates as she hears Bruno's answer. BRUNO So few murderers are caught. The Judge moves out of the way. Bruno smiles blandly at the ladies. One of them speaks to him. MRS. CUNNINGHAM Well, Mr. Antony, you seem very interested in the subject of murder. Anne looks more troubled, then moves on out of the scene. BRUNO No more than anyone else. No more than you, for instance. MRS. CUNNINGHAM Me? I'm not interested in murder. Bruno pulls up a chair to face the two woman on the settee, sits down, straddling the seat, to look at them over the back of the chair and settle down for a nice conversation. Converted to PDF by www.screentalk.org 81. BRUNO (his tone is teasing) Oh, come now, everyone's interested in that. Everyone would like to put someone out of the way. Now surely, Madame, you're not going to tell me that there hasn't been a time when you wanted to dispose of someone. Your husband, for instance. MRS. CUNNINGHAM (laughs) Good heavens, no! BRUNO (playfully) Ah ah! (shaking a finger at her) Are you sure? Do you mean to tell me there wasn't a tiny moment - when you'd been made really angry? And what did you say? MRS. CUNNINGHAM (squirms, giggling) Well... BRUNO There you are, you see! There you are! All right, now you're going -- to do a murder. How are you going to do it? This is the fascinating part -- how are you going to do it...I didn't get your name? MRS. CUNNINGHAM Mrs. Cunningham. BRUNO Mrs. Cunningham, how are you going to do it? MRS. CUNNINGHAM (entering into the spirit of the play) Well, I suppose I'll have to get a gun from somewhere. BRUNO (shakes his head) Tssk, tssk. Oh no, Mrs. Cunningham. (MORE) Converted to PDF by www.screentalk.org 82. BRUNO (CONT'D) Bang, bang, all over the place. Blood everywhere? The other woman joins in: MRS. ANDERSON What about a little poison? BRUNO Ah! That's better, that's better. Mrs.....? MRS. ANDERSON Anderson. BRUNO (he is thoroughly enjoying himself) That's better, Mrs. Anderson. But Mrs. Cunningham is in a dreadful hurry. Poison could take...let's see...ten to twelve weeks, if poor Mr. Cunningham is to die from natural causes. MRS. CUNNINGHAM I have a wonderful Idea! I can take him out in the car and when I get to a lonely spot, knock him on the head with a hammer, pour gasoline over him and over the car and start the whole thing ablaze. BRUNO (looks at her deprecatingly) And then have to walk all that way home? Mrs. Anderson laughs. BRUNO No, I have the best way, and the best tools. (he holds out his hands and shows them) Simple, silent, and quick. The silent part being the most important. Let me show you what I mean. (MORE) Converted to PDF by www.screentalk.org 83. BRUNO (CONT'D) (he raises his hands toward Mrs. Cunningham's throat, then stops a moment to ask) You don't mind if I borrow your neck for a moment do you? MRS. CUNNINGHAM (giggles) Well, it's not for long. BRUNO Oh! no. (he takes a drink and puts his glass down) Now, when I nod my head, just see if you can cry out, and I bet you can't. (he places his hands around Mrs. Cunningham's neck) Now with my two thumbs...you see that's where I'll be able to prevent any sound coming from you. Now, just wait for the nod of my head. CLOSEUP BRUNO As he starts to Press her neck, his eyes wander from the face of his "victim" to someone else off scene. MED. SHOT BARBARA She is watching this rather unorthodox demonstration. The CAMERA MOVES UP until her head fills the screen. Her glasses glint in the light. CLOSEUP BRUNO He is now transfixed. His breathing becomes heavy. A strange expression comes over his face. He still stares off at Barbara. MED. SHOT BARBARA We see the whirling merry-go-round spinning around her head. Converted to PDF by www.screentalk.org 84. BIG HEAD CLOSEUP BRUNO He now seem to have almost gone into a trance. Over the shot we begin to HEAR a strangled cry, and a broken exclamation, then Mrs. Anderson's voice. MRS. ANDERSON'S VOICE Mr. Antony! Mr. Antony! ANOTHER VOICE Stop him! Stop him! CLOSEUP Bruno's wrists and hands and the neck of his victim. We can just see Mrs. Cunningham's chin at the top of the screen. Her head is tossing from side to side. Her hands are clutching at Bruno's wrists. The hands of the other two women, also in the picture, are pulling at Bruno's wrists. Mrs. Cunningham's hands begin to slide off. Her head drops back. Over this we HEAR cries of: VOICES Stop him! Help, somebody! Pull him off! Mr. Antony! Mr. Antony! CLOSEUP BRUNO His body is swaying slightly at the various efforts to drag him away from Mrs. Cunningham. His eyes begin to close, and slowly he falls away from the picture in a dead faint on the floor. MEDIUM SHOT There is a rush of people around Mrs. Cunningham, who is breathing frantically, her eyes opening and closing. A couple of women are feebly slapping her hands, someone else is fanning her face. MEDIUM SHOT The Senator and Guy rush into the picture. They look at the fallen Bruno. They search around for an explanation. Other man come in ad they start to pick Bruno up. Converted to PDF by www.screentalk.org 85. GUY Bring him this way. Guy gives a quick look in direction of Mrs. Cunningham, sees that she is being attended to. MEDIUM SHOT Anne rushes into the picture. She sees Bruno being helped to his feet; then turns her attention to Mrs. Cunningham, who has now somewhat recovered. Mrs. Cunningham is helped to the settee. There is a babble of women's voices trying to explain what has happened. ANNE (thru the babble) Bring her upstairs. As the two groups pass off in different directions, the few people who ran into the scene late are asking the others what the disturbance is. "What's wrong?" "Did she faint?" "I didn't see anything." "What happened to him?" "Somebody hurt?" But one small figure stands in the clear. It is Barbara, She is still transfixed by what she has seen. Her hands are trembling. CAMERA MOVES SLOWLY IN ON HER. We see that her lips are trembling, too, and in her eyes frightened tears are welling. Her breath is heavy. INT. STUDY Bruno is stretched out on a settee. He is completely out. His collar and tie are open. Two or three of the male guests are just leaving the room. The Senator remains behind for a moment with Guy. SENATOR I thought he was a bit weird when he arrived. Who is he? GUY I hardly know him, sir. SENATOR Get him out of here as soon as you decently can -- will you. This is a nice item for the gossips. First thing you know, they'll be talking about orgies. I'd better get back... GUY Yes, sir. Converted to PDF by www.screentalk.org 86. The Senator leaves. Guy stands over Bruno's outstretched figure. MEDIUM SHOT Bruno is now half awake. Almost without seeing Guy, he staggers to his feet and begins to make his way to the door. Guy advances, and with a sharp thrust, pushes Bruno back on the settee. Bruno looks and sees Guy clearly for the first time. BRUNO What happened? I was on a merry-go- round somewhere. It made me dizzy. Guy moves forward, and thrusting his hand in Bruno's open shirt, pulls him to his feet. Bruno ignores Guy's violence and remain puzzled. GUY (disgusted) You're a mad, crazy maniac, and you ought to be locked-up! Now will you get out of here and let me alone? BRUNO But, Guy -- Guy smashes Bruno in the jaw, in utter disgust, and knocks him back onto the settee. Bruno looks up from his sprawled position, a dull look in his eye. BRUNO You shouldn't have done that, Guy. GUY (subsiding) Come on -- pull yourself together. Do your tie up. Bruno staggers to his feet. He fumbles at his collar. As he crosses to him, CAMERA MOVES IN to a CLOSER SHOT. GUY Here -- let me. He fixes Bruno's shirt and collar together and quickly ties his white bow. Bruno stands swaying like a small boy as Guy does this. Converted to PDF by www.screentalk.org 87. CAMERA PANS WITH THEM as Guy starts to escort Bruno from the room. GUY Have you got a car here? BRUNO (mumbling) Driver's outside. They pass trough door into the hallway. INT. HALL MEDIUM SHOT One or two of the guests turn their heads as Guy takes Bruno across to the front door. CLOSE SHOT Barbara appears in the hallway, coming from the crowded sitting room. She watches the two men go out the front door. MEDIUM SHOT Bruno and Guy going out the front door. The man-servant does not close it immediately, so we are able to HEAR the call for Mr. Antony's car. CLOSEUP BARBARA She turns her head and looks up the stairs. Barbara has not quite recovered from her ordeal. She hurries forward to greet Anne who is hurrying down the stairs. TWO SHOT CAMERA PANS DOWN with Anne as she descends the last few steps. Barbara enters the picture and the two girls meet at the foot of the stairs. ANNE What's the matter, Barbara? Did you see it happen? Did you see it -- all? Converted to PDF by www.screentalk.org 88. CLOSEUP BARBARA BARBARA (still shaken) He looked at me! His hands were on her throat, but, he was strangling me! CLOSEUP ANNE ANNE (aghast) How do you mean? TWO SHOT BARBARA He was looking at her first. Then he looked over at me. He went into a sort of trance (shudders) He looked horrible! (reflectively) He thought he was murdering me. CLOSEUP ANNE She looks away, with growing consciousness of the situation TWO SHOT BARBARA Anne, why me? Why me? What did I have to do with it? Anne is extremely concerned and thoughtful. Suddenly she gets an idea and with a pat on Barbara's arm, asks hurriedly: ANNE Do you know where Guy is? BARBARA He went out with that man! Anne hurries to the front door and passes through. Converted to PDF by www.screentalk.org 89. EXT. HOUSE Anne comes out onto the steps and looks around. She stops short as she sees: LONG SHOT EXT. STREET FROM ANNE'S VIEWPOINT There are cars lined up outside on the street. One limousine is pulling up in the center, two figures at the passenger door. One is climbing in. The other is Guy. CLOSEUP ANNE She calls out urgently: ANNE Guy! CLOSE SHOT Guy turns and closes the door. MEDIUM SHOT FROM ANNE'S VIEWPOINT The limousine moves off and Guy comes toward her. MEDIUM SHOT Anne comes down the steps and intercepts Guy on the sidewalk. She leads him along a few paces and then stops and faces him. CLOSE TWO SHOT Anne nods off in the direction of the departed Bruno and speaks in a desperate, low voice. ANNE You didn't meet him for the first time the other day, did you, Guy? Guy stares at her for a moment. GUY You mean when you introduced us at the club? Converted to PDF by www.screentalk.org 90. ANNE Yes. Did you notice how he stared at Barbara that day? GUY (awkwardly) Well, I didn't -- particularly -- ANNE (breaks in) He stared at her again tonight -- while his hands were around Mrs. Cunningham's throat. Guy looks at Anne with an expression of growing fear and alarm. She goes on inexorably: ANNE What did Miriam look like, Guy. GUY (awkwardly) Well, why do you ask me? You've seen her pictures in the paper. ANNE Go on, I want you to tell me. GUY (haltingly) Well, she was dark, not too tall, rather pretty -- ANNE What else? GUY What else is there? ANNE She wore glasses, didn't she? GUY Yes. ANNE She looked a lot like Barbara, didn't she? Guy suddenly begins to realize what Anne is getting at. Anne lowers her head, deliberately avoids looking at Guy, as she asks: Converted to PDF by www.screentalk.org 91. ANNE How did you get him to do it, Guy. GUY I get him to do it? ANNE He killed Miriam, didn't he? Tell me, Guy! GUY Yes. (suddenly bursting out) He's a maniac. I met him on the train going to Metcalf. He had a crazy scheme about exchanging murders. I do his murder and he do mine. ANNE (quietly) What do you mean -- your murder, Guy? GUY Well, he'd read about me in the paper. He knew about Miriam -- and about you. He suggested that if he got rid of Miriam for me, I should kill his father. ANNE You must have realized he was talking a lot of nonsense! GUY Of course! I didn't give it another thought. And now a lunatic wants me to kill his father. ANNE (beginning to believe) It's too fantastic! GUY (grimly) Yes, isn't it? ANNE You mean you've known about Miriam all this time? Converted to PDF by www.screentalk.org 92. GUY Since the first night. He gave me her glasses. ANNE Why didn't you call the police? GUY (bitterly) And have them say what you did -- "Mr. Haines, how did you get him to do it?" And Bruno would say we'd planed it together. ANNE Oh, Guy -- what can we do? GUY I don't know, Anne...I don't know. ANNE (With an anxious look across the street) Guy, hadn't we better go inside? Your friend Hennessy's watching us. (she Shudders) GUY (sadly) You see, Anne, that's why I didn't want you to know anything about this. I wanted to protect all of you -- your father, Barbara. And now that you know, you're acting guilty, too. ANNE (desperately) Oh, if we could only talk to father or someone about it. GUY No, that's no good, Anne. I mustn't drag anyone else into this mess. Come on. Let's go in. They go toward the house. CUT TO: Converted to PDF by www.screentalk.org 93. TWO SHOT ACROSS THE STREET As Hennessy watches Anne and Guy go toward the house, his relief, HAMMOND, comes up. Hammond's a zealous, hard-eyed sleuth. HENNESSY (a little glum) Hello, Hammond. HAMMOND You look worried. What's the matter? HENNESSY You'd better keep on your toes. Something funny's going on. DISSOLVE TO: INT. GUY'S APARTMENT LATER THAT NIGHT Still in his dinner clothes, Guy is seated in deep thought near the telephone, wrestling with his problem. There is an open telephone directory in front of him. He comes to a decision, picks up the telephone and dials a number. He waits for the answer, then: GUY Bruno? Yes, yes, it's Guy...I've decided to do what you want. I'll make that little visit to father.... (listens a moment) Tonight. (listens another moment) Yes, I want to get this thing over with, can you leave the house again, Bruno? (pause) You'd better stay out till daylight. Guy hangs up, rises and starts to move with purpose for his night's activities. DISSOLVE TO: INT. GUY'S APARTMENT NIGHT Guy is sitting at the table. He is dressed differently, having changed from his dinner clothes to a sack suit. There is only one lamp lighted in the room. Guy presents a grim picture. Converted to PDF by www.screentalk.org 94. He is studying the plan of Bruno's house, and he picks up the key Bruno sent along with it. Finally he looks at his watch, then folds the plan and puts it in his pocket with the key. He rises, crosses to the chest of drawers, opens the top drawer. INSERT: THE OPEN DRAWER Guy's hands take out the Luger. His hand then picks up Miriam's glasses from the drawer, holds them a moment. He is about to put them back, then decides to take them along, puts them into his pocket. MED. SHOT CAMERA PANS GUY across to the window. He parts the curtains slightly and looks out. MED. SHOT ON STREET (FROM GUY'S VIEWPOINT) Hammond is lighting a cigarette as he strolls in front of the house. INT. GUY'S APARTMENT Guy crosses to his door, which he opens surreptitiously. MED. SHOT CORRIDOR Guy glances down the stairs, then closes the door behind him quietly and moves away to a window at the turn of the stairs. EXT. FIRE ESCAPE Guy comes out of the window onto the second floor fire escape. He creeps stealthily down and emerges into a narrow alleyway. He steps back into the shadows for a moment when he sees: LONG SHOT FROM GUY'S VIEWPOINT (PROCESS) The strolling figure of Hammond on the far side of the street. Converted to PDF by www.screentalk.org 95. MED. SHOT Guy turns away and is soon lost in the darkness of the street. LAP DISSOLVE TO: EXT. A TALL PAIR OF ELABORATE IRON GATES NIGHT We are on the inside of the gates. We see them swing open slightly and the figure of Guy edges through them. CLOSE SHOT Guy leaves the gates ajar and then, taking the plan of Bruno's house from his pocket, and the key, he looks toward the house. EXT. STEPS LONG SHOT NIGHT This is a long flight of steps. Moonlit. They are lined with tall black cypress trees which throw their shadows across the steps. Guy moves out of one shadow, into another and carefully starts up the stairs. AT THE DOOR He pauses, looks about for a moment and listens. Then he puts the key into the lock, finding it with his flashlight. The door opens a few inches. He turns off the flash, and enters. INT. ANTONY HOME ENTRANCE HALL As Guy moves in soundlessly and closes the door. He looks toward the stairs which are in shadow. MED. SHOT Guy starts up the stairs slowly. He carries his flashlight and the plan. AT THE TOP OF THE STAIRS THE DOG A huge shadow lies it the head of the stairs. As Guy comes slowly up the stairs, the Great Dane looks down at him. Converted to PDF by www.screentalk.org 96. GUY ON THE STAIRS He reacts to the sight of the dog, stops an instant, and turn on his flashlight. The heavy massive face of the dog looks straight down at him. Guy turns off the flashlight and after a moment of indecision starts slowly up the stairs once more, the dog watching every step he takes. UPPER HALLWAY Guy comes up the last few stairs and still the dog hasn't moved. Guy slowly edges past him and the Great Dane's head turns to watch him. GUY moving quietly along the hallway, approaches two doors. He takes out his flash and identifies the door with his plan. INSERT: The plan shows two doors in relation to the stairway. The first one is clearly marked: "MY room." The adjoining door is marked: "My FATHER'S room." CLOSE SHOT GUY He pauses at the first door, then passes it quietly, walking on to the next one. He turns the knob soundlessly and passes through into the room. INT. ANTONY BEDROOM LONG SHOT The room is in darkness except for the dim outline of the recumbent figure in the bed. We hear Guy's voice, in a loud whisper: GUY Mr. Antony! The figure stirs. ANOTHER ANGLE Guy takes a stop closer to the bed. Converted to PDF by www.screentalk.org 97. GUY (urgently) Mr. Antony! Don't be alarmed -- but I must talk to you about your son. About Bruno. Mr. Antony! The figure on the bed turns and a hand stretches out toward a bedside light. The light goes on with a sudden glare. CLOSEUP FACE OF BRUNO IN THE LIGHT (LOW CAMERA) The low CAMERA throws a vast shadow up on the wall behind him, creating a grimace of his smile. BRUNO Yes, Mr. Haines? CLOSEUP GUY His face is dead. MED. SHOT Bruno rises from the bed and sits on the and of it. He is fully dressed, just as he was at the party, in white tie and tails. BRUNO (politely) My father isn't home tonight, Mr. Haines. (smiles grimly at Guy's surprise) I was about to tell you that over the phone. But you came to such a sudden decision. I wondered why. GUY (recovering quickly) Since you sent me a key to your house, I decided to use it -- to make a little social call on your father. I thought he'd be Interested to know he his a lunatic son. The faintest flicker of Bruno's eyes indicates the intensity of his reaction. He stares hard at Guy. Converted to PDF by www.screentalk.org 98. BRUNO Then a I correct, Mr. Haines, in assuming that you have no intention of going ahead with our arrangement? GUY No intention whatsoever. I never had. BRUNO I see. You won't have any further use for the key, then, Mr. Haines. (he holds out his hand and Guy gives him the key) Thank you very such. As Bruno continues to stare at him, Guy takes out the Luger. For a moment a look of fear comes into Bruno's face as he thinks Guy will probably shoot him. After a pause, Guy tosses the gun on the bed. GUY Or this. Bruno's relief turns again to menace. He picks up the gun and fingers it nervously. GUY (kindly) Look, Bruno. You're terribly sick. (haltingly) I don't know whether it's possible for you to realize it or not. I don't know much about these things, Bruno. But why don't you go someplace where you can get some treatment? Not only for your own sake, Bruno, but you can't go on causing more and more destruction to anyone you happen to meet. Bruno pays no attention. He rises. TWO SHOT Guy's arguments have made no impression on Bruno whatsoever. He fingers the gun. BRUNO I don't like to be doublecrossed. (MORE) Converted to PDF by www.screentalk.org 99. BRUNO (CONT'D) I have a murder on my conscience, but it's not my murder, Mr. Haines -- it's yours. And as you're the one to profit, I think you should be the one to pay for it. For an instant his nervous hands seem to be struggling with the urge to kill Guy. GUY (gives up) Well, I guess it's no use, Bruno. We sees to have nothing further to discuss. Bruno goes to the door in silent acquiescence and opens it for Guy to pass through. INT. HALLWAY MED. SHOT Guy walks toward the stairs, tense and apprehensive. Bruno is following him, still holding the gun. When the Great Dane sees Bruno it gets to its feet, as if waiting for a command. Guy starts down the stairs but Bruno stays where he is, the dog beside him. Gay turns and looks back it this tableaux of menace. BRUNO Don't worry. I'm not going to shoot you, Mr. Haines. It might disturb mother. (with a feeling of power) I'm a very clever follow. I'll think of something better than that. Much better. LONG SHOT Bruno remains in the foreground of the scene as Guy proceeds on down the stairs. We see him open the front door and pass through. LAP DISSOLVE TO: Converted to PDF by www.screentalk.org 100. EXT. STREET ACROSS GUY'S APARTMENT EARLY CLOSE SHOT HENNESSY AND HAMMOND MORNING Hennessy is relieving Hammond who has kept watch on Guy's apartment night. HAMMOND (in the middle of his story) He came back at three twenty-five. I didn't even know he'd given me the slip until his 'phone kept ringing for about half an hour. Nobody sleeps that sound. So I got the janitor to let me in. No Haines. HENNESSY (to himself) Wonder where he went? HAMMOND We'll probably hear of another dame murdered. HENNESSY (puzzled) Shut up. I'd better contact Metcalf. I should think this calls for more questioning of Mr. Haines. HAMMOND Questioning? Nuts! Let's take him in. HENNESSY My dear Mr. Hamond, how many times do I have to tell you that we have nothing conclusive on Haines? There's no evidence that he was ever at the scene of the crime. Can't you get that into your thick head? (quietly) Now stay put till I get back. As he starts away -- FADE OUT. Converted to PDF by www.screentalk.org 101. FADE IN INT. ANTONY LIVING ROOM LATE MORNING Anne and Mrs. Antony are in the middle of a conversation. Anne's manner is tense and purposeful, Mrs. Antony's much less serious. MRS. ANTONY Oh, now, Miss Burton, really! I know Bruno's been in some very awkward scrapes, but nothing so ridiculous as a murder. (she gives a short little laugh) ANNE (desperately) But, Mrs. Antony, you've got to make him do something about this. Don't you see that just one word from him would extricate Guy from this dreadful situation? MRS. ANTONY (lightly) Oh, but Miss Burton, I'm sure this thing must be some practical joke. You know, Bruno sometimes goes too far. (girl to girl) Of course I shouldn't be saying this to an outsider, but sometimes he's terribly irresponsible and gets into all kinds of escapades. ANNE But don't you understand, Mrs. Antony -- your son is responsible for a woman's death. MRS. ANTONY (drawing herself up with some hauteur) Did Bruno tell you this? ANNE Of course not, Mrs. Antony. MRS. ANTONY (that settles it) Well, there you are. (MORE) Converted to PDF by www.screentalk.org 102. MRS. ANTONY (CONT'D) (she sighs and rises, winding it up) Well, Miss Burton, it was very nice of you to call. You must excuse me now. I must get back to my painting. Do you care for painting, Miss Burton? I find it so soothing. (shakes Anne's hand) You must come again sometime. She goes out. Anne is left helpless, standing in the middle of the room. She picks up her purse and is about to go when she hears a voice: BRUNO'S VOICE Oh, Miss Burton! Anne turns back in direction of the voice. CAMERA PULLS BACK until we can see the feet of Bruno protruding from behind a chair in which he is sitting. He has obviously heard the entire conversation between Anne and his mother. Bruno rises. He is in dressing gown and pajamas. BRUNO I'm afraid mother wasn't very helpful, was she? (he strolls toward Anne) You know she hasn't been well for a long time. She's a little -- how shall I say -- confused. (shakes his head commiseratingly) Poor mother. Anne is too stunned to speak. BRUNO You know, I'm very upset with Guy. He shouldn't have sent you on an errand like this. ANNE Guy doesn't know I'm here, Mr. Antony. BRUNO He's been leading you up the garden path, I'm afraid. He must be very desperate to try to involve me. I've been protecting him ever since we had that conversation on the train and he told me how he hated his wife. Converted to PDF by www.screentalk.org 103. Bruno is now standing near the window a little apart from Anne, with his back to him. He takes something out of the pocket of his dressing gown and looks down at it in his hand. It is Guy's lighter. Suddenly he stuffs it back his pocket and turn back to Anne. BRUNO Why, do you know, Miss Burton, he tried to get me to go back to the island one night after dark and pick up his lighter so the police wouldn't find it? He dropped it there, you know, when -- well, that night. Anne's horror is growing. BRUNO The whole thing's been worrying me so much. But of course I couldn't do it, Miss Burton. It would have been too risky. And besides, it would have made me an accessory. Anne stares at this insane man and sinks on the settee. She starts to cry in sheer frustration. Bruno goes to her sympathetically. BRUNO Miss Burton, I know how you feel. He puts his hand on her shoulder. Anne flings it off. There is an awkward pause as Bruno looks down at her. Then he begins to look around restlessly. BRUNO Miss Burton, you must excuse me. I have an urgent appointment. (looks it his watch) I must go up and change. Now, I really must go...if you'll excuse me... He turns, starts out of the room and up the stairs in the hall. Anne watches him. STAIRWAY FROM ANNE'S VIEWPOINT Bruno turns and waves to Anne from the landing, then goes on up the stairs. Converted to PDF by www.screentalk.org 104. INT. LIVING ROOM MED. SHOT Anne slowly rises, a lonely figure in the large room, and makes her way out. DISSOLVE TO: LONG SHOT FOREST HILL STADIUM Grouped. A game is in progress. MED. SHOT A TERRACE NEAR THE MAIN STADIUM (PROCESS) where people get refreshments. There are various table with umbrellas. MED. SHOT AT ONE OF TABLE (PROCESS) Anne and Guy are seated at the table. ANNE ...And he said that if the police found your lighter there, that's all they'd need -- something to prove you were at the scene of the murder. GUY (grimly) That big lie about my wanting him to get it back means he's going to put my lighter on that island! ANNE (urgently) Guy, you'll have to get there before he does. You won't have time to play, You'd better tell them. (she nods her head in the direction of the center court) GUY Darling, if that loudspeaker announces that I'm not going to play, Hennessy bound to be suspicious He'd keep me from ever getting near Metcalf. ANNE Then I'll go. Converted to PDF by www.screentalk.org 105. GUY (quickly) No, darling. (he puts his hand on hers and speaks firmly, with concern for her safety as well as for his own situation) You stay right here and help me give Hennessy the slip after the match. ANNE But, Guy, that'll be too late! GUY (getting a thought) Didn't Bruno say that I wanted him together there one night after dark? ANNE Yes. GUY Well, that's what's in his mind now. He's not going to expose himself in broad daylight, If I can finish off this match in three sets, I'll still get there in time. REYNOLDS, Guy's opponent, enters scene behind Guy's chair. REYNOLDS We're on in a few minutes, Guy. (to Anne) How are you, Miss Morton. Anne acknowledges his greeting with a nod. GUY Okay, Tim. Be right with you. Reynolds leaves Anne and Guy rise, and as they walk toward the stadium, we can see Guy start to speak to Anne in a whisper. ENTRANCE TO COVERED STAND ALREADY SHOT Hennessy and Hammond are standing by. Converted to PDF by www.screentalk.org 106. HAMMOND Well, if Turley said to pick him up for questioning, let's pick him up! HENNESSY Let him have his game first, Hammond. HAMMOND (sourly) This is the first time I ever waited for a murder suspect to play tennis before I pulled him in. When the boys it headquarters heir about this they'll send me orchids. Guy and Anne come into the scene just as the players from the previous match emerge. They pass through, nodding to Hennessy. HENNESSY Good luck, Guy. Guy is so preoccupied with his grim doesn't nod to Hennessy until Anne nudges him. INSIDE THE STAND MED. SHOT Anne is reluctant to leave Guy who must now join his opponent, Reynolds. GUY You got it straight? (ANNE nods) Just make sure Barbara has everything ready as soon as the third set starts. He goes on to the court, and Anna goes to her box. MED. SHOT Anne joins Barbara in the box. She starts to whisper something to her. LONG SHOT Guy and Reynolds complete their warm-up as the umpire announces that Guy is to serve. The game starts. Converted to PDF by www.screentalk.org 107. EXT. ANTONY HOME A taxi is at the front door. Bruno is descending the steps. He gets into the cab, which moves off. FOREST HILLS MED. SHOT ANNOUNCER'S BOOTH (PROCESS) Over the shoulder of the announcer WE SEE the game in progress through the window of his booth. ANNOUNCER --It looks like an interesting match with Haines constantly charging the net -- not like Haines at all -- to press so early in the game... MED. SHOT TEN COURT Guy and his opponent, Reynolds, in play. Guy scores a point. CLOSEUP THE UMPIRE He announces game to Haines. MED. LONG SHOT We see the two men change ends and come toward the Umpiri's chair. Reynolds stops to take a drink of water. Guy, with an impatient glance at him, moves over to the passing line and waits, the CAMERA going with him. EXT. WASHINGTON STREET A taxicab is seen coming along. MED. SHOT INSIDE CAB (PROCESS) Bruno is sitting with in unlighted cigarette in his mouth. CAMERA MOVES IN until he is in big CLOSEUP. His eyes look down. There is the SOUND of a click, then, Guy's lighter comes up into the picture held against the cigarette. LAP DISSOLVE TO: Converted to PDF by www.screentalk.org 108. FADE IN LONG SHOT FOREST HILLS STADIUM Grouped. A game is in progress. MED. SHOT A terrace where people get refreshments. There are various tables with umbrellas. MED. SHOT AT ONE OF TABLE (PROCESS) Guy is seated. He has his rackets with him and is waiting his turn to start his match. An official is talking to him but Guy keeps looking around as if expecting someone. OFFICIAL Well, at least there'd be a trip to Australia, if you made it. GUY (absently) We'll know more about that by the end of the week... (his face brightens as he sees Anne) Anne hurries in, nods briefly to the official who has started to leave, and sits down. OFFICIAL They're close to the finish, Guy GUY Be right there. (turns to Anne) I was afraid you wouldn't get here. Wish me luck, darling. He makes a move as if to follow official toward the stadium, but Anne puts hand on his arm. ANNE (quickly and urgently) Guy, listen to me, If I sound all mixed up I can't help it. I -- I'm scared. GUY What about? Converted to PDF by www.screentalk.org 109. ANNE That's just it. I don't know. It's Bruno. I talked to him, Guy -- Guy stares at her, takes a quick look toward the stadium, then gives Anne his full attention. ANNE He acted peculiar -- as if he could put the murder right in your lap, and not involve himself at all. GUY (shaking his head) He'd drag himself into it, -- and Bruno loves Bruno. I'm all right so long as he thinks I have an alibi for that night. (noticing the stricken look on Anne is face) He knows? Anne nods slowly. GUY (grimly) Then he'll think of something. He said he would. ANNE Guy, has he anything that the police could trace to you -- (quoting Bruno) Any little thing. GUY My cigarette lighter. He said once he could have left it on the islands as evidence (a pause) But he wouldn't do that. Not in broad day light. ANNE (trying to think) But he's going somewhere, Guy. He told his mother -- GUY (tensely) Metcalf -- did he say Metcalf? Converted to PDF by www.screentalk.org 110. ANNE No, -- I don't think so. Oh, why can't I remember -- he said such crazy things! GUY (tensely) Try to think, Anne! VOICE (OFFSCENE) Guy Haines! -- Reynolds! While Anne is frantically trying to remember, Guy turns toward, the stadium and gives a signal of "Be right there." ANNE Something about the moon -- he said he had an appointment with the moon. Guy's shoulders droop with disappointment. GUY That's no help. But I can't take any chances. I've got to get that lighter -- somehow. REYNOLDS, Guy's opponent, ENTERS SCENE behind Guy's chair. REYNOLDS Okay, Guy. We're on. He walks away. Anne and Guy rise, following him. GUY I'll have to default. ANNE And have Hennessy and that other one right at your heels? Guy's expression says she's right, as they walk toward the stadium. ENTRANCE TO COVERED STAND Hennessy and Hammond, the two detectives, are standing by. HAMMOND First time I ever waited for a killer to play tennis before I nabbed him! (MORE) Converted to PDF by www.screentalk.org 111. HAMMOND (CONT'D) When the boys at headquarters hear about this they'll send me an orchid! HENNESSY We got our orders. We take him in -- after the match. Guy and Anne come INTO THE SCENE just as the players from the previous match emerge. They pass through, nodding to Hennessy. HENNESSY (a little sadly) Good luck, Guy! Guy gives him a thank-you nod. Hammond rolls his eyes in disgust at Hennessy's politeness. INSIDE THE STAND MED. SHOT Anne is about to turn to her box but she is reluctant to leave Guy, who must now join his opponent, Reynolds. As their eyes hold, in mutual helplessness, Guy suddenly stares at her with realization. GUY The moon! You said he had an appointment -- Anne looks puzzled as Guy looks up at the sun, then at his watch. GUY Then he is going to Metcalf. But he has to wait until it gets dark -- (with frantic haste, he thinks quickly, then murmurs to Anne) Listen, Anne, as soon as the third set starts, tell Barbara -- MED. CLOSE SHOT REYNOLDS waiting at the bottom of steps to the stand. Guy joins his opponent, and Anne goes to her box. Guy and Reynolds move onto the court amid the rounds of applause that greet them. Converted to PDF by www.screentalk.org 112. MEDIUM SHOT ANNE JOINS BARBARA In the box. She starts to whisper something to her. LONG SHOT Guy and Reynolds complete their warm-up as the umpire announces that Guy is to serve. The game starts. EXT. ANTONY HOME A taxi is at the front door. Bruno is descending the steps. He gets into the cab, which moves off. FOREST HILLS MED. SHOT ANNOUNCER'S BOOTH Over the shoulder of the announcer WE SEE the game in progress through the window of his booth. ANNOUNCER It looks like an interesting match -- with Haines out to blast Reynolds into a fast fight, -- not like Haines at all -- to press so early in the game... MED. SHOT THE COURT Guy and his opponent, Reynolds, in play. Guy scores a point. CLOSEUP THE UMPIRE He announces game to Haines. MED. LONG SHOT We see the two men change ends and come toward the Umpire's chair. Reynolds stops to take a drink of Water. Guy, with an impatient glance it him, moves over to the passing line and waits, the CAMERA going with him. EXT. WASHINGTON STREET A taxicab is seen coming along. Converted to PDF by www.screentalk.org 113. MED. SHOT INSIDE TAXI CAB Bruno is sitting with an unlighted cigarette in his mouth. CAMERA MOVES IN until he is in big CLOSEUP. His eyes look down. There is the SOUND of a click, then Guy's lighter comes up into the picture held against the cigarette. LAP DISSOLVE TO: INT. ANNOUNCER'S BOOTH FOREST HILL The announcer is broadcasting the progress of the match and we learn from him that this first set is nearly finished. LONG SHOT THE COURT Guy and Reynolds in play. MED. SHOT Anne and Barbara sitting in their box watching the play anxiously. MED. SHOT At the entrance to the covered stand. The two detectives Hennessy and Hammond, are watching. Hammond is bored by this game. HAMMOND Stupid game. You'd never get me into them short pants. I'd feel naked. HENNESSY (his eyes intent on the game) You'd feel naked in an Eskimo suit -- if you weren't wearing your badge. MED. SHOT Guy playing hard but holding his own. MED. SHOT Reynolds, his opponent, playing back. Converted to PDF by www.screentalk.org 114. LONG SHOT The big crowd watching. MED. SHOT Guy scores point over Reynolds. MED. SHOT There is general applause from the crowd in the covered stand as we HEAR the Umpire's announcement. UMPIRE'S VOICE (O.S.) Mr. Haines wins the first set. EXT. UNION STATION WASHINGTON D.C. We see Bruno get out of a cab and pass into the depot. LONG SHOT FOREST HILLS The game in process. MED. SHOT A nearer view of the game. CLOSE SHOT GUY IN PLAY volleying with Reynolds. CLOSE SHOT Reynolds playing the covered stand people are concentrating. MED. SHOT Guy misses a point and the game. He and Reynolds make for the Umpire's chair. We HEAR the Umpire announce. UMPIRE'S VOICE Game to Mr. Reynolds. Games are two all...Second set. Converted to PDF by www.screentalk.org 115. INT. UNION STATION WASHINGTON, D.C. Bruno is casually waiting for the train. He stands near a news-stand reading a paper. INSERT: We see that the paper is open at the sports page. There is a picture of Guy among other tennis players. WITH A DISSOLVE the whole character of this page changes with the exception of Guy's picture, which becomes surrounded with large type, announcing the arrest of Guy Haines for the murder of his wife Miriam. A sub-heading tells of Guy's cigarette lighter found at the scene of the crime. All this DISSOLVES AWAY and the page becomes once more the sports section. CLOSEUP Bruno looks up with satisfaction. LONG SHOT FOREST HILLS The crowd watching. MED. SHOT Guy and Reynolds in play. MED. SHOT Guy playing hard. MED. SHOT Reynolds playing back. CLOSEUP The Umpire watching the game. Suddenly he announces: UMPIRE Game to Mr. Reynolds. Games are three all... second set. Converted to PDF by www.screentalk.org 116. INT. CLUB CAR ON TRAIN Bruno is now seated in his accustomed place in the club car. His gloved fingers are quietly toying with Guy's lighter. A passenger next to him asks: PASSENGER May I have a light, please? Bruno looks at him for a moment and then at the lighter. With great deliberation he puts the lighter away in his pocket and takes out book-matches. Lighting a match, he holds it to his fellow passenger's cigarette. LONG SHOT FOREST HILLS The game as seen from under the covered stand. MED. SHOT Anne and Barbara very tense. CLOSEUP GUY about to serve, looks anxiously across the court. CLOSEUP THE CLOCK CLOSEUP GUY as he serves. CLOSEUP REYNOLDS returns. CLOSEUP BALL hits the net. CLOSEUP UMPIRE announces. Converted to PDF by www.screentalk.org 117. UMPIRE Second set to Haines. Haines leads two sets to love. There is a round of applause. We see the heads of the two players reach the Umpire's chair. Guy, very anxious still, as he wipes his neck with a towel. INT. COVERED STAND CLOSE SHOT ANNE BARBARA Anne is speaking. ANNE If he wins this next set -- you'd better have everything ready. (takes bill from her purse and hands it to Barbara) Here -- give the driver this ten dollars. BARBARA (puzzled) I wish understood what this is all about! ANNE (urgently) You don't have to understand, just do it. And for heaven's sake, act natural. Barbara nods and goes along. ENTRANCE TO COVERED STAND Barbara smiles winningly at Hennessy as she goes through. Her interpretation of "acting natural" is exaggerated and rather comical. Hammond's eyes narrow as he looks after her suspiciously. LONG SHOT The game in progress. Guy starts the next set. He serves. MED. SHOT Reynolds returns. Converted to PDF by www.screentalk.org 118. MED. SHOT Guy volleys. MED. SHOT Reynolds puts the ball in the air. CLOSE SHOT Guy smashes. CLOSE SHOT The ball hits the net. CLOSEUP UMPIRE UMPIRE Love fifteen. LONG SHOT THE CROWD We HEAR the smash of the ball and the voice of the Umpire. UMPIRE'S VOICE (O.S.) Love thirty. CLOSEUP ANNE looking very worried. Again the call of the Umpire. UMPIRE'S VOICE (O.S.) Double fault. Love forty. INT. THE ANNOUNCER'S BOOTH The announcer telling his listeners that Guy Haines seems to be a little reckless. ANNOUNCER -- Haines hasn't let up his terrific pace for an instant, smashing every (MORE) Converted to PDF by www.screentalk.org 119. ANNOUNCER (CONT'D) ball with a recklessness we've never seen in his playing. It's beginning to look as if he doesn't care whether he wins or loses because he's in a hurry - an awfully big hurry --- LAP DISSOLVE TO: EXT. METCALF STATION We see Bruno alight from the train. He makes his way in the direction of the town. MED. SHOT METCALF STATION As Bruno comes toward us, he stands on the sidewalk and then takes the lighter from his pocket once more. At this moment a hurrying passenger on his way to the depot accidentally jogs Bruno's elbow. The lighter flies from his hand. CLOSE SHOT We see it fall through the bars of a grating by the sidewalk. CLOSEUP BRUNO looks down in dismay. FOREST HILLS MED. SHOT The game in progress. Guy and his opponent playing hard. Guy misses a point. We HEAR the Umpire's call. UMPIRE'S VOICE (O.S) Game to Mr. Reynolds. Mr. Reynolds leads five games to three in the third set. EXT. METCALF STATION Bruno is leading a porter toward the grating, pulling him by the arm. They reach the drain. Converted to PDF by www.screentalk.org 120. BRUNO Down there -- my -- my cigarette -- (catches himself -- not wanting to say "cigarette lighter") case. It's very valuable. PORTER (peering down) Down here? BRUNO You've got to get this grating up right away. Two passersby enter. FIRST PASSERBY What's the trouble? BRUNO (yelling) Can't we do something...! (to passerby) I dropped my cigarette case. PORTER (looking down) Mightn't be any good, mister. Probably gone down the storm drain. BRUNO (horrified) Storm drain? FIRST PASSERBY On the other hand, it might have lodged on the edge. SECOND PASSERBY Don't they have a trap down there -- like under a sink? BRUNO (excited) Don't just stand here -- do something! PORTER (calmly) Guess we could phone the city engineer, all right. (MORE) Converted to PDF by www.screentalk.org 121. PORTER (CONT'D) Worst he could do would be to tell me to take a running jump and -- (Bruno grabs his arm. Porter shakes Bruno off) Relax, mister. BRUNO I don't want to relax. He goes on his knees and forces his arm down the drain. INT. THE ANNOUNCER'S BOOTH FOREST HILLS ANNOUNCER (with great excitement) This is more than a tennis game, ladies and gentlemen -- it's a desperate fight with Guy Haines playing as if his life depended on it! MED. SHOT Guy is volleying. MED. SHOT Reynolds lobs. CLOSEUP Guy smashes. CLOSE SHOT Reynolds lobs again. CLOSE SHOT Guy smashes. CLOSE SHOT Reynolds misses and the ball hits inside the line. Converted to PDF by www.screentalk.org 122. CLOSEUP The Umpire calling. UMPIRE Game to Mr. Haines. Mr. Reynolds leads five games to four...third set. EXT. METCALF STATION MED. SHOT A few more passersby have stopped to watch Bruno, whose arm is pushed through the grating. CLOSEUP Bruno's face -- straining. CLOSEUP Under the grating Bruno's hand is groping. His fingers are a long way from the lighter. LONG SHOT FOREST HILLS with the game in progress. MED. SHOT EXT. CLUB A taxi has pulled up. Barbara gets out. CLOSE SHOT She takes the ten dollar bill from her purse and passes it to the driver. She gives a final look inside the cab. CLOSEUP On the seat are Guy's everyday pants, laid out. MED. SHOT Barbara hurries out of the picture toward the club. Converted to PDF by www.screentalk.org 123. LONG SHOT The crowd watching. CLOSEUP The tense face of Anne. CLOSEUP The Umpire is somewhat impressed. INT. THE ANNOUNCER'S BOOTH CLOSEUP The announcer is telling his listeners that the score is now six-five in favor of Haines. That he has pulled up wonderfully and only needs one more game to win the match. EXT. COVERED STAND ENTRANCE Barbara, very nervous but trying to "act natural", passes Hennessy and Hammond. Hammond's eyes again follow her, but Hennessy is intent on the game. MED. SHOT FEATURING BOX As Barbara joins Anne, she gives her a surreptitious signal by ringing her thumb and forefinger, indicating everything is set. CLOSE SHOT Guy now playing hard. CLOSEUP His racket smashing at the ball. CLOSEUP Reynolds and his racket hitting the ball back. Converted to PDF by www.screentalk.org 124. CLOSEUP THE UMPIRE CALLING UMPIRE Advantages, Mr. Haines. CLOSEUP Guy serving. CLOSEUP His ball hitting the racket. CLOSEUP The ball in the net. CLOSEUP A second ball hitting the net. The Umpire's voice calling: UMPIRE'S VOICE (O.S) Duece! EXT. METCALF STATION A LOW SHOT ON Bruno bent over the grating and the onlookers behind him. BIG HEAD CLOSEUP BRUNO straining and panicky. CLOSEUP Under the grating, Bruno's fingers get near the lighter, and in their groping, they knock the lighter off the ledge, onto the ledge below. CLOSEUP Bruno's horror-stricken face. Converted to PDF by www.screentalk.org 125. FOREST HILLS MED. SHOT Guy still playing. CLOSE SHOT Barbara standing with Hennessy, watching. We HEAR the score. UMPIRE'S VOICE (O.S) Advantage, Mr. Reynolds. CLOSEUP ANNE unable to bear the suspense. She glances O.S. MED. SHOT The waiting cab. CLOSE SHOT Guy and Reynolds in play. UMPIRE'S VOICE (O.S) Score is deuce. CLOSE SHOT Reynolds serves. CLOSE SHOT Guy volleys. He waits for the return ball. He misses it. UMPIRE'S VOICE (O.S) Advantage, Mr. Reynolds. EXT. METCALF STATION ANGLE SHOOTING THROUGH the grating at CLOSEUP BRUNO'S HEAD AND SHOULDERS staining. Converted to PDF by www.screentalk.org 126. CLOSEUP Under the grating, Bruno's fingers go lower and lower, straining to reach the lighter, which is still a few inches out of reach. FOREST HILLS MED. SHOT Guy is volleying with Reynolds. INT. ANNOUNCER'S BOOTH He is very excited. ANNOUNCER -- Haines hasn't let up for a moment. If he wins this set, he wins the whole match! CLOSEUP ANNE AND BARBARA in their box. They are extremely tense. MED. SHOT Guy slams hard a shot that wins him the game. LONG SHOT CROWD applauding and shouting. CLOSE SHOT ANNE AND BARBARA At an urgent signal from Anne, Barbara hurries out as if she knew what she had to do. LONG SHOT Guy shakes hands with his opponent, and then hurries across to Anne in the stand. He leans over the front of the box. While congratulating him outwardly, she whispers something to him. He leaves his racket with her and hurries away. Converted to PDF by www.screentalk.org 127. MED. SHOT STAND ENTRANCE A block of people leaving cut off Hennessy's view. Barbara tries desperately to keep his attention off Guy. BARBARA (breathlessly) Isn't it wonderful, Mr. Hennessy? He won! It calls for a celebration. Anne says you must have dinner with us. Just the family, and you, and Guy. HENNESSY (awkwardly) Sorry I can't make it. Business. BARBARA But Guy is your business. You'll be with him, won't you? HENNESSY (a little grimly) Yeah -- I'll be with Guy. MED. SHOT Guy moving along the front of the stand making for another exit. CLOSE SHOT Barbara takes it for granted that Hennessy will accept her invitation. BARBARA Guy says you love steak -- rare, Medium, or well-done? HENNESSY I sure wish I could -- SEMI CLOSEUP Hammond is looking off. He calls into the stand. HAMMOND Hennessy! He points off toward Guy. Converted to PDF by www.screentalk.org 128. MED. SHOT Guy is hurrying toward the public entrance of the stand. SEMI CLOSEUP Hennessy and Hammond move off, leaving a dismayed Barbara. SEMI LONG SHOT Guy hurrying under the stand toward the waiting cab. MED. SHOT The two men hurrying after him. EXT. CLUB Guy goes to the waiting cab and gets in. The cab moves off. MED. SHOT The two men hurry out of the club and stand helplessly looking after the departing cab. They hurry out of the picture. CLOSE SHOT We see them grab another car. It is a chauffeur-driven limousine. Hammond jumps in front and seats himself beside the driver. Hennessy hops in the back. The car moves off. INT. LIMOUSINE TWO SHOT Hennessy finds himself seated by an old dowager about seventy- five years of age. She looks startled for a moment and almost recoils from him. He shows her his badge. HENNESSY If you'll pardon us, madam, we need your help. We're chasing a man. The old lady's eyes light up. DOWAGER How exciting. (MORE) Converted to PDF by www.screentalk.org 129. DOWAGER (CONT'D) (she leans forward and calls to the chauffeur) Hurry, O'Toole! Hurry! She leans back and maintains her air of excitement as she looks across at Hennessy. CLOSE SHOT INSIDE THE TAXI Guy is busy changing his pants. He glances over his shoulder. INT. CAR The two men looking ahead toward Guy. EXT. METCALF STATION CLOSEUP BRUN0'S FACE - ANGLE SHOOTING UP to get the peering faces behind him. Bruno still frantically trying to reach the lighter. CLOSEUP Under the grating Bruno's fingers slowly closing in on the lighter. They barely manage to grasp it. CLOSEUP BRUN0'S FACE -- triumphant. CLOSEUP Bruno's fist, holding the lighter, comes through the grating. CLOSE SHOT Bruno straightens up. CAMERA BACK as all the onlookers turn their heads in his direction. ONLOOKER You sure must think a lot of that -- Whatever it is. Converted to PDF by www.screentalk.org 130. Bruno doesn't answer. With the lighter in his closed fist, he darts through the crowd, the people looking after him. LONG SHOT The sun is much lower. INT. CLUB CAR Guy is now glancing at his watch. The sun is behind him and very much lower. EXT. AMUSEMENT PARK Bruno is looking at his watch and then across at the sky. LONG SHOT FROM HIS VIEWPOINT The last trace of the setting sun has gone. EXT. METCALF STATION MED. SHOT Guy is stepping off the train. He crosses to a waiting taxi, CAMERA FOLLOWING him. CLOSE SHOT GUY (to the driver) The amusement park, quick. As he gets in the Cab, we go to -- CLOSE SHOT MAN watching Guy get into taxi. As we hear the taxi drive away, the man hurries across to a waiting police car. CLOSE SHOT He puts his head in the side window and tells the two waiting detectives where Guy has gone. MAN Amusement park. Converted to PDF by www.screentalk.org 131. We see one of the detectives take up a microphone as the car drives off. EXT. AMUSEMENT PARK It is now getting dark. MED. SHOT Bruno leaves his spot at the side of the tent and ambles over toward the queue of people waiting for boats. CLOSE SHOT BRUNO joining the queue. He glances ahead of him. MED. SHOT FROM HIS VIEWPOINT We see the light above the pay booth go on, shedding a downward glare. CLOSE SHOT BRUNO pulls his hat a little further over his, eyes. Some new arrivals join the queue behind him. INT. TAXI Guy looking anxiously ahead on his way to the amusement park. AMUSEMENT PARK ENTRANCE We see a police car arrive. One uniformed man and two detectives get out of the car and make their way toward the entrance. One of to detectives stands at the entrance while the other two hurry into the grounds. MED. SHOT Guy's taxi arrives. MED. SHOT Across the street, another police car arrives. Converted to PDF by www.screentalk.org 132. MED. SHOT At Guy is paying his cab fare, he glances around him. MED. SHOT FROM HIS VIEWPOINT He sees one police car. CLOSE SHOT GUY gives a furtive glance around while waiting for his change. MED. SHOT ANOTHER POLICE CAR MED. SHOT Guy cautiously makes his way toward the entrance to the Amusement Park. MED. SHOT Guy passes the waiting detective and looks off. From his viewpoint we see: MED. SHOT THE TWO DETECTIVES who were at the station indicate Guy is the man. MED. SHOT One detective turns away and starts to follow Guy. CLOSE SHOT BRUNO in the queue of people. He is edging slowly along. He is about ten people away from the entrance. He suddenly looks ahead and sees. FROM HIS VIEWPOINT The uniformed man and the detective are talking casually to the boat men in charge of the concession. Converted to PDF by www.screentalk.org 133. DETECTIVE (to boatman) The killer is here tonight. So keep your eyes open and the minute you see him, let us know. CLOSE SHOT The boatman looks at them with an expression of alarm. CLOSE SHOT BRUNO begins to look a little uneasy. We see him begin to mentally deliberate. MED. SHOT Guy, threading his way through the crowds, conscious that he is being followed, but nevertheless, on the lookout for Bruno. CLOSE SHOT BRUNO moving along the line. CAMERA MOVES IN until his head and shoulders fill the screen. He is now coming within range of the flood-lit pay-box. The light seems to creep up across his chest and slowly reveal his face. He lowers his head. MED. SHOT The boatman begins to look along the queue. There is an expression of growing recognition on his face. MED. SHOT Bruno sees this, makes a decision and casually deserts the queue of people. MED. SHOT The boatman hurries across to the uniformed man and begins to talk to him excitedly, looking in Bruno's direction. MED. SHOT GUY Coming along and looking for Bruno. His eyes light up. Converted to PDF by www.screentalk.org 134. SEMI-LONG SHOT FROM HIS VIEWPOINT We see Bruno making his way from the queue of people. CLOSE SHOT GUY calls to Bruno. GUY Hey, Bruno. CLOSE SHOT BRUNO gives a quick glance back, sees Guy then he turns and looks off in another direction. SEMI-LONG SHOT The uniformed man and the boatman approaching him. CLOSE SHOT Bruno hurries on. He stop short as he sees. SEMI-LONG SHOT FROM HIS VIEWPOINT Another uniformed man. MED. SHOT Bruno starts to run. MED. SHOT Guy starts to run after him. MED. SHOT Bruno is seen to jump on a merry-go-round, which is just starting. Its pace is already pretty fast. MED. SHOT Guy runs toward Bruno. Converted to PDF by www.screentalk.org 135. CLOSE SHOT DETECTIVE Haines! Hold it! Hold it! The detective pulls out his gun and starts to run after Guy. SEMI-LONG SHOT Guy jumps on the merry-go-round after Bruno. Its speed is so great that he nearly gets flung off. CLOSE SHOT The detective fires at Guy. CLOSE SHOT The man running the machine in the center of the merry-go- round is suddenly hit in the shoulder. CLOSE SHOT His hand, which is on the starting lever, jerks it down. MED. SHOT The detective, after Guy, jumps on the machine but is flung off on his back. FULL SHOT The merry-go-round has now started to increase the speed. CLOSE SHOT Bruno at the far side is trying to jump off, but it's going too fast. LONG SHOT FROM HIS VIEWPOINT We see the hard ground whizzing past him. Everything seems to be a blur. We get a glimpse of screaming women and the crowds rushing up from the midway. Converted to PDF by www.screentalk.org 136. CLOSE SHOT BRUNO He turns and glances over his shoulder. MED. SHOT FROM HIS VIEWPOINT Guy is threading his way between the rising and falling horses. Guy gets right up close to him. TWO SHOT As Guy comes near to Bruno, the latter turns on him and starts to attack him. BRUNO I want to get off of here! Let me off of here! It makes me dizzy. GUY Stop it, Bruno. Give me my lighter, Bruno! MED.SHOT Against the whirling background of the merry-go-round, Turley and Campbell rush up as the detective struggles to his feet, slightly hurt. The noise from the calliope is very loud. CAMPBELL (to Turley, puzzled; indicating the merry- go-round) Who's the man he's fighting with on there? At this moment the boatman rushes up. BOATMAN (excited) There he is! That's the one! That's the one who killed her! TURLEY Of course he is. We know that. CLOSE SHOT ON MERRY-GO-ROUND Guy and Bruno in a struggle. Guy has to protect himself from a madman whose hands attempt to reach his throat. Converted to PDF by www.screentalk.org 137. They are staggering across between the rising and falling horses. MED. SHOT OUTSIDE MERRY-GO-ROUND A detective turns to the group around hIm. DETECTIVE Get somebody to come and stop that thing! An elderly man in soiled work clothes speaks up. WORKMAN I'll handle it. Immediately the workman heads straight for the merry-go-round and starts to crawl under it on his stomach. DETECTIVE (calls after him) Hey! Be careful! Stop! A second detective speaks to him quizzically. 2ND DETECTIVE Well, do you want to do it yourself? The first detective leans over and looks off toward the workman who is continuing his slithering way under the machine, then straightens up. 1ST DETECTIVE (changing his mind) No. I think he'll make it all right. MED. SHOT GUY AND BRUNO Bruno swings around till his back is to us. He pushes Guy toward the edge, but Guy manages to grab the rein of the nearest horse. The momentum of the machine swings Guy around against the horse, whose big head towers in the f.g. Bruno, on this side of the horse pushes forward and tries to grab the reins from Guy's hand. He tries to slash at Guy's face. The back of Bruno's head is toward us during this. Guy suddenly leans out across the horse and smashes his fist against Bruno's face. Bruno's head goes back until it is in the f.g. in a upside-down position. Converted to PDF by www.screentalk.org 138. MED. SHOT THE CAMERA IS LOW so that we get the effect of Bruno falling into the CAMERA from Guy's blow! MED. SHOT In the f.g. is a young boy of four years. He is excited by the speed of the ride and laughs at the fight with great enjoyment. He sees this by suddenly glancing over his shoulder. In the b.g. Guy and Bruno are continuing their fight. Bruno rises. Guy staggers after him. Bruno again leaps upon Guy. The two men sway toward the CAMERA until Bruno gets alongside the little boy. The boy now shows some anxiety. The three figures now fill the screen with the horses' heads in the f.g. Bruno is forced against the little boy, who now, alarmed, beats Bruno on the cheek with one hand, the other holding onto the brass rail in front. Bruno stops and with a sweep of his arm, knocks the little boy off the horse onto the floor below. The little boy, in falling, grabs the horse's rein or stirrup. CLOSE SHOT Guy breaks away from Bruno and dives around the back of the horse to grab the little boy. CLOSE SHOT As Guy grabs the boy, he staggers forward with him to a small gondola. Bruno leaps onto his back but Guy manages to put the boy in the gondola. CLOSE SHOT UNDERNEATH THE WHIRLING MERRY-GO-ROUND The boat man is making slow progress. FROM HIS VIEW POINT We see his goal. It is the wounded mechanic in the center, who is slightly stirring. All during this the base of the merry-go-round is skimming above the back and head of the boat man. Converted to PDF by www.screentalk.org 139. BACK ON MERRY-GO-ROUND The two men are now in a clinch. Guy tries to fight off the maddened Bruno. They are flung between the horses, bouncing one against the other, almost half way around the merry-go- round. CLOSE SHOT BRUNO AND GUY Again they struggles between two horses. On each side of them are two young screaming girls. The two bounce from one horse to the other. CLOSE SHOT The calliope has little figures and these boat away on their cymbals almost as though they are applauding what's going on. CLOSE SHOT Underneath the merry-go-round, the boat man has made further progress. He is creeping inch by inch. His nose starts to run. He starts to fumble for a dirty piece of handkerchief. He blows his nose and then moves on. CLOSE SHOT Back above the two men swinging past the two girls on their horse and they both crash to the floor underneath another horse, upon which is riding side-saddle, a mother and her three-year-old little girl. CLOSEUP The two big heads of the men, battling. The two men roll underneath the horse's hoofs, which are seen rising and falling. They get right underneath one horse. CLOSEUP Guy has turned over on his back and his eyes look up. Converted to PDF by www.screentalk.org 140. CLOSEUP FROM HIS VIEWPOINT We see the big horse's head above and its hoofs coming down toward the CAMERA and filling the screen. We get a faint impression of the screaming mother hugging her child to her breast, above. BIG CLOSEUP THE HORSE'S HOOFS striking Guy's head. CLOSE SHOT Guy wrenches himself out of this position. He rolls away from the CAMERA right to the edge of the merry-go-round. He manages to grab a rail. MED. SHOT Guy's body is flung out horizontally. We see the crowd behind back-up for fear of being knocked over. The screw of tension increase. Over this comes the sound of an approaching ambulance siren. CLOSE SHOT Bruno edges himself toward Guy. He is hanging on to the reins of a horse. His feet manage to roach Guy's knuckles. CLOSEUP BRUNO'S VICIOUS EXPRESSION CLOSEUP BRUNO'S FEET kicking at Guy's knuckles. CLOSEUP GUY'S AGONIZED EXPRESSION MED. SHOT A flash of the horror-stricken faces of the spectators seen through the whirling machine. Converted to PDF by www.screentalk.org 141. CLOSEUP Machinery and the lever that was pulled on too fast. The Boatman's hand comes up into the picture and pulls the lever over. LONG SHOT The sudden braking causes the whole merry-go-round to topple over with a grinding roar. LONG SHOT FROM HIGH ANGLE The merry-go-round his keeled over. For a moment we don't know who has survived. There is a surge of people milling and shouting. Those who have jumped back out of the way when the merry-go-round toppled, now rush forward again as the cloud of dust settles. From the midway in the background others are running forward. MED. LONG SHOT Distraught parents try to force their way to their children who were on the merry-go-round, but are hold back from the wreckage by police. MED. CLOSE SHOT Guy is somewhat stunned from his fall. He is helped to his feet by some men in the crowd. His knuckles are bleeding. In the background people are rushing about. The crowd is in uproar as women and children are helped from the wreckage. Officials and uniformed policemen pushing back the surge of the crowd. AD LIBS Get back. Get back there. Give us room here. Turley and Campbell rush in to Guy. TURLEY Are you all right, Haines? GUY Yes, I think so. Converted to PDF by www.screentalk.org 142. Guy is surrounded by police and Campbell stands at his elbow. At this moment the boatman runs in. One of the detectives is with him. DETECTIVE Mr. Turley! Mr. Turley! (indicating boatman) He says this isn't the man we want. (with a nod in Guy's direction) It's the other one -- the one he was fighting with. TURLEY (stops to give his full attention to this unexpected bit of information) What do you mean, this isn't the -- (turns to Guy, not quite taking it in) Not Haines? (back to boatman) But you said he was. You pointed him out. BOATMAN No, I didn't, sir. I've never seen this man before in my life. I meant the other one. The detective who was holding Guy instinctively relax his hold on Guy's arm. Turley turns to Guy, puzzled. TURLEY What is this all about, Haines? Did you know he killed your wife? GUY (nods) He has my cigarette lighter and wanted to plant it there on the island to pin the whole thing on me. (urgently) Let me talk to him. Let me show you. Where is he? ANOTHER DETECTIVE Over here. He leads the way. They follow. Converted to PDF by www.screentalk.org 143. MED. CLOSE SHOT as Guy and Turley enter to the spot where Bruno is pinned under the overturned machine. He is caught between two of the horses, the head of one of them across his chest. Bruno's head sags back somewhat, but is resting on pieces of debris. A uniformed policeman looks up from Bruno to Turley: POLICEMAN This one's in a pretty bad way, Mr. Turley. Guy is shocked at the sight of Bruno. GUY (looking down at Bruno) Can't you get that stuff off him? POLICEMAN No, they've done everything they can until the crane comes. Bruno opens his eyes and sees Guy. BRUNO Hello, Guy. Turley has leaned forward to look at the helpless Bruno. BRUNO (weakly nodding at Turley) Who's that? GUY This is Mr. Turley, Chief of Police. BRUNO (with a half smile) So they got you at last, eh, Guy? Guy looks around desperately, frustrated for a moment as Turley eyes him stonily. Then he turns again to Bruno. GUY (rather gently) Can you talk a little? Can you tell the chief you have my lighter? Converted to PDF by www.screentalk.org 144. BRUNO (with a faint, quizzical smile) I haven't got it. It's still on the island where you left it. Guy looks around helplessly to Turley, who looks back at him suspiciously. DETECTIVE (looking down at Bruno) I think he's going. Turley leans over to look. CLOSE SHOT BRUNO'S FIST FROM TURLEY'S VIEWPOINT As Bruno is dying, his closed fist slowly starts to open. DETECTIVE'S VOICE He's finished. Guy's lighter is now revealed in Bruno's open hand. MED. SHOT GROUP Turley takes the lighter from the dead Bruno's hand. Guy is watching him. Turley straightens up and holds the lighter out to him. TURLEY Is this your lighter, Haines? Guy nods without speaking, and with a half look in Bruno's direction. TURLEY Well, you were right. (sticks the lighter into his own pocket) I'd better keep this for the time being. (in a friendly tone) We can clear the whole thing out the morning. How about staying in town over night, Haines? I imagine you have a lot to tell me. Nine o'clock, all right? Converted to PDF by www.screentalk.org 145. GUY (nods) Okay, Mr. Turley. Thanks. Turley turns back to the group around Bruno. Guy looks down for a moment at Bruno, then speaks to the boatman, who is standing nearby. GUY Can you tell me where there's a telephone? BOATMAN (indicating) There's one up near the entrance. (with a look back to the dead Bruno) Who was he, Bud? Guy looks back sympathetically in Bruno's direction, speaks without looking at the boatman. GUY Bruno. Bruno Antony. (reminiscently and a little compassionately, remembering what Bruno had said of himself) A very clever fellow. He moves off through the crowd. DISSOLVE TO: INT. BURTON STUDY NIGHT Anne, Barbara and the Senator are sitting silently in the attitudes of waiting. The telephone rings. Anne is instantly on her feet. Barbara and the Senator watch her anxiously as she goes to answer it. ANNE (into phone) Hello... (impatiently) Yes, operator, yes! (waits a moment, then eagerly:) Guy? (MORE) Converted to PDF by www.screentalk.org 146. ANNE (CONT'D) (a pause, then she closes her eyes with heartfelt relief. Another pause, then:) Yes, darling, yes. Of course I'll be there...Goodbye. She hangs up, turns slowly, to face Barbara and her father. Her expression is one of intense relief. ANNE Guy'll be back tomorrow. (overcome with emotion she has difficulty in speaking) He wants me to take him some things. With a sob, Barbara flings herself into Anne's arms. As she cries, Anne strokes her head comfortingly. Then with a half- choked sobs Anne, too, begins to cry. She speaks through her tears, looking over Barbara's shoulder at her father. ANNE He says he looks silly in his tennis clothes. The Senator eyes them a moment, then speaks a little wryly: SENATOR I presume from all those tears that you have had good news. DISSOLVE TO: INT. PARLOR OF TRAIN NEXT DAY Anne and Guy are sitting quietly together. Opposite them is a man in a clerical collar who is reading a sports magazine. On the cover is a picture of a tennis player in action. The man looks over the top of his magazine at Guy, with recognition. He leans forward. CLERIC I beg your pardon, but aren't you Guy Haines? GUY (uncomfortably) Yes. Converted to PDF by www.screentalk.org 147. Guy and Anne exchange a brief look, rise hurriedly and start to walk away before the conversation can go any farther. The cleric looks after them with a frown and a puzzled shrug of his shoulders, as if to say, "Did I say something wrong?" FADE OUT. THE END | 1 | 5.3% |
| FROZEN Written by Jennifer Lee Final Shooting Draft 9/23/13 OPEN ON: ICE. We're underwater looking up at it. A saw cuts through, heading right for us. EXT. SNOW-CAPPED MOUNTAINS -- DUSK ICE HARVESTERS, dressed in traditional Sami clothing, score a frozen lake. They SING. "The Frozen Heart (Ice Worker's Song)" ICE HARVESTERS BORN OF COLD AND WINTER AIR AND MOUNTAIN RAIN COMBINING, THIS ICY FORCE BOTH FOUL AND FAIR HAS A FROZEN HEART WORTH MINING. The men drag giant ice blocks through channels of water. ICE HARVESTERS (CONT'D) CUT THROUGH THE HEART, COLD AND CLEAR. STRIKE FOR LOVE AND STRIKE FOR FEAR. SEE THE BEAUTY SHARP AND SHEER. SPLIT THE ICE APART! AND BREAK THE FROZEN HEART. Hup! Ho! Watch your step! Let it go! A young Sami boy, KRISTOFF (8), and his reindeer calf, SVEN, share a carrot as they try to keep up with the men. ICE HARVESTERS (CONT'D) Hup! Ho! Watch your step! Let it go! Young Kristoff struggles to get a block of ice out of the water. He fails, ends up soaked. Sven licks his wet cheek. ICE HARVESTERS (CONT'D) BEAUTIFUL! POWERFUL! DANGEROUS! COLD! ICE HAS A MAGIC CAN'T BE CONTROLLED. A sharp ice floe overtakes the workers, threateningly. They fight it back. ICE HARVESTERS (CONT'D) STRONGER THAN ONE, STRONGER THAN TEN STRONGER THAN A HUNDRED MEN! Massive fjord horses drag heavy ice plows. 2 FROZEN - J. Lee ICE HARVESTERS (CONT'D) BORN OF COLD AND WINTER AIR AND MOUNTAIN RAIN COMBINING The sun sets. Lanterns are lit. ICE HARVESTERS (CONT'D) THIS ICY FORCE BOTH FOUL AND FAIR HAS A FROZEN HEART WORTH MINING. CUT THROUGH THE HEART, COLD AND CLEAR. In the dark, Kristoff and Sven finally manage to get a single block of ice out of the water. ICE HARVESTERS (CONT'D) STRIKE FOR LOVE AND STRIKE FOR FEAR. THERE'S BEAUTY AND THERE'S DANGER HERE. SPLIT THE ICE APART! BEWARE THE FROZEN HEART. The workers pile onto the giant horse-drawn ice sled as it pulls away. Left behind, Kristoff and Sven push their ice block onto a dinky little sled then head off. We sweep up from them to the Northern Lights filling the sky...then move across the mountains...beneath the snowline...and descend upon... EXT. THE KINGDOM OF ARENDELLE -- NIGHT A humble castle, built of wood, nestled in a deep fjord. INT. CASTLE, NURSERY -- NIGHT ELSA (8) sleeps in her bed. Her little sister ANNA (5) pops up beside her. YOUNG ANNA Elsa. Psst. Elsa! Psst. Elsa doesn't stir. Anna sits on Elsa and bounces. YOUNG ANNA (CONT'D) Wake up. Wake up. Wake up. YOUNG ELSA (grumbling) Anna, go back to sleep. Anna rolls onto her back and spreads all her weight on Elsa. 3 FROZEN - J. Lee YOUNG ANNA (drama queen-ish) I just can't. The sky's awake, so I'm awake, so we have to play. YOUNG ELSA ...Go play by yourself. Elsa shoves Anna off the bed. Anna lands butt to floor, sighs, defeated. But then she gets an idea. She hops back on the bed and lifts one of Elsa's eyelids. YOUNG ANNA (mischievously) Do you want to build a snowman? Elsa's eyes both pop open. She smiles. INT. CASTLE STAIRCASE -- NIGHT Anna, now wearing snow boots, pulls Elsa by the hand. YOUNG ANNA Come on, come on, come on, come on. Elsa tries to shush her, but Anna's too excited. INT. BALLROOM -- NIGHT The girls sneak into the ballroom. Elsa shuts the door. YOUNG ANNA Do the magic! Do the magic! Elsa laughs and waves her hands together. Snowflakes suddenly burst forth and dance between her palms, forming a snowball. Elsa throws the snowball high into the air. Snow bursts out and flurries around the room. Anna dances about, catching flakes in her palms and mouth. YOUNG ANNA (CONT'D) This is amazing! YOUNG ELSA Watch this! Elsa stomps her little slippered foot and a layer of ice suddenly coats the floor, forming a giant ice rink. Anna slides off, laughing. 4 FROZEN - J. Lee PLAY MONTAGE: -Anna and Elsa roll giant snowballs and build a snowman together. Elsa moves his stick arms around. YOUNG ELSA (CONT'D) (goofy voice) Hi, I'm Olaf and I like warm hugs. Anna jumps up and hugs him. YOUNG ANNA I love you, Olaf. -Anna and Olaf appear to be dancing. REVEAL: Elsa is actually propelling them across the ice floor with her magic. -The girls slide down snowbanks together! -Anna fearlessly jumps off a snow peak into mid air. YOUNG ANNA (CONT'D) Catch me! Elsa makes another peak to catch Anna. YOUNG ELSA Gotcha! Anna keeps jumping. Elsa keeps casting magic. YOUNG ANNA (jumping faster) Again! Again! YOUNG ELSA (struggling to keep up) Slow down! Elsa suddenly slips. Her magic accidentally STRIKES Anna in the head. Anna tumbles down a snowbank and lands, unconscious. YOUNG ELSA (CONT'D) ANNA! Elsa runs to Anna and takes her in her arms. A streak of Anna's hair, where struck, turns white. YOUNG ELSA (CONT'D) MAMA! PAPA! The room around them fills with frightening ice spikes. 5 FROZEN - J. Lee The parents burst through the frozen door. GASP at the sight of the room. KING Elsa, what have you done? This is getting out of hand! QUEEN (seeing Anna) Anna! The King and Queen rush to Anna and take her in their arms. ELSA It was an accident. I'm sorry, Anna. QUEEN (about Anna) She's ice cold. KING ...I know where we have to go. SLAM CUT TO: INT. DARK ROOM -- NIGHT The King sifts through a shelf to find an ancient book inscribed with Old Norse runes. He opens the book, scrambles to a page with an ancient map. EXT. ARENDELLE -- NIGHT Carrying the girls, the King and Queen ride their horses out of the kingdom. Snow streams from Elsa's hands, leaving a trail of ice behind them. EXT. FJORD MOUNTAIN FOREST -- NIGHT A sleepy Kristoff and Sven travel alone through the dark woods. All of a sudden, the King and Queen race by with the girls, leaving the wake of ice. KRISTOFF Ice? SLAM CUT TO: 6 FROZEN - J. Lee EXT. BLACK MOUNTAINS -- NIGHT Kristoff rides Sven as they follow the trail of ice. YOUNG KRISTOFF Faster, Sven! EXT. THE VALLEY OF THE LIVING ROCK -- NIGHT Kristoff hops off Sven at the edge of a deep valley. They hide behind a rock and peek out. Down below, the King holds a frightened Elsa. The Queen holds the still unconscious Anna. KING Please, help. My daughter! Suddenly, a bunch of rocks tumble down the valley toward them. It looks as though they'll be crushed! But, luckily, the rocks stop at their feet. The rocks then unfold, revealing bright faces. YOUNG KRISTOFF Trolls...? The rock in front of Kristoff "wakes up." Meet BULDA. BULDA Shush. I'm trying to listen. She grabs Kristoff and Sven by hand and hoof and hugs them close. Sven licks her face and she eyes them both. BULDA (CONT'D) Cuties. I'm gonna keep you. Back below, the crowd parts for a troll as old as the Earth. They call him GRAND PABBIE. He approaches arthritically, but determined. He nods respectfully to the king. GRAND PABBIE Your Majesty. (referring to Elsa) Born with the powers or cursed? KING Born. And they're getting stronger. Grand Pabbie motions for the Queen to bring Anna to him. She does. He examines her. 7 FROZEN - J. Lee GRAND PABBIE (about Anna) You are lucky it wasn't her heart. The heart is not so easily changed, but the head can be persuaded. KING Do what you must. GRAND PABBIE I recommend we remove all magic, even memories of magic to be safe.... But don't worry, I'll leave the fun. Grand Pabbie pulls out a glowing blue energy from Anna's head. We see her memories floating right above her. Grand Pabbie changes all of her magical memories to ordinary memories -- snowy play indoors with the girls in their nightgowns changes to outdoors on the winter fjords with the girls in winter gear. He puts the ordinary memories back in her head. GRAND PABBIE (CONT'D) She will be okay. YOUNG ELSA But she won't remember I have powers? KING It's for the best. PABBIE Listen to me, Elsa, your power will only grow. As he speaks, he conducts the Northern Lights to show a silhouette of an adult Elsa creating magical snowflakes. PABBIE (CONT'D) There is beauty in your magic.... But also great danger. The snowflakes turn to sharp spikes. PABBIE (O.S.) (CONT'D) You must learn to control it. In the Northern Lights display, the sharp spikes cause human figures to panic and attack Elsa. PABBIE (CONT'D) Fear will be your enemy. 8 FROZEN - J. Lee Elsa gasps and buries her face in the King's chest. The King wraps his arms around Elsa, protectively. KING No. We'll protect her. She can learn to control it. I'm sure. Over the King's words we... DISSOLVE TO: -The Arendelle castle gates shutting. KING (O.S.) (CONT'D) Until then, we'll lock the gates. We'll reduce the staff. We will limit her contact with people and keep her powers hidden from everyone... including Anna. -The castle shutters close. -Anna sits on her bed as Elsa's furniture disappears. -Anna rushes to the hall to see Elsa shut the door to her new room. Anna watches, confused and sad. DISSOLVE TO: INT. CASTLE WINDOW -- DAY We look out on a gentle snowfall. Little Anna skips up to the window. She lights up at the sight of the snow and rushes down the hall. INT. HALLWAY, ELSA'S DOOR -- DAY Anna knocks on Elsa's door and SINGS. "Do You Want to Build a Snowman?" YOUNG ANNA DO YOU WANT TO BUILD A SNOWMAN? COME ON LET'S GO AND PLAY. Anna peeks under the door. YOUNG ANNA (CONT'D) I NEVER SEE YOU ANYMORE. COME OUT THE DOOR. IT'S LIKE YOU'VE GONE AWAY. 9 FROZEN - J. Lee -INT. ANNA'S ROOM -- Anna plays with two dolls, gives up, sad. YOUNG ANNA (CONT'D) WE USED TO BE BEST BUDDIES AND NOW WE'RE NOT. I WISH YOU WOULD TELL ME WHY. -ELSA'S DOOR. Anna peeks through the key hole. YOUNG ANNA (CONT'D) DO YOU WANT TO BUILD A SNOWMAN? -Anna calls through the keyhole. YOUNG ANNA (CONT'D) IT DOESN'T HAVE TO BE A SNOWMAN. YOUNG ELSA (O.S.) Go away, Anna. YOUNG ANNA (hearbroken) ...OKAY BYE. -BEHIND THE DOOR -- DAY. Elsa sits at the window looking out, longingly. Suddenly, her icy hands freeze the windowsill. -LATER. The King slips leather gloves onto Elsa's hands. KING The gloves will help. He pats her gloved hand. KING (CONT'D) See? You're good.... (starting their mantra) Conceal it. YOUNG ELSA Don't feel it. YOUNG ELSA & KING Don't let it show. -INT. HALLWAY, ELSA'S DOOR -- DAY. Anna, now 9, knocks on Elsa's door. ANNA (9) DO YOU WANT TO BUILD A SNOWMAN? -INT. HALLWAY -- DAY. Alone, Anna rides a bicycle built for two in the hall by standing on the back seat. 10 FROZEN - J. Lee ANNA (9) (CONT'D) OR RIDE OUR BIKE AROUND THE HALL? I THINK SOME COMPANY IS OVERDUE... -INT. PORTRAIT ROOM -- DAY. Anna runs around the portrait room, gaining momentum to flip over the arm of the couch. ANNA (9) (CONT'D) I'VE STARTED TALKING TO THE PICTURES ON THE WALLS. Anna lands PLOP on the cushions, then looks up at the painting above her of the courageous Joan of Arc. ANNA (9) (CONT'D) Hang in there, Joan. -INT. EMPTY LIBRARY -- DAY. Looks like no one's around. ANNA (9) (CONT'D) IT GETS A LITTLE LONELY ALL THESE EMPTY ROOMS. But then we find Anna, laying at the base of the grandfather clock, playing with her braids, bored out of her mind. ANNA (9) (CONT'D) JUST WATCHING THE HOURS TICK BY. Anna's eyes follow the grandfather clock's pendulum. ANNA (9) (CONT'D) TICK TOCK. TICK TOCK. TICK TOCK. -INT. ELSA'S ROOM -- NIGHT. Elsa (now 12) paces as she panics. The entire wall is frozen behind her. ELSA (12) I'm scared. It's getting stronger. KING Getting upset only makes it worse. The King goes to hug her. ELSA (12) No. Don't touch me. I don't want to hurt you. He and the Queen look at each other with alarmed sadness. -INT. LIBRARY -- DAY. Anna, now a teenager, slides past Elsa's room without stopping. 11 FROZEN - J. Lee -INT. KING AND QUEEN'S QUARTERS -- DAY. Anna runs into the room and throws herself into her parents' arms. TEEN ANNA See you in two weeks. -INT. ELSA'S ROOM -- DAY. Elsa curtsies in front of her parents, formally, not touching them. TEEN ELSA Do you have to go? KING You'll be fine, Elsa. -EXT. DOCKS -- DAY. The King and Queen leave on a ship. -EXT. ROUGH SEAS -- NIGHT. Lightning flashes. The sea rages in a storm. The King and Queen's ship is lost in the waves. -INT. CASTLE -- DAY. A portrait of the King and Queen is covered in mourning cloth. -EXT. CEMETERY -- DAY. Anna looks small, standing before her people, beside burial stones. -INT. HALLWAY, ELSA'S DOOR. Anna, still in her mourning clothes, approaches and knocks. ANNA (singing) Elsa? PLEASE I KNOW YOU'RE IN THERE PEOPLE ARE ASKING WHERE YOU'VE BEEN THEY SAY HAVE COURAGE AND I'M TRYING TO I'M RIGHT OUT HERE FOR YOU. PLEASE LET ME IN. Anna slides down the door and sits with her head against it. ANNA (CONT'D) WE ONLY HAVE EACH OTHER. IT'S JUST YOU AND ME. WHAT ARE WE GONNA DO? (weak, internal) DO YOU WANT TO BUILD A SNOWMAN? We move through the door... -INT. ELSA'S ROOM -- DAY. Elsa is sitting in the exact same pose as Anna. Her bedroom is frozen with ice. Snowflakes hang in the air, suspended by grief. FADE OUT. 12 FROZEN - J. Lee EXT. THE KINGDOM OF ARENDELLE -- MORNING A new dawn rises over the fjords. Ships pull up to the docks. Guests pile out. DOCK MASTER Welcome to Arendelle! A BOY tries to get away as his MOTHER tries to stuff him in his bunad jacket. BOY Why do I have to wear this? MOTHER Because the Queen has come of age. It's Coronation Day! BOY That's not my fault. They pass the May Pole being raised and a Sami ice harvester chatting with his reindeer. We recognize them as Kristoff and Sven, all grown up. Sven hops around excitedly like a dog and nuzzles Kristoff's chest. KRISTOFF What do you want, Sven? Kristoff leans in and speaks for Sven, as if he can. KRISTOFF (AS SVEN) (CONT'D) Give me a snack. KRISTOFF (CONT'D) What's the magic word? KRISTOFF (AS SVEN) (CONT'D) Please! Kristoff pulls a carrot out of his shirt pocket and hands it to Sven. Sven tries to bite the whole thing. KRISTOFF (CONT'D) Hey, hey, hey! Share! Sven takes a smaller bite. Kristoff then has a bite himself, not seeming to care that it's covered in reindeer slobber. We move on to PERSI and AGGIE, a super-excited couple who rush towards the castle. 13 FROZEN - J. Lee PERSI I can't believe they're finally opening up the gates! AGGIE And for a whole day! Faster, Persi! They pass a tiny but menacing DUKE, who wears taps on his shoes to "enhance" his presence. Two THUG guards follow close behind him. DUKE Ah, Arendelle, our most mysterious trade partner. Open those gates so I may unlock your secrets and exploit your riches. (catching himself) ...Did I just say that out loud? We leave him and head down the bridge towards the castle gates, passing an Irishman and a Spanish Dignitary. IRISHMAN Oh, me sore eyes can't wait to see the Queen and the Princess. I bet they're absolutely lovely. SPANISH DIGNITARY I bet they are beautiful. We move past them, to a particular castle window. CUT TO: INT. CASTLE, ANNA'S BEDROOM -- DAY Anna, 18, snores. Drools. KNOCK. KNOCK. KAI (O.S.) Princess Anna...? Anna sits up. She's got major bedhead. She coughs. Snorts. Pulls a hair from her mouth. ANNA ...Huh? Yeah? KAI (O.S.) Sorry to wake you, ma'am but-- ANNA No, you didn't. I've been up for hours. 14 FROZEN - J. Lee She falls back asleep while sitting. She snores. Her head drops, startling her awake. ANNA (CONT'D) Who is it? KAI (O.S.) It's still me, ma'am. Time to get ready. ANNA Ready for what? KAI (O.S.) Your sister's coronation, ma'am. ANNA My sister's cor-neration... One eye opens enough to catch sight of her coronation dress. She bolts, wide awake in excitement. ANNA (CONT'D) Coronation Day! Ha ha! SLAM CUT TO: EXT. CASTLE HALL -- DAY Anna bursts out of her room, wearing her coronation dress. She finishes pinning ribbons in her hair. Seeing the hustle and bustle of preparations, she can't help but SING. "For the First Time in Forever" ANNA THE WINDOW IS OPEN! SO'S THAT DOOR! I DIDN'T KNOW THEY DID THAT ANYMORE. WHO KNEW WE OWNED 8000 SALAD PLATES...? -Anna slides along the floor of the ballroom in her socks. ANNA (CONT'D) FOR YEARS I HAVE ROAMED THESE EMPTY HALLS WHY HAVE A BALLROOM WITH NO BALLS? FINALLY, THEY'RE OPENING UP THE GATES! -She shakes hands with a suit of armor. Breaks it. Hides the evidence. 15 FROZEN - J. Lee ANNA (CONT'D) THERE'LL BE REAL, ACTUAL PEOPLE - IT'LL BE TOTALLY STRANGE. BUT WOW AM I SO READY FOR THIS CHANGE! -Anna comes to a window and jumps out onto a window washer's pulley. She raises herself up to see the ships arriving. ANNA (CONT'D) FOR THE FIRST TIME IN FOREVER, THERE'LL BE MUSIC, THERE'LL BE LIGHT. FOR THE FIRST TIME IN FOREVER, I'LL BE DANCING THROUGH THE NIGHT. -Anna walks through the garden and follows a family of geese. ANNA (CONT'D) DON'T KNOW IF I'M ELATED OR GASSY, BUT I'M SOMEWHERE IN THAT ZONE 'CAUSE FOR THE FIRST TIME IN FOREVER, I WON'T BE ALONE. (speaking) I can't wait to meet everyone.... (GASP) What if I meet THE ONE? -Anna twists herself in a velvet drape like it's a gown. She acts like she looks gorgeous, but she looks ridiculous. ANNA (CONT'D) TONIGHT, IMAGINE ME GOWN AND ALL- FETCHINGLY DRAPED AGAINST THE WALL. THE PICTURE OF SOPHISTICATED GRACE. -She notices the bust of a man across the room. ANNA (CONT'D) (google-eyed) I SUDDENLY SEE HIM STANDING THERE, A BEAUTIFUL STRANGER TALL AND FAIR. (mouth full of chocolate) I WANNA STUFF SOME CHOCOLATE IN MY FACE! -She grabs the bust of the man and swings it around. ANNA (CONT'D) BUT THEN WE LAUGH AND TALK ALL EVENING, WHICH IS TOTALLY BIZARRE. NOTHING LIKE THE LIFE I'VE LED SO FAR. The bust goes flying and lands on the top of the cake. -Anna bursts into the portrait room, bounces on the furniture, and interacts with the paintings. 16 FROZEN - J. Lee ANNA (CONT'D) FOR THE FIRST TIME IN FOREVER, THERE'LL BE MAGIC, THERE'LL BE FUN. FOR THE FIRST TIME IN FOREVER, I COULD BE NOTICED BY SOMEONE. AND I KNOW IT IS TOTALLY CRAZY TO DREAM I'D FIND ROMANCE. BUT FOR THE FIRST TIME IN FOREVER, AT LEAST I'VE GOT A CHANCE! -INT. LIBRARY. ELSA, now a very poised 21, watches out the window as the coronation guests arrive. ELSA DON'T LET THEM IN. DON'T LET THEM SEE. BE THE GOOD GIRL YOU ALWAYS HAVE TO BE. Elsa moves to a painting of her father's coronation. She takes off her gloves and mimics the painting by holding a candlestick and ornament in place of an orb and scepter. ELSA (CONT'D) CONCEAL. DON'T FEEL. PUT ON A SHOW. MAKE ONE WRONG MOVE AND EVERYONE WILL KNOW. The candlestick and ornament ice over. Elsa gasps, slams them back down onto the table. She tries to reassure herself. ELSA (CONT'D) BUT IT'S ONLY FOR TODAY. We cut between Anna's excitement and Elsa's nerves. ANNA IT'S ONLY FOR TODAY! ELSA IT'S AGONY TO WAIT. ANNA IT'S AGONY TO WAIT!!! ELSA TELL THE GUARDS TO OPEN UP THE GATE. ANNA THE GATE!!! -Finally, the gates are open! Anna moves through the crowd, admiring the people around her. 17 FROZEN - J. Lee ANNA (CONT'D) ELSA FOR THE FIRST TIME IN DON'T LET THEM IN FOREVER. DON'T LET THEM SEE ANNA ELSA I'M GETTING WHAT I'M DREAMING BE THE GOOD GIRL OF YOU ALWAYS HAVE TO BE ANNA ELSA A CHANCE TO LEAVE MY SISTER'S CONCEAL. WORLD CONCEAL. DON'T FEEL. A CHANCE TO FIND TRUE LOVE DON'T LET THEM KNOW. -Anna hurries over the bridge and into the village square. ANNA (CONT'D) I KNOW IT ALL ENDS TOMORROW, SO IT HAS TO BE TODAY!! `CAUSE FOR THE FIRST TIME IN FOREVER. . . FOR THE FIRST TIME IN FOREVER! NOTHING'S IN MY WAY!!! -Anna SLAMS right into the breast of a HORSE! She falls back and lands in a small wooden boat. It tips off of the dock. She's heading overboard. But just then, the horse slams his hoof into the boat and steadies it. ANNA (CONT'D) (frustrated) Hey! HANS I'm so sorry. Are you hurt? The rider, HANS, sure is handsome and regal. ANNA (gentler) Hey. I-ya, no. No. I'm okay. HANS Are you sure? ANNA Yeah, I just wasn't looking where I was going. But I'm okay. He hops down from his horse and steps into the boat. ANNA (CONT'D) I'm great, actually. 18 FROZEN - J. Lee HANS Oh, thank goodness. He offers her a hand and their eyes meet. Chemistry. He helps her to her feet. HANS (CONT'D) (bowing) Prince Hans of the Southern Isles. ANNA (curtseying) Princess Anna of Arendelle. HANS Princess...? My Lady. He drops to his knees, head bowed. The horse bows too, curling his hoof up and out of the boat. The boat tips. Hans tumbles on top of Anna. Awkward. ANNA Hi...again. The horse slams his foot back into the boat to stabilize it. Anna and Hans tumble the other way. Anna lands on top of him. HANS Oh boy. ANNA Ha. This is awkward. Not you're awkward, but just because we're-- I'm awkward. You're gorgeous. (did she just say that?) Wait, what? Hans quickly gets to his feet and helps Anna up again. HANS I'd like to formally apologize for hitting the Princess of Arendelle with my horse...and for every moment after. ANNA No. No-no. It's fine. I'm not THAT Princess. I mean, if you'd hit my sister Elsa, that would be-- yeash! `Cuz, you know... (patting the horse) Hello. (MORE) 19 FROZEN - J. Lee ANNA (CONT'D) (to Hans) But, lucky you, it's-it's just me. HANS Just you? Hans smiles, amused. She smiles back. The bells RING. She doesn't notice at first; she's too busy drinking in Hans's handsomeness. ANNA ...The bells. The coronation. I-I-I better go. I have to...I better go. She hurries off, stops, turns back. Gives Hans a little wave. ANNA (CONT'D) Bye! As she rushes off again, Hans waves back. The horse waves too, once again taking his hoof out of the boat. HANS Oh no. The boat falls, with Hans in it. SPLASH! It lands upside down in the water. Hans raises it up off of him, gasping for air. CUT TO: INT. CHURCH CHAPEL -- DAY Elsa stands at the alter. Anna stands off to one side. She peeks out to the audience. Hans waves at her from the pews. He's changed his clothes. The crown is placed on Elsa's head. The scepter and orb are presented to Elsa on a pillow. She slowly reaches for them. BISHOP (a whisper) Your Majesty, the gloves. Elsa hesitates. She breathes nervously, removes her gloves, places them on the pillow. Her hands shake. She takes the orb and scepter, then turns to the people. BISHOP (CONT'D) (formal, in Old Norse) Sehm hon HELL-drr IN-um HELL-gum AYG-num ok krund ee THES-um HELL- gah STAHTH, ehk teh frahm FUR-ear U- thear... 20 FROZEN - J. Lee The scepter and orb start to freeze over. BISHOP (CONT'D) ...Queen Elsa of Arendelle. CROWD Queen Elsa of Arendelle. Just in time. Elsa manages to set the orb and scepter back down on the pillow before anyone notices the ice. She picks up her gloves and slips them on. She made it. CUT TO: INT. GREAT HALL -- NIGHT Springy music fills the Great Hall. Guests dance. Eat. Laugh. TRUMPETS SOUND. KAI (announcing) Queen Elsa of Arendelle. Elsa enters, poised and looking surprisingly content. She stands under a formal awning. KAI (CONT'D) Princess Anna of Arendelle! Anna runs into the room, waves awkwardly. Kai ushers her over to stand right next to Elsa. ANNA Here? Are you sure? She and Elsa sneak awkward peeks at each other. ELSA ...Hi. ANNA Hi me...? Oh. Um. Hi. ELSA ...You look beautiful. ANNA Thank you. You look beautifuller. I mean, not fuller. You don't look fuller, but more beautiful. 21 FROZEN - J. Lee ELSA Thank you. They look out at the celebration. ELSA (CONT'D) So, this is what a party looks like? ANNA It's warmer than I thought. ELSA And what is that amazing smell? They both close their eyes and inhale. ANNA AND ELSA (TOGETHER) ...Chocolate. Their eyes pop open. They laugh. Elsa looks back out at the party. Anna looks at Elsa. She wants to say so much, but she can't think of where to start. Just as she finds her way, Kai interrupts. KAI Your Majesty. The Duke of Weaseltown. DUKE Weselton. The Duke of Weselton. (to Elsa) Your Majesty, as your closest partner in trade, it seems only fitting that I offer you your first dance as queen. The Duke does a funny flitter of his feet, a hitch-kick, and a deep bow. DUKE (CONT'D) (whispers to himself) One, two, three. Jump. As he holds out his hand, head down, his toupee dips forward. Anna giggles. Elsa looks at Anna, stifles a giggle herself. ELSA (to the Duke) Thank you...only I don't dance. 22 FROZEN - J. Lee DUKE (offended) Oh...? ELSA But my sister does. ANNA What? DUKE Lucky you.... ANNA Oh, I don't think-- The Duke grabs Anna's arm and yanks her away before she can protest. DUKE If you swoon, let me know, I'll catch you. Anna looks back at Elsa, desperately. ELSA Sorry. OUT ON THE DANCE FLOOR: The Duke showboats, but he's just awful. Anna tries to make the best of it. DUKE Like an agile peacock... CLUCK- CLUGGLE-CLUCK! He lands on her feet. ANNA Ow. Ow. DUKE Speaking of, so great to have the gates open. Why did they shut them in the first place? Do you know the reason? Hmm? He gets in her face, suspicious. ANNA ...No. 23 FROZEN - J. Lee DUKE Oh, all right. Hang on. They don't call me the little dipper for nothing. He dips Anna back. Elsa peeks through the crowd, can barely hold in her laughter. Anna shoots Elsa funny, help-me looks. DUKE (CONT'D) (groove fully on) Like a chicken...with the face of a monkey...I fly. JUMP CUT TO: MOMENTS LATER... Anna limps back to Elsa. DUKE (O.S.) Let me know when you're ready for another round, M'Lady. ELSA Well, he was sprightly. ANNA (rubbing her sore feet) Especially for a man in heels. ELSA Are you okay? ANNA (loving Elsa's attention) I've never been better. This is so nice. I wish it could be like this all the time. ELSA (sincere) Me too.... But then Elsa catches herself. She stiffens up, looks away. ELSA (CONT'D) But it can't. ANNA Why not? If-- ELSA It just can't. 24 FROZEN - J. Lee Anna's smile drops. She tries not to get emotional. ANNA Excuse me for a minute. She walks away. Elsa watches her go, saddened. Moving through the crowd, Anna gets bumped by a bowing man's butt. She falls. Just before she hits the floor, Hans catches her. He smiles perfectly. HANS Glad I caught you. ANNA Hans. He smoothly sets his drink down on a passing tray. He lifts her up and leads her in a romantic dance. DISSOLVE TO: LATER: Anna and Hans drink and chat. ANNA (CONT'D) I often had the whole parlor to myself to slide... Oops. Sorry. She hits him in the face by mistake with her hand. He laughs. DISSOLVE TO: -THE CASTLE DOORS: Anna and Hans stroll out of the castle. ANNA (CONT'D) ...Your physique helps I'm sure. DISSOLVE TO: -THE ROSE GARDEN... Hans notices her white streak. HANS (about her white streak) What's this? ANNA I was born with it, although I dreamt I was kissed by a troll. HANS I like it. DISSOLVE TO: 25 FROZEN - J. Lee EXT. BALCONY -- NIGHT Anna teaches Hans how to eat krumkake. ANNA Yeah, the whole thing! You got it. They laugh as the krumkake crumbles in his face. ANNA(CONT'D) Okay wait, wait. So you have how many brothers? HANS Twelve older brothers. Three of them pretended I was invisible... literally...for two years. ANNA That's horrible. HANS It's what brothers do. ANNA ...And sisters. Elsa and I were really close when we were little. But then, one day she just shut me out, and I never knew why. He takes her hand. Leans in close. HANS I would never shut you out. ANNA Okay, can I just say something crazy? HANS I love crazy. "Love is an Open Door" ANNA (singing) ALL MY LIFE HAS BEEN A SERIES OF DOORS IN MY FACE. AND THEN SUDDENLY I BUMP INTO YOU. HANS I was thinking the same thing, because like. . . (MORE) 26 FROZEN - J. Lee HANS (CONT'D) I'VE BEEN SEARCHING MY WHOLE LIFE TO FIND MY OWN PLACE. AND MAYBE IT'S THE PARTY TALKING, OR THE CHOCOLATE FONDUE. ANNA BUT WITH YOU- HANS BUT WITH YOU, I FOUND MY PLACE. ANNA I SEE YOUR FACE. BOTH AND IT'S NOTHING LIKE I'VE EVER KNOWN BEFORE. They jump to the neighboring balcony and enter a door. They come out on top of one of the castle's towers. BOTH (CONT'D) LOVE IS AN OPEN DOOR! LOVE IS AN OPEN DOOR! Cut to them sliding across an empty hallway in their socks. BOTH (CONT'D) LOVE IS AN OPEN DOOR ANNA WITH YOU! HANS WITH YOU! ANNA WITH YOU! HANS WITH YOU! BOTH LOVE IS AN OPEN DOOR. They hop up on the castle roof and watch a shooting star. HANS I MEAN IT'S CRAZY. ANNA What? 27 FROZEN - J. Lee HANS WE FINISH EACH OTHER'S- ANNA SANDWICHES! HANS That's what I was gonna say! They slide down the back of the roof out of sight. We next find them strutting on a bridge ledge. ANNA I'VE NEVER MET SOMEONE- BOTH WHO THINKS SO MUCH LIKE ME. BOTH (SPOKEN) (CONT'D) Jinx.. . .jinx again. Are they doing the robot? No. They're imitating the mechanical figures on the clock tower. BOTH (CONT'D) OUR MENTAL SYNCHRONIZATION CAN HAVE BUT ONE EXPLANATION, HANS YOU- ANNA AND I- HANS WERE- ANNA JUST- BOTH MEANT TO BE. Anna and Hans dance on top of the lighthouse and cast dancing shadows across the sails of ships in the docks. ANNA SAY GOODBYE- HANS SAY GOODBYE- 28 FROZEN - J. Lee BOTH TO THE PAIN OF THE PAST. BOTH (CONT'D) WE DON'T HAVE TO FEEL IT ANYMORE! LOVE IS AN OPEN- They play hide and seek amongst the stable doors. BOTH (CONT'D) DOOR! LOVE IS AN OPEN DOOR! They climb to the waterfall looking out over the kingdom. Anna raises up her hands to frame the moon. Hans puts his hands on top of hers. Together their hands form a heart. BOTH (CONT'D) LIFE CAN BE SO MUCH MORE- ANNA WITH YOU! HANS WITH YOU! ANNA WITH YOU! HANS WITH YOU! BOTH LOVE IS AN OPEN HANS DOOR. ANNA DOOR. HANS Can I say something crazy...? Will you marry me? ANNA Can I just say something even crazier? Yes. CUT TO: 29 FROZEN - J. Lee INT. BALL -- NIGHT Anna pushes through the crowd towards Elsa, Hans in tow. ANNA Oops! Pardon. Sorry. Can we just get around you there? Thank you. Oh, there she is. Elsa! Elsa turns to Anna. Anna curtseys awkwardly. ANNA (CONT'D) I mean...Queen.... Me again. Um. May I present Prince Hans of the Southern Isles. HANS (bowing) Your Majesty. Elsa gives a polite but reserved curtsey. ANNA We would like-- HANS --your blessing-- ANNA --of-- ANNA/HANS --our marriage! ELSA Marriage...? ANNA Yes! ELSA I'm sorry, I'm confused. ANNA Well, we haven't worked out all the details ourselves. We'll need a few days to plan the ceremony. Of course we'll have soup, roast, and ice cream and then-- Wait. Would we live here? ELSA Here? 30 FROZEN - J. Lee HANS Absolutely! ELSA Anna-- ANNA Oh, we can invite all twelve of your brothers to stay with us-- ELSA What? No, no, no, no, no. ANNA Of course we have the room. I don't know. Some of them must-- ELSA Wait. Slow down. No one's brothers are staying here. No one is getting married. ANNA Wait, what? ELSA May I talk to you, please. Alone. Anna sees Hans's worried face. Hooks arms with him. ANNA No. Whatever you have to say, you- you can say to both of us. ELSA Fine. You can't marry a man you just met. ANNA You can if it's true love. ELSA Anna, what do you know about true love? ANNA More than you. All you know is how to shut people out. ELSA You asked for my blessing, but my answer is no. Now, excuse me. 31 FROZEN - J. Lee HANS Your Majesty, if I may ease your-- ELSA (flustered) No, you may not. And I-I think you should go. Elsa walks away. As she passes the Royal Handler-- ELSA (CONT'D) The party is over. Close the gates. ANNA What? Elsa, no. No, wait! Anna grabs Elsa's hand. She pulls off Elsa's glove. Elsa gasps, spins around and reaches for the glove in panic. ELSA Give me my glove! Anna holds the glove away from Elsa. ANNA (desperate) Elsa, please. Please. I can't live like this anymore. Elsa fights tears. ELSA (weak) ...Then leave. Elsa sees Anna's hurt face. It's too much. She can't hold it in. She turns and rushes away. ANNA (heartbroken) ...What did I ever do to you?! The party goes silent as everyone watches the sisters. ELSA Enough, Anna. ANNA No. Why? Why do you shut me out?! Why do you shut the world out?! What are you so afraid of?! ELSA I said, enough! 32 FROZEN - J. Lee Ice shoots from Elsa's hand, spikes across the floor! Guests cry out in shock, back away. DUKE (ducking behind his men) ...Sorcery. I knew there was something dubious going on here. ANNA Elsa...? Elsa rushes out of the room. CUT TO: EXT. COURTYARD -- NIGHT Elsa bursts out of the castle door. The CITIZENS CHEER! CROWD There she is. Your Majesty! Long live the Queen! Queen Elsa.... Come drink with us. Elsa ducks through the crowd, holding her bare hand. BOWING TOWNSMAN Queen Elsa. TOWNSWOMAN WITH BABY Your Majesty? Are you all right? Elsa backs away from the baby. She knocks into the fountain, grabs its edge. The waters freeze at her touch. GASPS of shock and fear sweep over the crowd. The Duke and thugs come out the door. DUKE There she is! Stop her! ELSA (to the Duke) Please, just stay away from me. Stay away! Magic accidentally shoots from her hand and turns the staircase into ice. The thugs and the Duke fall. DUKE Monster.... Monster! 33 FROZEN - J. Lee The crowd panics. A snowstorm begins. Elsa flees. Anna runs out of the palace doors, carrying the glove. ANNA Elsa! Hans follows closely behind her. GATES TO THE KINGDOM: Elsa runs out of the gates and down to the water's edge. The shoreline freezes under her feet. Anna calls to her from the gates. ANNA (CONT'D) Elsa! Wait, please! Elsa glances back at Anna, but turns away. She tentatively steps out onto the fjord. It freezes instantly. She breaks into a run, as the water freezes over with each step. ANNA (CONT'D) Elsa, stop! Anna rushes out onto the fjord ice, slips, falls. HANS Anna! Hans rushes to Anna's side. Elsa reaches the far shore. She doesn't look back. She just scrambles into the mountains. ANNA No. HANS (shocked) Look.... The fjord. The ice spreads out until the entire fjord is frozen, locking the ships in place. INT. CASTLE COURTYARD -- NIGHT Snow falls. Hans and Anna move through the panicking crowd. CROWD WALLAH Snow? It's...snow...in July. 34 FROZEN - J. Lee HANS ...Are you all right? ANNA (in shock) No. HANS Did you know? ANNA No. Nearby, the Duke flutters about in fright. DUKE Look! It's snowing! It's snowing! The Queen has cursed this land! She must be stopped! (to his thugs) You have to go after her. Anna rushes up to the Duke. ANNA Wait, no! The Duke hides behind his thugs and points out at Anna. DUKE You! Is there sorcery in you, too? Are you a monster, too? ANNA No. No. I'm completely ordinary. HANS That's right she is... (realizing how that sounds) ...in the best way. ANNA ...And my sister's not a monster. DUKE She nearly killed me. HANS You slipped on ice. DUKE Her ice! 35 FROZEN - J. Lee ANNA It was an accident. She was scared. She didn't mean it. She didn't mean any of this.... Tonight was my fault. I pushed her. So I'm the one that needs to go after her. DUKE Yes. Fine. Do. HANS What? ANNA (to the Royal Handler) Bring me my horse, please. HANS Anna, no. It's too dangerous. ANNA Elsa's not dangerous. I'll bring her back, and I'll make this right. The Royal Handler brings Anna her horse and a cloak. HANS I'm coming with you. ANNA No, I need you here to take care of Arendelle. He sees the desperation in her eyes. HANS ...On my honor. She throws on the cloak and hops right onto the horse, coronation dress and all. ANNA (to the crowd) I leave Prince Hans in charge! HANS (before letting her go) Are you sure you can trust her? I don't want you getting hurt. ANNA She's my sister; she would never hurt me. 36 FROZEN - J. Lee She snaps the reins and rides out. Hans watches after her. The snow picks up and overtakes our view. We push through a blizzard...lose our way...then finds ourselves... EXT. HIGH UP IN THE MOUNTAINS -- NIGHT Well above the snow-line, a small figure climbs the highest peak. It's Elsa. Finally, she stops, looks around. Catches her breath and sings... "Let It Go" ELSA THE SNOW GLOWS WHITE ON THE MOUNTAIN TONIGHT, NOT A FOOTPRINT TO BE SEEN. A KINGDOM OF ISOLATION AND IT LOOKS LIKE I'M THE QUEEN. THE WIND IS HOWLING LIKE THIS SWIRLING STORM INSIDE. COULDN'T KEEP IT IN, HEAVEN KNOWS I TRIED. . . DON'T LET THEM IN, DON'T LET THEM SEE, BE THE GOOD GIRL YOU ALWAYS HAVE TO BE. CONCEAL, DON'T FEEL, DON'T LET THEM KNOW. WELL, NOW THEY KNOW. Elsa takes off her glove and throws it into the air. ELSA (CONT'D) LET IT GO. LET IT GO. CAN'T HOLD IT BACK ANYMORE. Elsa creates a snowman, just like the one she made with Anna when they were children. ELSA (CONT'D) LET IT GO. LET IT GO. TURN AWAY AND SLAM THE DOOR. I DON'T CARE WHAT THEY'RE GOING TO SAY. LET THE STORM RAGE ON. THE COLD NEVER BOTHERED ME ANYWAY. Elsa lets her cape fly back into the wind. 37 FROZEN - J. Lee ELSA (CONT'D) IT'S FUNNY HOW SOME DISTANCE MAKES EVERYTHING SEEM SMALL. AND THE FEARS THAT ONCE CONTROLLED ME CAN'T GET TO ME AT ALL. IT'S TIME TO SEE WHAT I CAN DO, TO TEST THE LIMITS AND BREAK THROUGH. NO RIGHT, NO WRONG, NO RULES FOR ME...I'M FREE! Elsa creates ice steps and climbs them. ELSA (CONT'D) LET IT GO! LET IT GO! I AM ONE WITH THE WIND AND SKY. LET IT GO! LET IT GO! YOU'LL NEVER SEE ME CRY. HERE I STAND AND HERE I'LL STAY. Elsa slams her foot down and forms a giant snowflake. ELSA (CONT'D) LET THE STORM RAGE ON.... In a flurry of creative release, she raises the snowflake on ice beams, builds walls, archways, a glistening chandelier, and an intricate ceiling that leaves the sky visible. ELSA (CONT'D) MY POWER FLURRIES THROUGH THE AIR INTO THE GROUND. MY SOUL IS SPIRALING IN FROZEN FRACTALS ALL AROUND. AND ONE THOUGHT CRYSTALLIZES LIKE AN ICY BLAST- Standing firmly in her mighty ice palace, Elsa removes her crown and throws it. ELSA (CONT'D) I'M NEVER GOING BACK, (back to resolve) THE PAST IS IN THE PAST! She takes down her hair and creates a new dress made of ice. ELSA (CONT'D) LET IT GO! LET IT GO! AND I'LL RISE LIKE THE BREAK OF DAWN. LET IT GO! LET IT GO! The sun rises. Elsa struts onto out onto a balcony and into the light. She's free. 38 FROZEN - J. Lee ELSA (CONT'D) THAT PERFECT GIRL IS GONE. HERE I STAND IN THE LIGHT OF DAY. LET THE STORM RAGE ON!! THE COLD NEVER BOTHERED ME ANYWAY. She turns and slams her ice palace door on us. CUT TO: EXT. THE FJORD FOREST -- DAY Anna rides her horse through two feet of snow. She shivers. ANNA (shivering) Elsa! Elsa! It's me, Anna...your sister who didn't mean to make you freeze the summer. I'm sorry. It's all my f-f-f-f-f-f-fault. DISSOLVE TO: LATER: Anna and the horse struggle through a wooded area. ANNA (CONT'D) (hearing a wolf howl) Of course, none of this would have happened if she'd just told me her secret...ha...she's a stinker. A branch of a nearby tree snaps and startles the horse. Anna goes flying off, lands face down in the snow. She sits up. Spits out snow. Sees the horse running away. ANNA (CONT'D) Oh no. No. No. No. Come back. No. No. No. No.... Oooo-kay. He doesn't come back. Anna grabs onto a branch of a leaning conifer, tries to pull herself to her feet, but the tree snaps upright and releases all its snow onto her. GROAN. DISSOLVE TO: EXT. MOUNTAIN -- NIGHT The Northern Lights shine as Anna struggles, out of breath, reaching the top of a hill. 39 FROZEN - J. Lee ANNA Snow, it had to be snow, she couldn't have had tr-tr-tropical magic that covered the f-f-fjords in white sand and warm -- She sees smoke rising up in the distance. ANNA (CONT'D) Fire! WHOA! Anna goes tumbling down the hill. She lands with a crash in an icy stream at the bottom. ANNA (CONT'D) (from inside the snowball) Cold, cold, cold, cold, cold... EXT. A SMALL BUILDING AND STABLE -- NIGHT Anna shuffles up to the building, her dress frozen stiff. She shakes the snow off a sign and reads: ANNA Wandering Oaken's Trading Post. Snow drops off a smaller sign. She reads it, happily. ANNA (CONT'D) Ooh! And Sauna... INT. WANDERING OAKEN'S TRADING POST & SAUNA -- NIGHT Anna steps cautiously through the door--which hits her frozen butt and knocks her into the center of the shop. She looks around, sees only summer supplies. OAKEN (O.S.) Hoo hoo. Anna turns to see a bright-faced fellow sitting low behind the counter, fingers tapping tip to tip. OAKEN (CONT'D) Big summer blow out. Half off swimming suits, clogs, and a sun balm of my own invention, yah? ANNA Oh, great. For now, how about boots. Winter boots...and dresses? 40 FROZEN - J. Lee OAKEN (slight disappointment) That would be in our winter department. The winter department contains one outfit, a pick ax, and a lonely pair of boots. ANNA Oh. Um, I was just wondering; has another young woman, the Queen perhaps, I don't know, passed through here? She brings the clothes and boots to the counter. OAKEN Only one crazy enough to be out in this storm is you, dear? The front door suddenly blows open and in walks a mass of a man covered in ice. Underneath is KRISTOFF. OAKEN (CONT'D) You and this fellow.... Hoo hoo. Big summer blow out. Kristoff walks right up to Anna. KRISTOFF (in her face) Carrots. ANNA Huh? KRISTOFF Behind you. ANNA Oh, right. Excuse me. Anna moves out of Kristoff's way. He grabs a bunch of carrots, tosses them on the counter, then moves through the place, gathering other supplies. OAKEN (to Kristoff) A real howler in July, yah? Where ever could it be coming from? KRISTOFF The North Mountain. 41 FROZEN - J. Lee ANNA (to herself) North Mountain. Kristoff brings his supplies to the counter. Oaken counts on his fingertips. OAKEN That'll be forty. KRISTOFF Forty? No, ten. OAKEN (sweet as pie) Oh dear, that's no good. See these are from our winter stock, where supply and demand have a big problem. KRISTOFF You want to talk about a supply and demand problem? I sell ice for a living. Kristoff motions out the window, where we see the blocks of ice on his sled, covered in snow. ANNA Ooh, that's a rough business to be in right now. I mean, that is really... (he shoots her a look) Ahem. That's unfortunate. OAKEN Still forty. But I will throw in a visit to Oaken's sauna. Hoo hoo! Hi, family. Kristoff and Anna turn to see a naked family waving through the window of the steaming sauna. NAKED FAMILY Hoo hoo! KRISTOFF ...Ten's all I got. Help me out. OAKEN (isolating the carrots) Ten will get you this and no more. Kristoff seethes. Stalemate. 42 FROZEN - J. Lee ANNA Okay, just tell me one thing; what was happening on the North Mountain? Did it seem magical? Kristoff pulls down his scarf and gives Anna a firm answer. KRISTOFF Yes! Now, back up while I deal with this crook here. Oaken stands up, revealing his seven-foot stature. OAKEN What did you call me? EXT. WANDERING OAKEN'S TRADING POST AND SAUNA -- NIGHT Oaken stomps out the door, carrying Kristoff with one arm. KRISTOFF Okay. Okay, I'm- Ow! Whoa! Oaken throws Kristoff, who face-plants in the snow. OAKEN Bye bye. Oaken slams the door. Kristoff sits up. His reindeer, Sven, canters over, snorts, and nudges him, expectantly. KRISTOFF No Sven, I didn't get your carrots. Sven huffs in his face. Kristoff turns away and sees something. He points to a dilapidated barn. KRISTOFF (CONT'D) But I did find us a place to sleep. And it's free. INT. WANDERING OAKEN'S TRADING POST AND SAUNA -- NIGHT Anna stands watching Oaken and all his great height as he squeezes behind the counter and sits down low again. OAKEN (teddy bear) I'm sorry about this violence. I will add a quart of lutefisk, so we'll have good feelings. Just the outfit and boots, yah? 43 FROZEN - J. Lee Anna looks between Kristoff's supplies and the door. CUT TO: INT. OAKEN'S STABLES - NIGHT Kristoff, now unfrozen, relaxes on a bed of hay, playing his lute and singing to (and for) Sven. "Reindeer(s) are Better than People" KRISTOFF REINDEERS ARE BETTER THAN PEOPLE. SVEN, DON'T YOU THINK THAT'S TRUE? KRISTOFF (AS SVEN) (CONT'D) (throwing his voice) YEAH, PEOPLE WILL BEAT YOU & CURSE YOU & CHEAT YOU. EVERY ONE OF EM'S BAD, EXCEPT YOU. (speaking) Oh, thanks, Buddy. (singing, as Kristoff) BUT PEOPLE SMELL BETTER THAN REINDEERS. SVEN, DON'T YOU THINK I'M RIGHT? (As Sven) THAT'S ONCE AGAIN TRUE, FOR ALL EXCEPT YOU. (As Kristoff) YOU GOT ME. LET'S CALL IT A NIGHT. (As Sven) GOOD NIGHT. (As Kristoff) DON'T LET THE FROSTBITE BITE. The door opens. Anna enters. ANNA Nice duet. Kristoff sits up with a start...sees who it is. KRISTOFF Oh, it's just you. What do you want? ANNA I want you to take me up the North Mountain. 44 FROZEN - J. Lee KRISTOFF I don't take people places. He lays back down, closes his eyes. ANNA Let me rephrase that... A sack of supplies lands in Kristoff's lap. KRISTOFF Umph. He sits up. Looks in the bag. ANNA Take me up the North Mountain.... Please. He eyes her. He clearly doesn't take orders. ANNA (CONT'D) Look, I know how to stop this winter. He considers, lies back down, pulls his hat over his eyes. KRISTOFF We leave at dawn.... And you forgot the carrots for Sven. A bag of carrots hits Kristoff in the face. KRISTOFF (CONT'D) Ugh! ANNA Oops. Sorry. Sorry. I'm sorry. I didn't-- (catching herself) We leave now. Right now. She steps back outside and waits, anxiously. Annoyed, Kristoff offers Sven a carrot. Sven has a bite. Then Kristoff has a bite, contemplating. SLAM CUT TO: EXT. MOUNTAIN HIGH -- NIGHT Sven races, top speed, up a narrow cliff, pulling the sled, which skids precariously. Kristoff mans the reins. Anna sits beside him. 45 FROZEN - J. Lee KRISTOFF (trying to scare Anna) Hang on! We like to go fast! ANNA (fearless) I like fast! Anna leans back and puts her feet up on the dashboard. KRISTOFF Whoa, whoa! Get your feet down. He pushes her feet down. KRISTOFF (CONT'D) This is fresh lacquer. Seriously, were you raised in a barn? Kristoff spits on the dash to clean it. The spit flies back and hits Anna in the face. ANNA (grossed out) Ew. No, I was raised in a castle. She wipes off her face. KRISTOFF So tell me, what made the Queen go all ice-crazy? ANNA ...Oh well, it was all my fault. I got engaged but then she freaked out because I'd only just met him, you know, that day. And she said she wouldn't bless the marriage-- KRISTOFF Wait. You got engaged to someone you just met? ANNA Yeah. Anyway, I got mad and so she got mad and then she tried to walk away, and I grabbed her glove-- KRISTOFF Hang on. You mean to tell me you got engaged to someone you just met?! 46 FROZEN - J. Lee ANNA Yes. Pay attention. But the thing is she wore the gloves all the time, so I just thought, maybe she has a thing about dirt. KRISTOFF Didn't your parents ever warn you about strangers? Anna eyes Kristoff up and down, then slides away from him. ANNA Yes, they did.... But Hans is not a stranger. KRISTOFF Oh yeah? What's his last name? ANNA ...Of-the-Southern-Isles? KRISTOFF What's his favorite food? ANNA ...Sandwiches. KRISTOFF Best friend's name? ANNA Probably John. KRISTOFF Eye color. ANNA Dreamy. KRISTOFF Foot size...? ANNA Foot size doesn't matter. KRISTOFF Have you had a meal with him yet? What if you hate the way he eats? What if you hate the way he picks his nose? ANNA Picks his nose? 47 FROZEN - J. Lee KRISTOFF And eats it. ANNA Excuse me, sir. He's a prince. KRISTOFF All men do it. ANNA Ew. Look it doesn't matter; it's true love. KRISTOFF Doesn't sound like true love. ANNA Are you some sort of love expert? KRISTOFF No. But I have friends who are. ANNA You have friends who are love experts.... I'm not buying it. Sven suddenly stops, ears perked in alarm. KRISTOFF (to Anna) Stop talking. ANNA No, no, no. I'd like to meet these-- Kristoff clamps his hand over Anna's mouth. KRISTOFF I mean it. SHHH. Kristoff stands, looks into the dark woods surrounding them. Sensing something behind them, he holds up his lantern. Its light reflects off...EYES. Several. KRISTOFF(CONT'D) Sven, go. Go! Sven takes off. ANNA What are they? KRISTOFF Wolves. 48 FROZEN - J. Lee Flashes of white dart through the woods. Kristoff hops into the back of the sled, grabs a torch. Lights it. ANNA Wolves. What do we do? KRISTOFF I've got this. You just...don't fall off and don't get eaten. ANNA But I wanna help. KRISTOFF No. ANNA Why not? KRISTOFF Because I don't trust your judgement. ANNA Excuse me?! A wolf jumps at them, but Kristoff kicks it off. KRISTOFF Who marries a man she just met? Anna grabs the lute, swings it right at Kristoff's head. ANNA It's true love! He screams, as she...BAM!...swings past Kristoff and knocks a wolf away. KRISTOFF (shocked) Whoa. Just then Kristoff is yanked off the sled by another wolf. The torch goes flying. Anna catches it, shocked. ANNA Christopher! Kristoff grabs onto a loose rope hanging from the back of the sled and holds on for dear life as he's dragged behind. KRISTOFF It's Kristoff! 49 FROZEN - J. Lee A wolf jumps on Kristoff's back. KRISTOFF (CONT'D) AH! Anna thinks fast, uses the torch to light a blanket on fire. ANNA Duck! Anna throws the flaming blanket right at him. He ducks. The blanket hits the wolves. They tumble off Kristoff. KRISTOFF You almost set me on fire! Anna reaches out a hand, pulls Kristoff back onto the sled. ANNA But I didn't. Sven cries out. There is a massive gorge ahead. ANNA (CONT'D) Get ready to jump, Sven! KRISTOFF You don't tell him what to do! Kristoff shoves a satchel into her arms then scoops her up. KRISTOFF (CONT'D) I do! Kristoff tosses Anna onto Sven, then unhooks Sven's harness from the sled. KRISTOFF (CONT'D) Jump, Sven! Sven jumps the gorge with Anna on his back. Kristoff goes flying off behind them, still on the sled. Anna and Sven land safely on the other side of the gorge. Kristoff's sled loses momentum. It's not going to make it. He leaps off. He flaps his arms, claws at the air. He slams into the snowy edge of the cliff. Hanging by his hands, he looks down to see his sled hit the ground far below and burst into flames. 50 FROZEN - J. Lee KRISTOFF (CONT'D) (shocked sadness) ...But I just paid it off. Suddenly, he starts to slip. He claws at the loose snow, but it's clearly hopeless. He's going down. KRISTOFF (CONT'D) Uh-oh. No, no, no. To make matters worse, an AXE comes flying right at his face. KRISTOFF (CONT'D) AH! NO, NO, NO! The axe slams into the snow, inches from his nose. ANNA (O.S.) Grab on! Kristoff grabs on. ANNA (CONT'D) Pull, Sven! Pull! REVEAL: The axe is tied to a rope, then wrapped around Sven. Anna helps Sven pull Kristoff to safety. Kristoff rolls onto his back, exhausted. Anna peeks down at the burning sled. ANNA (CONT'D) Whoa.... I'll replace your sled and everything in it. Kristoff groans. ANNA (CONT'D) And I understand if you don't want to help me anymore. Anna walks off, sadly. Sven comes over and nuzzles Kristoff. KRISTOFF Of course I don't want to help her anymore. In fact, this whole thing has ruined me for helping anyone ever again. KRISTOFF (AS SVEN) (CONT'D) But she'll die on her own. KRISTOFF (AS SELF) (CONT'D) I can live with that. 51 FROZEN - J. Lee Through their conversation, they watch Anna go the wrong way...turn, go the other wrong way, turn, trip... KRISTOFF (AS SVEN) (CONT'D) But you won't get your new sled if she's dead. KRISTOFF (CONT'D) (knowing he's got a point) ...You know sometimes I really don't like you. Sven licks Kristoff happily. KRISTOFF (AS SELF) (CONT'D) (to Anna) Hold up. We're coming?! ANNA (excited) You are?! (catching herself) I mean, sure. I'll let you tag along. DISSOLVE TO: EXT. SHARP MOUNTAIN RIDGE -- DAWN Kristoff, Sven and Anna walk on a narrow rim of a mountain. DISSOLVE TO: EXT. MOUNTAIN FOREST CLEARING -- DAY As they step out of the thick trees, Anna catches sight of something far below. ANNA Arendelle. KRISTOFF It's completely frozen. ANNA ...But it'll be fine. Elsa will thaw it. KRISTOFF Will she? 52 FROZEN - J. Lee ANNA (uncertain) ...Yeah. Now come on. This way to the North Mountain? She points straight ahead. KRISTOFF More like this way. He points her finger up towards a perilously mighty mountain. DISSOLVE TO: INT. FROZEN WILLOW TREES -- DAY Anna, Kristoff, and Sven walk beneath frozen willows. The hanging branches glisten like Christmas lights. Sven knocks them with his antlers. They tinkle like chimes. ANNA I never knew winter could be so beautiful. Suddenly, a voice comes in from nowhere. We'll call that voice OLAF. OLAF (O.S.) YEAH...It really is beautiful, isn't it? But it's so white. You know, how about a little color? Must we bleach the joy out of it all? I'm thinking like maybe some crimson, chartreuse... While this is going on, Anna and Kristoff look around for the source of the rambling. They look at Sven - could he actually be talking? Sven looks back at them, his antlers tangled in branches, just as baffled as they are. In the meantime, a nose-less snowman, Olaf, wanders up behind them. OLAF (CONT'D) How `bout yellow--no, not yellow. Yellow and snow? Brrrr...no go. He stops between Kristoff and Anna. They look down at him. How did he get there? He suddenly looks up at Anna. OLAF (CONT'D) Am I right? 53 FROZEN - J. Lee Anna SCREAMS! Reflexes take over and she kicks Olaf's head, sending it flying off his body and into Kristoff's arms. OLAF (CONT'D) (cheery, to Kristoff) Hi! KRISTOFF You're creepy. Kristoff tosses the head back to Anna and they commence a game of hot potato. ANNA I don't want it! KRISTOFF Backatchya! OLAF Please don't drop me. ANNA Don't! KRISTOFF Come on, it's just a head. ANNA No! Olaf's body runs at Anna, arms waving. OLAF (O.S.) All right, we got off to a bad start. ANNA Ew, ew, the body! Anna slams Olaf's head back on the body, upside down. Olaf smiles happily, then looks confused. OLAF Wait, what am I looking at right now? Why are you hanging off the earth like a bat? ANNA (sympathetic) ...Okay. Wait one second. Anna kneels in front of Olaf and rights his head. 54 FROZEN - J. Lee OLAF Oooh! Thank you! ANNA You're welcome. OLAF Now I'm perfect. She looks over his innocent face, gets an idea. ANNA Well, almost. She digs into Kristoff's satchel, holds up a carrot just as Olaf turns toward her. The carrot accidentally slams all the way through his head. OLAF Woo! Head rush! ANNA Oh! Too hard. I'm sorry! I-I, I was just.... Are you okay? Olaf sees a tiny piece of carrot sticking out between his eyes. He lights up. OLAF Are you kidding me? I am wonderful! I've always wanted a nose. (going cross-eyed to look at his tiny nose) So cute. It's like a little baby unicorn. Anna reaches behind Olaf to the bulk of the carrot sticking out the back of his head, and pushes it forward. OLAF (CONT'D) What? Hey! Whoa. (seeing his now big nose) Oh, I love it even more! Hah.... All right, let's start this thing over. Hi everyone. I'm Olaf. And I like warm hugs. Olaf opens his arms wide to Anna. That triggers a memory. It takes her a moment to place it, but then she does. ANNA Olaf?...That's right, Olaf. 55 FROZEN - J. Lee OLAF ...And you are? ANNA Oh, um...I'm Anna. OLAF And who's the funky-looking donkey over there? ANNA That's Sven. OLAF Uh-huh. And who's the reindeer? ANNA ...Sven. Olaf looks from Kristoff to Sven, confused. OLAF Oh. They're--oh, okay.... (accepting it) Makes things easier for me. Sven tries to bite Olaf's nose. OLAF (CONT'D) Ha. Aw, look at him tryin' to kiss my nose. (gushes) I like you, too! ANNA Olaf, did Elsa build you? OLAF Yeah. Why? Curious, Kristoff takes one of Olaf's twig arms off, studies it. It seems to be moving in sync with his other arm. ANNA Do you know where she is? KRISTOFF (studying the arm) Fascinating... OLAF Yeah. Why? 56 FROZEN - J. Lee ANNA Do you think you could show us the way? OLAF Yeah. Why? KRISTOFF (bending the arm) How does this work? Olaf's dismembered arm slaps Kristoff across the face. OLAF Stop it, Sven. Trying to focus here. (to Anna) Yeah, why? KRISTOFF I'll tell you why. We need Elsa to bring back summer. OLAF (shocked) Summer? (sinking into wistfulness) Oh, I don't know why but I've always loved the idea of summer, and sun, and all things hot. KRISTOFF Really? I'm guessing you don't have much experience with heat. OLAF Nope. But sometimes I like to close my eyes and imagine what it'd be like when summer does come. DISSOLVE TO: OLAF'S FANTASY WORLD -- PERFECT SUMMER DAY Olaf walks through a grassy meadow with the sun shining behind him. He SINGS. "In Summer" OLAF BEES'LL BUZZ / KIDS'LL BLOW DANDELION FUZZ / AND I'LL BE DOING WHATEVER SNOW DOES IN SUMMER. 57 FROZEN - J. Lee -Olaf now lies in the sand on a beach. OLAF (CONT'D) A DRINK IN MY HAND / MY SNOW UP AGAINST THE BURNING SAND / PROB'LY GETTING GORGEOUSLY TANNED IN SUMMER. -Olaf sails in a boat. OLAF (CONT'D) I'LL FINALLY SEE A SUMMER BREEZE / BLOW AWAY A WINTER STORM / -Olaf floats in the water. All his pieces begin to separate. OLAF (CONT'D) AND FIND OUT WHAT HAPPENS TO SOLID WATER / WHEN IT GETS WARM. -Olaf tumbles on a sandy beach with sand-snowmen. OLAF (CONT'D) AND I CAN'T WAIT TO SEE / WHAT MY BUDDIES ALL THINK OF ME / JUST IMAGINE HOW MUCH COOLER I'LL BE IN SUMMER . . ! -Olaf and the seagull break out into a tap-dance. OLAF (CONT'D) DA DA . . . DA DOO / AH BAH BAH BAH BAH BAH BOO. -Olaf and another snowman drink hot chocolate in a hot tub. OLAF (CONT'D) THE HOT AND THE COLD ARE BOTH SO INTENSE / PUT `EM TOGETHER, IT JUST MAKES SENSE! -Olaf tap dances with a gaggle of seagulls. OLAF (CONT'D) RATDADAT DAD DADA DOO . . . -Olaf bounds down a grassy hill. OLAF (CONT'D) WINTER'S A GOOD TIME TO STAY IN AND CUDDLE / BUT PUT ME IN SUMMER AND I'LL BE A... He stops at a puddle, looks down at it. Smiles. Hops over it. 58 FROZEN - J. Lee OLAF (CONT'D) HAPPY SNOWMAN! -Olaf runs with a checkered blanket that he spreads out. He relaxes and stares at the blue sky. OLAF (CONT'D) WHEN LIFE GETS ROUGH I LIKE TO HOLD ON TO MY DREAM / OF RELAXING IN THE SUMMER SUN JUST LETTING OFF STEAM! Sven, Anna, Kristoff and Olaf have a picnic. OLAF (CONT'D) OH THE SKY WILL BE BLUE / AND YOU GUYS'LL BE THERE TOO / WHEN I FINALLY DO WHAT FROZEN THINGS DO IN SUMMER! KRISTOFF I'm gonna tell him. ANNA Don't you dare. OLAF IN SUMMER! Olaf sings the final note. We swing around him and return to: REALITY. He then straightens up and smiles. OLAF (CONT'D) So, come on! Elsa's this way. Let's go bring back summer! Olaf grabs Anna's hand and pulls her along up the mountain. ANNA (laughing) I'm coming! Sven hops along, happily following them. Kristoff watches all of them like they're nuts. KRISTOFF Somebody's got to tell him. DISSOLVE TO: 59 FROZEN - J. Lee EXT. ARENDELLE, VILLAGE -- DAY A layer of solid ice coats everything. People huddle around weak fires. Anxiety runs high amongst the villagers and guests. We pass two CITIZENS fighting over a woodpile. CITIZEN ONE No. No. You've got the bark facing down. The bark needs to be face-up. CITIZEN TWO Bark down is drier. CITIZEN ONE Bark up. CITIZEN TWO Bark down. CITIZEN ONE Bark up. Like a light in the dark, Hans moves through the crowd. HANS Cloak. Does anyone need a cloak? GERDA Arendelle is indebted to you, Your Highness. HANS The castle is open. There's soup and hot glogg in the Great Hall. He hands the stack of cloaks to a guard. HANS (CONT'D) Here. Pass these out. Just then the Duke approaches Hans. DUKE Prince Hans, are we just expected to sit here and freeze while you give away all of Arendelle's tradable goods? HANS (tall and confident) Princess Anna has given her orders and-- 60 FROZEN - J. Lee DUKE And that's another thing; has it dawned on you that your princess may be conspiring with a wicked sorceress to destroy us all? Hans's nice eyes turn to threatening slits. HANS Do not question the Princess. She left me in charge, and I will not hesitate to protect Arendelle from treason. DUKE (flabbergasted, offended) Treason?! Suddenly they hear the alarmed whinny of Anna's horse. It returns alone, bucking and kicking. Hans grabs its reins. HANS Whoa! Whoa! Whoa, boy. Easy. Easy. CROWD (various) Princess Anna's horse. What happened to her? Where is she? Hans steadies the horse, looks up at the mountain. He sees all the panicked faces of the kingdom looking to him. HANS ...Princess Anna is in trouble. (calling out) I need volunteers to go with me to find her! Volunteers, some from Arendelle, some from other lands, rush up to offer their services. DUKE I volunteer two men, my Lord! (quietly to his thugs) Be prepared for anything, and should you encounter the Queen, you are to put an end to this winter. Do you understand? His two thugs sneer. CUT TO: 61 FROZEN - J. Lee EXT. THE NORTH MOUNTAIN -- DAY Anna, Kristoff, Sven, and Olaf move through hostile terrain. Wind-swept icicles face horizontal. KRISTOFF So how exactly are you planning to stop this weather? ANNA (confident) Oh, I am gonna talk to my sister. KRISTOFF That's your plan? My ice business is riding on you talking to your sister. ANNA Yup. Kristoff, so stunned by her casual plan, doesn't look where he's going and ends up with an ice-spike to the nose. He stops short, GULP, moves carefully around the spike. KRISTOFF So you're not at all afraid of her? ANNA Why would I be? OLAF (oblivious) Yeah. I bet Elsa's the nicest, gentlest, warmest person ever. Olaf backs right into an icicle. It runs through his torso. OLAF (CONT'D) Oh, look at that. I've been impaled. He laughs it off. DISSOLVE TO: EXT. STEEP MOUNTAIN FACE -- DAY Anna and Kristoff hit what looks like a dead end. The face of the mountain goes straight up. ANNA What now? 62 FROZEN - J. Lee Kristoff looks around, sighs. Digs in his rucksack. KRISTOFF ...It's too steep. I've only got one rope, and you don't know how to climb mountains. ANNA (O.S.) Says who? Sven nudges Kristoff, who looks up to see Anna trying to climb the cliff's flat face. KRISTOFF (finding her ridiculous) What are you doing? ANNA (straining) ...I'm going to see my sister. KRISTOFF You're going to kill yourself. Kristoff watches her searching for footholds and hand-holds. KRISTOFF (CONT'D) I wouldn't put my foot there. ANNA (O.S.) You're distracting me. KRISTOFF Or there. How do you know Elsa even wants to see you? ANNA (O.S.) I'm just blocking you out cause I gotta concentrate here. KRISTOFF You know, most people who disappear into the mountains want to be alone. ANNA (O.S.) Nobody wants to be alone. Except maybe you-- KRISTOFF I'm not alone.... I have friends, remember? Anna kicks a foot above her head to catch a foot hold. 63 FROZEN - J. Lee ANNA You mean the love experts? KRISTOFF Yes, the love experts! Anna realizes she's stuck. ANNA ...Please tell me I'm almost there. REVEAL: she's only about six feet up. Her muscles shake. ANNA (CONT'D) ...Does the air seem a bit thin to you up here? Kristoff smiles, getting a kick out of her. KRISTOFF Hang on. He pulls the rope from his bag. Just then Olaf steps out from behind a rock and waves to Kristoff. OLAF Hey, Sven? Not sure if this is going to solve the problem, but I found a staircase that leads exactly where you want it to go. ANNA Ha ha. Thank goodness. Catch! Anna drops off the cliff. Kristoff catches her. ANNA (CONT'D) Thanks! That was like a crazy trust exercise. She hops down, brushes off her dress, and bounds off. Kristoff watches after her, digging her fearless pluck. EXT. BASE OF THE ICE PALACE -- DAY Anna, Kristoff, and Olaf approach Elsa's elegant ice palace. ANNA Whoa. KRISTOFF (in awe) Now that's ice. I might cry. 64 FROZEN - J. Lee ANNA Go ahead. I won't judge. Anna climbs the steps with Olaf. Sven tries to follow. His hooves slip out. He scrambles but can't get traction. Kristoff runs to his aide. KRISTOFF All right, take it easy. I gotcha. Kristoff settles Sven back down the stairs and pats him. KRISTOFF (CONT'D) You stay right here, buddy. Sven obediently plops his reindeer butt down and wags his tail. Kristoff climbs the stairs, admiring the ice details. KRISTOFF (CONT'D) ...Flawless. Anna arrives at the door. Hesitates. OLAF ...Knock.... (she doesn't) Just knock.... (she doesn't. To Kristoff) Why isn't she knocking...? Do you think she knows how to knock? Anna finally KNOCKS. The sound echoes inside. The ice doors slide open. ANNA Ha. It opened. That's a first. Anna goes to step in. Kristoff follows. She gets a thought, stops him. ANNA (CONT'D) You should probably wait out here. KRISTOFF What? ANNA Last time I introduced her to a guy, she froze everything. KRISTOFF But, it's a palace made of ice. Ice is my life. 65 FROZEN - J. Lee OLAF Bye, Sven. Olaf starts to head inside. Anna stops him. ANNA You too, Olaf. OLAF Me? ANNA Just give us a minute. OLAF Okay. As Anna walks inside. Olaf starts counting. OLAF (CONT'D) One...two... Kristoff joins in. OLAF AND KRISTOFF Three...four... INT. ELSA'S PALACE -- DAY Anna walks into a great foyer. The place is beautiful, but also eerie. ANNA Elsa? It's me...Anna?! Anna slips. Steadies herself. ELSA (O.S.) Anna. Elsa steps out of the shadows onto a balcony. She sees Anna, looks to her longingly. Anna can't help but be struck by Elsa's beauty. ANNA Elsa, you look different.... It's a good different.... And this place is amazing. 66 FROZEN - J. Lee ELSA (cautious, polite) Thank you, I never knew what I was capable of. Anna starts to climb the stairs. ANNA ...I'm so sorry about what happened. If I'd known-- Elsa backs up, away from Anna. ELSA (on guard) No, it's okay. You don't have to apologize.... But you should probably go, please. ANNA But I just got here. ELSA ...You belong in Arendelle. ANNA So do you. Anna takes another step up. Elsa backs up more. ELSA No, I belong here. Alone. Where I can be who I am without hurting anybody. ANNA ...Actually, about that-- OLAF (O.S.) 58...59...60. ELSA Wait. What is that? Olaf comes running in the front door. He waves. OLAF Hi, I'm Olaf and I like warm hugs. ELSA (shocked) Olaf? Olaf stops beside Anna, looks up at Elsa, intimidated. 67 FROZEN - J. Lee OLAF (bashful) You built me. You remember that? ELSA (astonished) And you're alive? OLAF Um...I think so? Anna kneels down beside Olaf. ANNA He's just like the one we built as kids.... We were so close. We can be like that again. Elsa smiles, but then a memory returns to her. FLASH CUT TO: FLASHBACK: Young Anna is struck by Elsa's powers. YOUNG ELSA Anna! Young Anna falls unconscious. Young Elsa races to her. FLASH CUT TO: THE PRESENT: Elsa's face sinks in pain. ELSA No, we can't. Elsa turns and heads up the second story steps. ELSA (CONT'D) Goodbye, Anna. ANNA Elsa, wait-- ELSA (calling back) I'm just trying to protect you. Elsa continues to flee. Anna pursues. ANNA You don't have to protect me. I'm not afraid. Please don't shut me out again. 68 FROZEN - J. Lee Anna SINGS. "First Time in Forever, Reprise" ANNA (CONT'D) PLEASE DON'T SLAM THE DOOR. YOU DON'T HAVE TO KEEP YOUR DISTANCE ANYMORE. `CAUSE FOR THE FIRST TIME IN FOREVER, I FINALLY UNDERSTAND. FOR THE FIRST TIME IN FOREVER, WE CAN FIX THIS HAND IN HAND. WE CAN HEAD DOWN THIS MOUNTAIN TOGETHER. YOU DON'T HAVE TO LIVE IN FEAR. `CAUSE FOR THE FIRST TIME IN FOREVER, I WILL BE RIGHT HERE. They arrive on the top floor, Elsa's main living space. Elsa turns back to Anna, grateful, but determined. ELSA Anna, PLEASE GO BACK HOME. YOUR LIFE AWAITS. GO ENJOY THE SUN AND OPEN UP THE GATES. ANNA Yeah, but-- ELSA I know! YOU MEAN WELL, BUT LEAVE ME BE. YES, I'M ALONE BUT I'M ALONE AND FREE. Elsa opens up the balcony doors. ELSA (CONT'D) JUST STAY AWAY AND YOU'LL BE SAFE FROM ME. ANNA ACTUALLY, WE'RE NOT. ELSA WHAT DO YOU MEAN YOU'RE NOT? 69 FROZEN - J. Lee ANNA I GET THE FEELING YOU DON'T KNOW? ELSA WHAT DO I NOT KNOW? ANNA ARENDELLE'S IN DEEP DEEP DEEP DEEP SNOW. ELSA What? Elsa looks past Anna's shoulder out white-peaked mountains. ANNA You kind of set off an eternal winter...everywhere. ELSA Everywhere? ANNA It's okay, you can just unfreeze it. ELSA No, I can't. I don't know how. ANNA Sure you can. I know you can. Snow starts to swirl around the room. ANNA (CONT'D) CUZ FOR THE FIRST TIME IN FOREVER, ELSA (panicking) I'M SUCH A FOOL! I CAN'T BE FREE! ANNA YOU DON'T HAVE TO BE AFRAID. ELSA NO ESCAPE FROM THE STORM INSIDE OF ME! The snow picks up. Anna tries to fight through it. ANNA WE CAN WORK THIS OUT TOGETHER. 70 FROZEN - J. Lee ELSA I CAN'T CONTROL THE CURSE! ANNA WE'LL REVERSE THE STORM YOU'VE MADE. ELSA ANNA, PLEASE, YOU'LL ONLY MAKE IT WORSE! ANNA DON'T PANIC. ELSA THERE'S SO MUCH FEAR! ANNA WE'LL MAKE THE SUN SHINE BRIGHT. ELSA YOU'RE NOT SAFE HERE! ANNA WE CAN FACE THIS THING TOGETHER... But as Anna sings, we lose sight of her in the thickening blizzard taking over the room. ELSA NO! ANNA (O.S.) WE CAN CHANGE THIS WINTER WEATHER, AND EVERYTHING WILL BE... Anna's voice disappears in the storm as Elsa cries out. ELSA I CAN'T! Elsa's fear, so strong, sucks the blizzard back into her and then it bursts out, unwittingly, like a sharp snowflake. Anna is STRUCK right in the heart. She grasps her chest in pain and stumbles back. She falls to her knees. Elsa gasps when she sees Anna. Just then, Olaf and Kristoff rush into the room to Anna's side. KRISTOFF Anna. Are you okay? 71 FROZEN - J. Lee ANNA I'm okay.... I'm fine. Anna gets to her feet, determined to hide the pain. ELSA (scared) Who's this? Wait, it doesn't matter. You have to go. ANNA No, I know we can figure this out together-- ELSA (desperate) How? What power do you have to stop this winter? To stop me? Anna doesn't have the answer. Kristoff sees spiky ice shadows creeping down the walls. Puts a protective arm around Anna. KRISTOFF Anna, I think we should go. ANNA (close to tears) No. I'm not leaving without you, Elsa. ELSA (heartbroken but decisive) Yes, you are. Elsa waves her arms and builds a giant, menacing snowman. We'll call him MARSHMALLOW. SLAM CUT TO: EXT. ICE PALACE -- DAY Marshmallow holds Anna and Kristoff by the scruff of their necks in one hand and Olaf in the other. ANNA Stop. Put us down! OLAF (to Marshmallow) You are a lot stronger than I think you realize. Marshmallow tosses Kristoff and Anna down the steps. 72 FROZEN - J. Lee MARSHMALLOW (like a bouncer) Go away! Anna and Kistoff slide past Sven, who's got his tongue stuck to the ice railing. OLAF (O.S.) Heads up! Olaf's head smashes into a snowbank nearby. ANNA Olaf! OLAF Watch out for my butt! Anna and Kristoff duck as the rest of Olaf slams into the snowbank. Marshmallow turns to go back into the castle. Incensed, Anna tries to march back up the stairs. ANNA It is not nice to throw people! Kristoff grabs her, pulls her back. KRISTOFF ANNA All right feisty pants. Calm Let me at him. I want to get down. Woaw. Just let the snow him. I.... Okay. I'm Calm. man be. Anna backs down...for a moment. Then she grabs a snowball and throws it at Marshmallow. The tiny little ball hits Marshmallow's back, not making even the slightest dent. But it's enough to infuriate him. He ROARS. Spikes shoot out of his joints. KRISTOFF Uh-oh. Now you made him mad! OLAF ...I'll distract him. You guys go. Kristoff pushes Anna along. Sven runs off in the opposite direction. Olaf's belly and butt fall and follow Sven. OLAF (CONT'D) No, no, not you guys. 73 FROZEN - J. Lee Marshmallow goes charging after Anna and Kristoff as Olaf's head falls and lands face down in snow. OLAF (CONT'D) (muffled) This just got a whole lot harder. Anna and Kristoff leap and slide down a steep slope. They tumble to a stop at the bottom just as Marshmallow lands hard right behind them. They're off again...through a maze of conifers that sag under the weight of the snow, Marshmallow hot on their trail. KRISTOFF This way! Anna grabs a branch of a sagging trees and releases all of the snow. The tree snaps upright, knocking Marshmallow back. KRISTOFF (CONT'D) (impressed) Ho-ho-ho! ANNA I got him! Anna and Kristoff burst out of the conifer forest and almost run right off a cliff. They stop short, toes on the edge. KRISTOFF Whoa, stop! ANNA It's a hundred foot drop. KRISTOFF It's two hundred. Kristoff ties the rope around Anna and pulls tight. ANNA Ow. He drops to his knees and starts digging a u-shape in the snow with a pick axe. ANNA (CONT'D) What's that for? KRISTOFF I'm digging a snow anchor. 74 FROZEN - J. Lee ANNA (not trusting) Okay. What if we fall? KRISTOFF There's twenty feet of fresh powder down there; it'll be like landing on a pillow.... Hopefully. They hear an angry ROAR coming closer. KRISTOFF (CONT'D) Okay, Anna. On three. Anna preps for the jump like a boxer getting ready to fight. ANNA Okay. You tell me when... KRISTOFF One... ANNA ...I'm ready to go.... KRISTOFF Two... ANNA (pumped up) ...I was BORN ready! Yes! KRISTOFF Calm down. A huge tree flies through the air toward them. ANNA (O.S.) TREE! Anna jumps and pulls Kristoff over the edge with her. They hang upside down over the cliff by the rope. The rope catches their fall. KRISTOFF Whoa! That happened. Back up top, Olaf emerges from the woods. He's a complete mess, all his body parts are in the wrong places. He huffs and puffs, struggling to run. OLAF Ah. Ah. Man, am I out of shape. 75 FROZEN - J. Lee He stops. Puts his body back together in the right order. OLAF (CONT'D) There we go. Hey, Anna! Sven! Where'd ya guys go? We totally lost Marshmallow back there! Marshmallow steps up behind Olaf. Olaf turns to face him. OLAF (CONT'D) (happily) Hey. We were just talking about you. All good things, all good things. Marshmallow roars and approaches Kristoff's snow anchor. OLAF (CONT'D) NO! Olaf jumps onto Marshmallow's leg trying to stop him, but not making much of a difference. OLAF (CONT'D) This is not making much of a difference! Marshmallow flicks Olaf off his leg and right over the cliff. OLAF (CONT'D) WHOA! Olaf passes Anna and Kristoff. ANNA Olaf! OLAF Hang in there, guys! Marshmallow starts yanking Kristoff and Anna's rope up. ANNA Wait, what? Kristoff's head hits the cliff. KRISTOFF Aargghh! Kristoff passes out and hangs like a rag doll. ANNA Kristoff! 76 FROZEN - J. Lee Marshmallow pulls them up. He roars and breathes snow all over them. MARSHMALLOW Don't come back! ANNA (grossed out by his snow breath) Ugh. We won't. Anna whips out a knife and cuts the rope. Kristoff comes to just as they fall. They both SCREAM! SLAM! REVEAL: Anna opens her eyes to find herself buried up to her shoulders in the soft thick snow. She laughs. ANNA (CONT'D) Hey, you were right. Just like a pillow. She looks up to see Olaf's upper half hanging onto Kristoff's boots, which are sticking out of the snow. OLAF (shaking the boots) I can't feel my legs! I can't feel my legs! Suddenly, Kristoff's head pops up. He spits out snow. KRISTOFF Those are my legs. Olaf's bottom goes running by. OLAF (to Kristoff) Ooh. Hey, do me a favor, grab my butt. Kristoff grabs Olaf's head and puts it on his body. OLAF (CONT'D) Oh, that feels better. Sven walks up and sniffs Olaf's nose. OLAF (CONT'D) Hey, Sven! 77 FROZEN - J. Lee Olaf turns to Anna and Kristoff just as Sven goes to bite off his nose -- and misses. OLAF (CONT'D) He found us. (to Sven, funny voice) Who's my cute little reindeer? KRISTOFF Don't talk to him like that. Kristoff goes over to help Anna, who is stuck in the snow. KRISTOFF (CONT'D) Here. He lifts her out easily. ANNA (impressed) Whoa! KRISTOFF You okay? ANNA Thank you. They meet eyes. Wait. Is that chemistry? ANNA (CONT'D) ...Um.... How's your head? She touches the spot where he banged his head. KRISTOFF (in pain) Ah! Ooh! He catches himself. Waves off the pain with a giggle. KRISTOFF (CONT'D) I mean, It's fine. Ah...I'm good. Ha. I've got a thick skull. OLAF I don't have a skull.... Or bones. KRISTOFF ...So.... The awkwardness is killing him. 78 FROZEN - J. Lee KRISTOFF (CONT'D) (shy) Now what? ANNA (shy) Now what? (then...panicking) Now what?! Oh! What am I gonna do? She threw me out. I can't go back to Arendelle with the weather like this. And then there's your ice business-- KRISTOFF Hey, hey, don't worry about my ice business... (noticing something) Worry about your hair?! She thinks he means it looks bad. She smooths it down. ANNA What? I just fell off a cliff. You should see your hair. KRISTOFF No, yours is turning white. She grabs her braid as a tendril turns white. ANNA White? It's what? KRISTOFF It's because she struck you; isn't it? ANNA Does it look bad? KRISTOFF (thinking) ...No. Olaf's head pops up. He's holding his head up off his body to join the conversation. OLAF You hesitated. KRISTOFF No, I didn't. Anna, you need help. Now, come on. 79 FROZEN - J. Lee He heads towards the sunset. Sven and Olaf follow. OLAF Okay! Where are we going? KRISTOFF To see my friends. ANNA (catching up) The love experts? OLAF Love experts?! KRISTOFF Yes. And don't worry; they'll be able to fix this. ANNA How do you know? He looks her over, remembering the moment he saw the trolls heal her as a child. KRISTOFF ...Because I've seen them do it before. As they round the bend, the sun sets and Olaf turns to Sven. OLAF I like to consider myself a love expert. CUT TO: INT. ELSA'S PALACE -- DAY Elsa paces, distraught. She talks to herself. ELSA (mantra-style) Get it together. Control it. Don't feel. Don't feel. Don't FEEL! She hears ice cracking. Stops. Looks around. She's left a sharp wake of ice spikes behind her on the floor. They grow up the wall, taking over the castle. DISSOLVE TO: 80 FROZEN - J. Lee EXT. BLACK MOUNTAINS -- NIGHT The Northern Lights are bright. Olaf stares at them in awe as he rides on Sven's back. OLAF Look, Sven. The sky's awake. Behind Olaf and Sven, Anna walks with Kristoff. She shivers. KRISTOFF Are you cold? ANNA ...A little. He reaches like he might put an arm around her, but decides against it. He looks around as if he doesn't know what to do, then gets a thought. KRISTOFF Wait. Come here. He takes her hand and pulls her around a bend into a rock- lined pass. Steam vents, powered by the volcanic activity, dot the path. He holds her hands over one of them. ANNA Oooh.... That's nice. They continue on the path, walking from vent to vent. KRISTOFF (taking a deep breath) So, about my friends...well, I say friends, they're more like family.... Anyway, when I was a kid, it was just me and Sven...until they took me in. ANNA (moved) They did? KRISTOFF (nervous ramble) Yeah. I don't want to scare you, they can be a little bit inappropriate...and loud...very loud...they're also stubborn at times, and a little overbearing. And heavy. Really, really heavy. (MORE) 81 FROZEN - J. Lee KRISTOFF (CONT'D) But they're fine.. You'll get it. They mean well. Anna touches Kristoff's arm, reassuringly. ANNA Kristoff, they sound wonderful. Kristoff smiles, appreciating her sincerity. KRISTOFF Okay then.... Mustering the courage, Kristoff steps forward and with a wave of the arms announces-- KRISTOFF (CONT'D) Meet my family. REVEAL: he's surrounded by rocks. KRISTOFF (CONT'D) (to the rocks) Hey, guys! As Kristoff and Sven move through the rocks, waving and greeting, Olaf and Anna stand frozen, dumbfounded. ANNA (to herself) ...They're rocks. OLAF (realizing) He's crazy. (covertly, to Anna) I'll distract them while you run. (Loud and slow to a rock) Hi, Sven's family! It's nice to meet you! (quietly to Anna) Anna, because I love you, I insist you run. (to the rock) I understand you're love experts! (to Anna) Why aren't you running? Anna snaps out of her shock and starts backing away. ANNA Okay. Um...I'm gonna go-- Just then the rocks around her start rolling. 82 FROZEN - J. Lee ANNA (CONT'D) (panicking) Kristoff! Olaf lights up and chases the rocks, who surround Kristoff and unfold as trolls. BULDA KRISTOFF'S HOME! TROLLS (VARIOUS) Kristoff! Kristoff's home! It's been too long! Kristoff's home! Olaf jumps around all excitedly. OLAF (excitedly) Kristoff's home. He then stops, confused, and looks to one of the trolls. OLAF (CONT'D) Wait? Kristoff? Anna watches, shocked and confused. The trolls all want Kristoff's attention. One troll yanks him down with a boulder's strength. TROLL ONE Oh, lemme look at you! Another troll tries to pull off his clothes. TROLL TWO Oh, take off your clothes, Kristoff; I wash them. KRISTOFF (holding up his pants) Ah! No. I'm gonna keep my clothes on, thank you. KRISTOFF (CONT'D) Great to see you all. Where's grandpa? MUSHROOM KID TROLL He's napping. But look, I grew a mushroom. TROLL SCOUT KID And I earned my fire crystal. 83 FROZEN - J. Lee KIDNEY STONE TROLL I passed a kidney stone. PICK ME UP TROLL Pick me up. The kid troll jumps up on Kristoff's arm. Kristoff sinks under the weight of him. Anna still stares, confused, then realizes... ANNA Trolls? They're trolls. Silence. All troll eyes turn to Anna. Blink. Blink. BULDA ...He's brought a girl! TROLLS (TOGETHER) He's brought a girl! Suddenly Anna is surrounded by trolls. They body-surf/roll Anna over to Kristoff. She falls into his arms. ANNA What's going on? KRISTOFF I've learned to just roll with it. Bulda climbs on top of her husband, Cliff, to get a good look at Anna. She studies her like she's a piece of cattle. BULDA Let me see. Bright eyes. Working nose. Strong teeth. Yes, yes, yes. She'll do nicely for our Kristoff. ANNA Wait. Oh. Um. No. KRISTOFF You've got the wrong idea. That's not why I brought her here. ANNA Right. We're not. I'm not-- Anna laughs, uncomfortable, not knowing what to say. 84 FROZEN - J. Lee BULDA (to Anna) What's the issue, dear? Why are you holding back from such a man? Bulda SINGS. "Fixer-Upper" TROLLS (VARIOUS) IS IT THE CLUMPY WAY HE WALKS? OR THE GRUMPY WAY HE TALKS? OR THE PEAR-SHAPED, SQUARE-SHAPED WEIRDNESS OF HIS FEET? AND THOUGH WE KNOW HE WASHES WELL HE ALWAYS ENDS UP SORTA SMELLY. BUT YOU'LL NEVER MEET A FELLA WHO'S AS SENSITIVE AND SWEET. TROLLS (CHORUS) (CONT'D) SO HE'S A BIT OF A FIXER UPPER, SO HE'S GOT A FEW FLAWS- HIS PECULIAR BRAIN, DEAR. HIS THING FOR THE REINDEER THAT OUTSIDE A FEW OF NATURE'S LAWS. SO HE'S A BIT OF A FIXER UPPER, BUT THIS WE'RE CERTAIN OF- YOU CAN FIX THIS FIXER UPPER UP WITH A LITTLE BIT OF LOVE. KRISTOFF Can we just stop talking about this?! We've got a real, actual problem here. BULDA I'll say-- (To Anna) IS IT THE WAY THAT HE RUNS SCARED? TROLLS (VARIOUS) OR THAT HE'S SOCIALLY IMPAIRED? KID TROLL OR THAT HE ONLY LIKES TO TINKLE IN THE WOODS? TROLLS (VARIOUS) ARE YOU HOLDING BACK YOUR FONDNESS DUE TO HIS UNMANLY BLONDENESS? OR THE WAY HE COVERS UP THAT HE'S THE HONEST GOODS? 85 FROZEN - J. Lee TROLLS (CHORUS) (CONT'D) HE'S JUST A BIT OF A FIXER UPPER- HE'S GOT A COUPLE A' BUGS. KRISTOFF No, I don't. TROLLS HIS ISOLATION IS CONFIRMATION OF HIS DESPERATION FOR HEALING HUGS. SO HE'S A BIT OF A FIXER UPPER, BUT WE KNOW WHAT TO DO. THE WAY TO FIX UP THIS FIXER UPPER IS TO FIX HIM UP WITH YOU. The girl trolls sweep Anna away. The boys take Kristoff. KRISTOFF (to the male trolls) Enough! She's engaged to someone else. Okay?! TROLLS beat. Blink. Blink. The boy trolls turn, huddle... TROLLS (VARIOUS) SO SHE'S A BIT OF A FIXER UPPER, THAT'S A MINOR THING. THIS QUOTE "ENGAGEMENT" IS A FLEX ARRANGEMENT. KID TROLL AND BY THE WAY, I DON'T SEE NO RING. TROLLS (VARIOUS) SO SHE'S A BIT OF A FIXER UPPER, HER BRAIN'S A BIT BETWIXT. GET THE FIANCE OUT OF THE WAY AND THE WHOLE THING WILL BE FIXED! GIRL TROLLS WE AREN'T SAYING YOU CAN CHANGE HIM TROLLS (VARIOUS) 'CAUSE PEOPLE DON'T REALLY CHANGE. WE'RE ONLY SAYING THAT LOVE'S A FORCE THAT'S POWERFUL AND STRANGE. PEOPLE MAKE BAD CHOICES IF THEY'RE MAD OR SCARED OR STRESSED. (MORE) 86 FROZEN - J. Lee TROLLS (VARIOUS) (CONT'D) BUT THROW A LITTLE LOVE THEIR WAY (THROW A LITTLE LOVE THEIR WAY) AND YOU'LL BRING OUT THEIR BEST! TRUE LOVE BRINGS OUT THE BEST! Kristoff looks over at Anna. She actually looks shockingly beautiful dressed in moss, lit by shimmering crystals. ALL TROLLS EVERYONE'S A BIT OF A FIXER UPPER, THAT'S WHAT IT'S ALL ABOUT FATHER, SISTER, BROTHER WE NEED EACH OTHER TO RAISE US UP AND ROUND US OUT By this time Kristoff and Anna are being ushered into a pit by the sheer force of numbers. TROLLS EVERYONE'S A BIT OF A FIXER UPPER, BUT WHEN PUSH COMES TO SHOVE- THE ONLY FIXER UPPER FIXER THAT CAN FIX A FIXER UPPER IS TRUE TRUE TRUE TRUE LOVE During this last bit Anna and Kristoff are looking at each other differently. Hmmm. Maybe those trolls are right? Sparks! Chemistry! TROLL PRIEST Do you, Anna, take Kristoff to be your trollfully wedded-- ANNA Wait, what?! TROLL PRIEST You're getting married. TROLLS LOVE! Just then, Anna collapses. Kristoff catches her. She's shivering something fierce. KRISTOFF Anna? He pulls off her cape and hat. 87 FROZEN - J. Lee KRISTOFF (CONT'D) She's as cold as ice. Just then Grand Pabbie pushes his way through the crowd. Trolls clear the way for Pabbie. He stops at the edge of the pit. GRAND PABBIE There's strange magic here! KRISTOFF Grand Pabbie! GRAND PABBIE Bring her to me, Kristoff. Kristoff helps Anna over. Pabbie looks into her weak eyes. GRAND PABBIE (CONT'D) Anna, your life is in danger. There is ice in your heart, put there by your sister. If not removed, to solid ice will you freeze, forever. ANNA What...? No. KRISTOFF So remove it, Grand Pabbie. GRAND PABBIE I can't. If it was her head, that would be easy. But only an act of true love can thaw a frozen heart. ANNA An act of true love? BULDA (googley, to her hubby) A true love's kiss, perhaps? A bunch of trolls give each other kisses. Anna shivers again, collapsing into Kristoff's arms. More of her hair turns white. KRISTOFF Anna, we've got to get you back to Hans. ANNA (still weak) ...Hans. 88 FROZEN - J. Lee KRISTOFF Help us out, Sven. Kristoff grabs Sven's antlers. Sven pulls them out. Kristoff helps Anna onto Sven and hops up behind her. KRISTOFF (CONT'D) Come on, Olaf! Sven takes off. Olaf grabs Sven's tail, rides with them. OLAF I'm coming! Let's go kiss Hans! Who is this Hans?! CUT TO: EXT. ELSA'S PALACE - DAWN Hans and the men tread cautiously towards the castle. HANS We are here to find Princess Anna. Be on guard, but no harm is to come to the Queen. Do you understand? The Duke's thugs exchange a look. Suddenly, a mass of snow rises from the ground behind Hans. It's Marshmallow, Elsa's snow guard. MARSHMALLOW Go away! He slams a fist inches from Hans. Hans deftly dodges out of the way. All of the guards take up arms against Marshmallow, who quickly knocks them over. Marshmallow throws down a guard and his horse, who topple over Hans. Marshmallow raises his foot to stomp on Hans, but Hans barrel-rolls himself to safety. He sees his sword, leaps, and grabs it. Just then, Elsa peeks out the front doors. The Duke's two thugs see her. DUKE'S THUG The Queen. The thugs charge up the stairs. 89 FROZEN - J. Lee INT. ELSA'S PALACE -- DAY They guards burst through the ice doors. Elsa flees to the top floor of her palace. The guards pursue. They trap her on the top floor, raise their crossbows. ELSA (scared) No. Please. One of the thugs shoots an arrow right at Elsa. At the last moment she creates an ice wall. It stops the arrow, inches from her face. The thugs reposition to take another shot. ELSA (CONT'D) Stay away! Elsa shoots ice at the thugs. They duck out of the way and continue the attack. THUG Get her! Get her! Elsa fights for her life. BACK OUTSIDE: Hans is nearly crushed by Marshmallow. He rolls away. Jumps to his feet. And with agile might, he slices Marshmallow's leg off with his sword. Marshmallow stumbles back, off balance. And falls off over the cliff, but not before striking Hans. Hans goes over the edge. REVEAL: Hans clings to the ice steps. His men help him up and they rush into the ice palace. INT. ICE PALACE -- DAY Elsa is surrounded. It's do or die. In two swift moves, Elsa traps one thug in a cage of spikes that threaten his neck. The other she pushes back with a wall of ice....up against the balcony doors...which BURST and CRACK. OUT ONTO THE BALCONY.... The balcony doors shatter. The thug is pushed to the edge. He's inches away from falling to his death. BACK INSIDE: Hans and his men run in. See the destruction and the thugs near death. 90 FROZEN - J. Lee HANS Queen Elsa! Don't be the monster they fear you are. Elsa snaps out of her rage. She sees the men, frightened, moments from death. She stops. Elsa looks to Hans, overwhelmed, frightened. The wall retreats from the thug on the balcony. The ice spikes lower from the second thug's neck. He takes advantage and aims his crossbow at Elsa's back. Seeing it. Hans runs and pushes the crossbow up just as the arrow releases. The arrow hits the ice chandelier, hanging directly above Elsa. The chandelier comes CRASHING DOWN. Elsa dives out of the way but she falls in the blast. All we see is ice smashing like glass, and all we hear is the sound of it shattering as it rings out. CUT TO BLACK. FADE IN ON: Elsa's face as her eyes flutter open. She sits up. She's surrounded by stone. INT. ARENDELLE, DUNGEON -- DAY Elsa looks to the nearby window. Tries to rush to it. She's pulled taut by giant shackles that fit like iron gloves. She's chained to the wall. Elsa strains to looks out a window... INSET WINDOW: Arendelle is outside, frozen solid and getting further buried under the ice and snow that is falling. ELSA No....What have I done? Hans enters. He hangs a torch by the door. ELSA (CONT'D) Why did you bring me here? HANS I couldn't just let them kill you. 91 FROZEN - J. Lee ELSA But I'm a danger to Arendelle. Get Anna. HANS Anna has not returned.... Elsa looks to the storm with worry. HANS (CONT'D) If you would just stop the winter, bring back summer...please. Elsa meets his eyes, desperate. ELSA Don't you see...I can't. Hans sees the sincerity in her eyes. ELSA (CONT'D) You have to tell them to let me go. Hans walks to the door. He takes the torch. HANS I will do what I can. He opens the door and leaves. Elsa, distraught, hears cracking. She looks down as her shackles begin to freeze over. The storm outside picks up. CUT TO: EXT. THE FJORDS -- DAY Sven charges down the mountain with Kristoff and Anna on his back. Olaf slides along beside them, penguin-style. Anna shivers in Kristoff's arms. She's weakening. Kristoff takes off his hat and puts it on her head. KRISTOFF Just hang in there. (to Sven) Come on, buddy, faster! They arrive at the walls of Arendelle. Olaf slides past them, out of control. OLAF I'll meet you guys at the castle! 92 FROZEN - J. Lee KRISTOFF Stay out of sight, Olaf! OLAF I will! He disappears into the village streets. OLAF (O.S.) (CONT'D) Hello! TOWNSWOMAN (O.S.) Ah! It's alive! CUT TO: EXT. CASTLE COURTYARD -- DAY Guards see Kristoff and Anna approaching. GUARD It's Princess Anna! Sven skids to a stop outside the gates. Kristoff slides off, holding Anna, and carries her to the gate. KRISTOFF I've got you. Anna looks up at him, gratefully. ANNA ...Are you g-gonna be okay? KRISTOFF (touched, reassuring) Don't worry about me. Just then the castle gates open. Gerda, Kai, and a handmaid rush to help Anna. GERDA Anna! Oh, you had us worried sick. KAI My Lady. You are freezing. GERDA You poor girl, you're freezing. Let's get you inside. 93 FROZEN - J. Lee KRISTOFF Get her warm and find Prince Hans, immediately. KAI We will. Thank you. Anna is swept away from Kristoff and into the palace grounds. KRISTOFF Make sure she's safe! Kristoff is shut out as the castle gates close on him. Kristoff stands there with Sven for a beat, staring with worry at the closed gates. Finally, he sighs, turns and walks off. Sven reluctantly follows. CUT TO: INT. LIBRARY -- DAY Hans stands with the dignitaries and guards. HANS I'm going back out to look for Princess Anna. FRENCH DIGNITARY You cannot risk going out there again. HANS If anything happens to her-- SPANISH DIGNITARY If anything happens to the Princess, you are all Arendelle has left. Hans hesitates, realizing how much this kingdom has come to depend on him. Is he really all they have left? Just then the door opens and Gerda and Kai bring in Anna. KAI He's in here. Prince Hans. HANS Anna. 94 FROZEN - J. Lee Hans rushes to Anna. She falls into his arms. HANS (CONT'D) You're so cold. ANNA (weak, but desperate) Hans, you have to kiss me. HANS What? ANNA Now. Here we go. She tries to kiss him, but is too weak to pull herself up in his arms. GERDA We'll give you two some privacy. Everyone shuffles out, leaving Hans and Anna alone. HANS What happened out there? ANNA Elsa struck me with her powers. HANS You said she'd never hurt you. ANNA I was wrong. Anna crumbles, weak. HANS Anna. Hans carries her to a couch, sets her down. ANNA (shivering more) She froze my heart and only an act of true love can save me. HANS (understanding) A true love's kiss. He takes her chin in his hand and gives her a tender smile. He leans in slowly...gently... 95 FROZEN - J. Lee Then he stops. HANS (CONT'D) Oh, Anna. If only there was someone out there who loved you. ANNA What? Hans gets up, leaving her there. ANNA (CONT'D) ...You said you did. He goes to the window and shuts the curtains. HANS As thirteenth in line in my own kingdom, I didn't stand a chance. I knew I'd have to marry into the throne somewhere-- ANNA What are you talking about? HANS (putting out the candles) As heir, Elsa was preferable, of course. But no one was getting anywhere with her. But you- ANNA Hans? HANS You were so desperate for love you were willing to marry me, just like that. Hans crosses the room, grabs a pitcher of water from a table and goes to the fireplace. HANS (CONT'D) I figured, after we married, I'd have to stage a little accident for Elsa. Hans pours the water on the fireplace, putting out the fire. Anna tries to stop him. She falls to the floor, weak. ANNA Hans. No, stop. 96 FROZEN - J. Lee HANS But then she doomed herself, and you were dumb enough to go after her. ANNA Please. HANS (chuckles) All that's left now is to kill Elsa and bring back summer. Hans approaches Anna. ANNA ...You're no match for Elsa. He bends down, takes her chin in his hand again, this time not so gently. HANS No, you're no match for Elsa. I, on the other hand, am the hero who is going to save Arendelle from destruction. She wrenches her face out of his hands. ANNA (anger) You won't get away with this. Hans rises and crosses to the door. HANS Oh, I already have. Hans leaves and shuts her in, locking the door. Anna struggles to the door, yanks on the locked handle. ANNA (hoarse and weak) Please, somebody help. The rest of her hair turns white and she crumbles to the floor. CUT TO: 97 FROZEN - J. Lee INT. COUNCIL CHAMBER -- NIGHT The Duke looks out the window at the growing snowstorm. He rubs his arms and shivers. DUKE It's getting colder by the minute. If we don't do something soon, we'll all freeze to death. Hans comes in, putting on his most distraught face. SPANISH DIGNITARY Prince Hans. HANS Princess Anna is...dead. VARIOUS DIGNITARIES What...? No.... Mon dieu. Hans stumbles, weak with grief. The men help him to a chair. DUKE What happened to her? HANS She was killed by Queen Elsa. DUKE Her own sister. HANS (really putting it on) At least we got to say our marriage vows...before she died in my arms. He bows his head in a brilliant display of teary grief. DUKE There can be no doubt now; Queen Elsa is a monster and we are all in grave danger. SPANISH DIGNITARY Prince Hans, Arendelle looks to you. Hans nods; he knows what he's being asked to do, and he'll do it with the perfect amount of authority and gravitas. 98 FROZEN - J. Lee HANS With a heavy heart, I charge Queen Elsa of Arendelle with treason and sentence her to death. INT. ELSA'S DUNGEON -- DAY The cell ices over. Elsa looks out at the storm that is devastating Arendelle, then hears the guards approaching. GUARD (O.S.) She's dangerous. Move quickly and with resolve. Elsa pulls at her shackles. They crack. Just as the door busts open, the weight of the ice crumbles the walls. The men duck out of the way. Hans pushes his way into the room...sees... The back wall is blown open. Broken shackles rest on the floor. Elsa is gone. CUT TO: EXT. MOUNTAIN SLOPE -- DAY Kristoff heads into the mountains. Sven lags behind, not wanting to follow. He looks back at the kingdom, then shakes his head. Enough. He runs past Kristoff. Stops and turns to face him. He snorts and grunts. KRISTOFF What is it, buddy? Sven nudges Kristoff with his antlers. KRISTOFF (CONT'D) Hey, watch it. What's wrong with you? Sven snorts with more conviction, moos, brays. KRISTOFF (CONT'D) (avoiding) ...I don't understand you when you talk like that. 99 FROZEN - J. Lee Kristoff tries to walk on ahead, but Sven uses his antlers to lift Kristoff off the ground. KRISTOFF (CONT'D) Ah! Stop it! Put me down! Sven drops him hard then "yells" at him once more. KRISTOFF (CONT'D) No, Sven! We're not going back! Sven shakes his head, angrily. KRISTOFF (CONT'D) She's with her true love. Sven makes an "of-course-she-isn't" face. Kristoff gets it; he's made his point. Just then the wind picks up. Kristoff looks back at the kingdom. Sees a violent winter storm swirling over the castle. Sharp ice claws its way up the castle, encasing it. KRISTOFF (CONT'D) Anna. Without hesitating, he dashes back down the mountain. Sven runs after him, catches up. Kristoff grabs Sven's harness and jumps onto his back. CUT TO: INT. LIBRARY -- NIGHT Anna shivers by the door. She looks up to see ice overtaking the ceiling. The door handle suddenly jiggles. Stops. Jiggles again. ANNA (barely a whisper) Help. CLICK. The door swings open. We see a carrot in the lock and hear a giggle of victory. Olaf takes the carrot, puts it back on his face. Then he sees Anna lying there. OLAF Anna. Oh no. He runs to the fireplace. Throws in some fresh wood, including one of his own arms, which he quickly rescues, before striking a match and relighting the fire. 100 FROZEN - J. Lee ANNA Olaf? Olaf. Get away from there. OLAF Whoa! So this is heat.... (considering) I love it. He reaches a twig finger toward the flames. It catches on fire. OLAF (CONT'D) Ooh! But don't touch it! He shakes the flame out, as he rushes over to help Anna to the fire. OLAF (CONT'D) So, where's Hans? What happened to your kiss? ANNA I was wrong about him. It wasn't true love. OLAF (confused innocence) Huh. But we ran all the way here? ANNA Please Olaf, you can't stay here; you'll melt. OLAF I am not leaving here until we find some other act of true love to save you. He sits down behind her, stubbornly. Leans his back against hers and thinks. OLAF (CONT'D) ...Do you happen to have any ideas? ANNA I don't even know what love is. OLAF (confident) That's okay, I do.... Olaf hops back up and puts a soothing hand on her shoulder. 101 FROZEN - J. Lee OLAF (CONT'D) Love is...putting someone else's needs before yours, like, you know, how Kristoff brought you back here to Hans and left you forever. ANNA ...Kristoff loves me? OLAF Wow, you really don't know anything about love, do you? His face starts to melt. ANNA Olaf, you're melting. OLAF (sweet and reassuring) Some people are worth melting for. But then...his face REALLY melts. He panics, pushes the snow back in place. OLAF (CONT'D) Just maybe not right this second. Suddenly, the window blows open, cold wind sweeps in. OLAF (CONT'D) Don't worry, I've got it! Olaf flitters to the window. He pulls one panel of it shut but struggles with the second panel. OLAF (CONT'D) (determined) We're going to get through-- (distracted) Oh, wait. Hang on. I'm getting something. He breaks an icicle off the window, uses it as a telescope and sees... Kristoff and Sven running back down the mountain. OLAF (CONT'D) It's Kristoff and Sven! They're coming back this way. ANNA ...They-they are? 102 FROZEN - J. Lee OLAF Wow, he's really moving fast. Huh.... I guess I was wrong. I guess Kristoff doesn't love you enough to leave you behind. Anna tries to get to her feet. ANNA Help me up, Olaf. Please. He hurries over, tumbling over the couch, knocking over the chess set and water jugs. OLAF No, no, no, no, no. You need to stay by the fire and keep warm. ANNA I need to get to Kristoff. OLAF (clueless) Why...? (realizing) Oh, oh, oh, I know why. He hops around in an excited display of hope. OLAF (CONT'D) There's your act of true love, right there, riding across the fjords like a valiant, pungent reindeer king! Come on! The walls crack under the ice pressure. OLAF (CONT'D) Look out! They rush out the room just as the ceiling collapses. INT. CASTLE HALLWAY -- DAY Anna and Olaf struggle down the hall. Ice spikes grow and block their path. OLAF We're trapped. Anna looks around desperately for a way out. 103 FROZEN - J. Lee EXT. FJORD -- DAY Elsa runs, but is nearly blinded by the snow and wind. EXT. CASTLE -- DAY Anna and Olaf bust open a window. The storm is so strong it sweeps the window panes away. OLAF Slide, Anna. It's a long, snowy way down. But what choice do they have? They slide down the iced-covered building. Anna arrives at the bottom, weak but uninjured. Olaf gathers snow along the way. He arrives at the bottom as a giant snowball. OLAF (CONT'D) We made it! He shakes off the extra snow as Anna struggles to her feet. EXT. FJORD -- DAY Kristoff and Sven bound off the mountain and sprint across the frozen fjord waters and right into the heart of the storm. Its white-out wind pushes them back. But they fight through. KRISTOFF Come on, buddy, faster. CUT TO: Anna and Olaf reach the shore of the fjords. ANNA Kristoff! The wind lifts Olaf up and pulls him apart. He goes swirling off into the storm. OLAF Keep going, Anna! Anna struggles on. 104 FROZEN - J. Lee ANNA Kristoff! PAN TO: Kristoff rides Sven past cracking, frozen ships. Sven struggles over the uneven surface. KRISTOFF Come on! Come on! Suddenly, a mangled ship, risen by ice, capsizes over them. They give it all they've got as debris falls all around them and the mast shatters. They make it past just as the entire ship slams down and cracks the thick ice beneath their feet. The ice opens up. Sven bravely jumps over a gap. But it's too wide. He bucks Kristoff to safety, but lands in the freezing water and disappears below. KRISTOFF (CONT'D) Sven? Sven! At first there's nothing but the wind and the tumbling icy water. But suddenly, Sven surfaces and claws his way to a floating ice chunk. He calls out, signalling for Kristoff to go on. KRISTOFF (CONT'D) Good boy. CUT TO: Anna moves blindly across the fjord. Anna's hands frost over an icy blue. She stumbles on, determined. But she's running out of time. She clutches her chest. The color in her eyes fades, the inevitable is coming. CUT TO: Kristoff, lost in the white-out, doesn't know which way to turn. But then he hears a faint-- ANNA (O.S.) Kristoff. KRISTOFF Anna...? Anna! WHITE OUT TO: 105 FROZEN - J. Lee Elsa struggles through her own storm, but the fear is consuming her. A dark shadow approaches. It's Hans. HANS Elsa. You can't run from this! Elsa backs away from him. ELSA ...Just take care of my sister. HANS Your sister? She returned from the mountain weak and cold. She said you froze her heart. ELSA What? No. HANS I tried to save her, but it was too late. Her skin was ice. Her hair turned white... Elsa's face sinks as she realizes what she has done. HANS (CONT'D) Your sister is dead... because of you. Elsa drops to her knees, emotionally broken. And with that, the swirling storm suddenly stops. The snow freezes mid-air, hangs suspended, trapped in grief. Citizens and dignitaries rush to the wall's edge and look out to see... Anna, barely able to move but now able to see across the fjords to... ANNA (a whisper) Kristoff. KRISTOFF Anna. Anna pushes on towards Kristoff. He runs top speed towards her. There's still a lot of fjord to cross, but Kristoff is giving it all he's got. He's going to make it. But then, Anna hears the sound of a sword being drawn from its scabbard. She turns and sees Hans, behind Elsa, as he raises his sword over his head. 106 FROZEN - J. Lee ANNA Elsa. Anna looks back at Kristoff as he runs for her. She gives him a longing look, but then turns away from him and then... Using all of her remaining strength, as Hans brings his sword down, Anna throws herself in front of Elsa. ANNA (CONT'D) No! In that instant, Anna freezes to solid ice. The sword hits her instead of Elsa. The sword shatters completely. The force of it sends Hans flying back and knocks him out. ELSA Anna! Elsa rushes to Anna and touches her sister's frozen face. ELSA (CONT'D) Oh, Anna...no...no, please no. Olaf walks up and sees Anna, frozen. OLAF (confused, sad) Anna? Elsa hugs Anna and cries. Kristoff watches in shocked despair. Sven steps up to his side. Citizens and dignitaries on the castle walls bow their heads. All of Arendelle is joined in somber silence. But then, Anna warms. She begins to thaw. Olaf looks up and gasps. Kristoff and Sven notice, light up. Anna bends her arm and embraces Elsa. ELSA Wha-? Anna? Anna opens her eyes. She smiles at Elsa, relieved. ANNA Oh, Elsa. They embrace. 107 FROZEN - J. Lee ELSA ...You sacrificed yourself for me? ANNA (weak) ...I love you. Olaf realizes what's happened. He's so excited about it, he lifts his head right off his body and exclaims-- OLAF An act of true love will thaw a frozen heart. ELSA (processing) Love...will thaw... (realizing) Love.... Of course. Elsa looks at Anna with confidence. ANNA Elsa? ELSA Love. Elsa lifts her arms, and the ground shakes and cracks. The ice and snow breaks away and rises high into the air. Beneath their feet the bow of a ship thaws. The entire fjord melts and other boats right themselves. The villagers come out to see the warmth returning. In one final wave, Elsa draws all of the snow into a giant snowflake in the sky, then waves it away, leaving only a warm summer day. ANNA I knew you could do it. OLAF (melting, good-naturedly) Hands down, this is the best day of my life...and quite possibly the last. ELSA Oh, Olaf. Hang on, little guy. 108 FROZEN - J. Lee Elsa waves her hand and surrounds Olaf with a swirl of cold air. He refreezes. Above his head she leaves a little, perpetually-snowing storm cloud. Olaf loves it. OLAF Hey, my own personal flurry. Kristoff sees Hans trying to get to his feet. He marches toward him, prepared for a fight. But Anna puts up a hand and stops him. ANNA Uh. Uh. Uh. She'll handle this. She goes over to Hans. HANS (confused) Anna? But she froze your heart. ANNA The only frozen heart around here is yours. She turns away from him, proud of her words. But not yet satisfied, she turns back and punches him right in the face. HANS Ah! Whoa, whoa, whoa! He falls overboard. Elsa comes over to Anna and hugs her. Over her shoulder, Kristoff meets Anna's eyes. She smiles brighter, happy. DISSOLVE TO: EXT. ARENDELLE -- DAY It's a beautiful summer day. The mighty ships have been repaired and are sailing away. On one of the ships, HANS is thrown into a brig. FRENCH DIGNITARY (to Kai) I will return this scoundrel to his country. We shall see what his twelve big brothers think of his behavior. KAI Arendelle thanks you, my Lord. 109 FROZEN - J. Lee Down on the dock, Arendelle guards lead the Duke and his two thugs to their ship. DUKE This is unacceptable. I am innocent. I'm a victim of fear. I've been traumatized. (bad acting) Ow! My neck hurts. Is there a doctor I could...No? And I demand to see the Queen! Kai steps down from the gangplank to the dock. KAI I have a message from the Queen. (reading a scroll) Arendelle will henceforth and forever no longer do business of any sort with Weaseltown. DUKE Weselton. It's Weselton! The guards usher him and his thugs onto their ship. EXT. VILLAGE SQUARE -- DAY Anna runs through the crowd, pulling a blindfolded Kristoff along behind her. She's so excited she can't stand it. ANNA Come on. Come on. Come on. Come on! She runs him right into a pole. KRISTOFF Pole. ANNA Oops. Sorry. EXT. ARENDELLE DOCKS -- DAY Anna skips to the perfect spot and stops. ANNA (stopping) Okay. Okay. Here we are. 110 FROZEN - J. Lee She takes off the blindfold. Kristoff opens his eyes. Before him sits the most beautiful, suped-up sled. Sven poses in front of it -- Vanna White-style. ANNA (CONT'D) I owe you a sled. KRISTOFF (blown away) Are you serious? ANNA Yes. And it's the latest model. KRISTOFF No. I can't accept this... ANNA You have to. No returns. No exchanges. Queen's orders. She's named you the official Arendelle Ice Master and Deliverer. Sven shows off the Ice-Master-and-Deliverer medal like he's king of the bucks. KRISTOFF What? That's not a thing. But he can't help but admire her enthusiasm. ANNA Sure it is. And it even has a cup holder.... Do you like it? KRISTOFF Like it? He sweeps her up high overhead and spins her around. KRISTOFF (CONT'D) I love it.... I could kiss you! He drops her, suddenly embarrassed. KRISTOFF (CONT'D) ...I could. I mean I'd like to. I'd... may I? We me....I mean, may we? Wait, what? She gives him a quick kiss on the cheek. ANNA We may. 111 FROZEN - J. Lee He smiles and goes for it. It's a true love's kiss, alright. We move past them to find Olaf enjoying the summer. With his snow cloud safely overhead, he's free to smell the flowers, which he does. Then sneezes his carrot nose off. Sven catches it between his teeth. Olaf gasps as Sven sucks the whole carrot into his mouth. It's gone. Olaf's face sinks in sadness. But not to fear, Sven spits the carrot back out and jams it into Olaf's face where it belongs. It's completely covered in reindeer spit, but Olaf doesn't seem to mind. He hugs Sven happily. CUT TO: EXT. CASTLE COURTYARD -- DAY The gates to the castle are wide open. In the courtyard, stands Elsa. ELSA Are you ready? Villagers cheer. Elsa stops and creates an ice rink. The people, skates at the ready, hope onto it and twirl about. Elsa then freezes the fountain in a beautiful design and adds some snow flurries for atmosphere. Anna comes slipping in. Elsa catches her. ANNA I like the open gates. ELSA We are never closing them again. Elsa then waves her hand and magical ice skates (literally made of ice) form on Anna's boots. ANNA What? Oh, Elsa, they're beautiful, but you know I don't ska-- Elsa grabs Anna's hands and pulls her along on the ice. Anna slips and slides, but laughs in delight. Sven goes slipping past. Kristoff runs after him. KRISTOFF Look out. Reindeer coming through! 112 FROZEN - J. Lee Olaf skates and helps Elsa coach Anna. OLAF That's it. Glide and pivot and glide and pivot. We pull away slowly, into the sky. We arrive at a bird's-eye view to see that where the castle had crumbled has been repaired with a ice. All is right in Arendelle. FINAL FADE OUT. THE END | 1 | 5.3% |
| ------------------------------------------------------------ FINDING NEMO Transcript v1.0 Copyright 2003 Walt Disney Pictures, Pixar Animation Studios ------------------------------------------------------------ Transcribed by BaD_BURN email : markgonzalez154@hotmail.com ------------------------------------------------------------------ | Okay, this is the work-in-progress FINDING NEMO film transcript. | | Why is it 'work-in-progress' you might ask? Well for one, this | | isn't a 100% accurate transcript: some words might be missing, | | may not be right. Second, some lines may or may not have been | | spoken by the right character. There are instances in the film | | where a line is spoken but the character isn't on screen, which | | makes things complicated. But I'd say this transcript is about | | 98-99% accurate. Dialogue for each scene is seperated by a line | | of equal signs (=). | | | | This transcript is open for corrections, additions if you have | | any. What you CAN'T do, however, is to edit it and take credit | | for it. Although I do not own the movie or it's screenplay, this | | transcript was made with no intention of copyright infringement | | and the like. Enjoy. And remember: 'Fish are friends, not food'. | ------------------------------------------------------------------ ====================================================================================== MARLIN Wow. CORAL Mmm. MARLIN Wow. CORAL Mmm-hmm. MARLIN Wow. CORAL Yes, Marlin. No, I see it. It's beautiful. MARLIN So, Coral, when you said you wanted an ocean view, you didn't think that we we're gonna get the whole ocean, did you? Huh? [sighs] Oh yeah. A fish can breath out here. Did your man deliver or did he deliver? 1 CORAL My man delivered. MARLIN And it wasn't so easy. CORAL Because a lot of other clownfish had their eyes on this place. MARLIN You better believe they did--every single one of them. CORAL Mm-hmm. You did good. And the neighborhood is awesome. MARLIN So, you do like it, don't you? CORAL No, no. I do, I do. I really do like it. But Marlin, I know that the drop off is desirable with the great schools and the amazing view and all, but do we really need so much space? MARLIN Coral, honey, these are our kids we're talking about. They deserve the best. Look, look, look. They'll wake up, poke their little heads out and they'll see a whale! See, right by their bedroom window. CORAL Shhh, you're gonna wake the kids. MARLIN Oh, right. Right. CORAL Aww, look. They're dreaming. We still have to name them. MARLIN You wanna name all of 'em, right now? All right, we'll name this half Marlin Jr. and then this half Coral Jr. Okay, we're done. CORAL I like Nemo. MARLIN Nemo? Well, we'll name one Nemo but I'd like most of them to be Marlin Jr. CORAL Just think that in a couple of days, we're gonna be parents! MARLIN Yeah. What if they don't like me? CORAL Marlin. MARLIN No, really. CORAL There's over 400 eggs. Odds are, one of them is bound to like you. CORAL What? MARLIN You remember how we met? CORAL Well, I try not to. MARLIN Well, I remember. 'Excuse me, miss, can you check and see if there's a hook in my lip?' CORAL Marlin! MARLIN 2 'Well, you gotta look a little closer because it's wiggling'. CORAL Get away! MARLIN Here he is. Cutie's here! Where did everybody go? MARLIN [gasps] Coral, get inside the house, Coral. No, Coral, don't. They'll be fine. Just get inside, you, right now. MARLIN No! MARLIN Coral! Coral? MARLIN Coral? Oh! MARLIN Ohh. There, there, there. It's okay, daddy's here. Daddy's got you. I promise, I will never let anything happen to you...Nemo. ====================================================================================== NEMO First day of school! First day of school! Wake up, wake up! C'mon, first day of school! MARLIN I don't wanna go to school. Five more minutes. NEMO Not you, dad. Me! MARLIN Okay...huh? NEMO Get up, get up! It's time for school! It's time for school! It's time for school! It's time for school! Oh boy! Oh boy! MARLIN All right, I'm up. NEMO Oh boy--whoa! MARLIN Nemo! NEMO First day of school! MARLIN [gasps] Nemo, don't move! Don't move! You'll never get out of there yourself. I'll do it. All right, where's the break? You feel a break? NEMO No. MARLIN Sometimes you can't tell 'cause fluid is rushing to the area. Now, any rushing fluids? NEMO No. MARLIN Are you woozy? NEMO No. MARLIN How many stripes do I have? 3 NEMO I'm fine. MARLIN Answer the stripe question! NEMO Three. MARLIN No! See, something's wrong with you. I have one, two, three--that's all I have? Oh, you're okay. How's the lucky fin? NEMO Lucky. MARLIN Let's see. MARLIN Are you sure you wanna go to school this year? 'Cause there's no problem if you don't. You can wait 5 or 6 years. NEMO Come on, dad. It's time for school. MARLIN Ah-ah-ah! Forgot to brush. NEMO Ohh... MARLIN Do you want this anemone to sting you? NEMO Yes. MARLIN Brush. NEMO Okay, I'm done. MARLIN You missed a spot. NEMO Where? MARLIN There. Ha ha! Right there. And here and here and here! ====================================================================================== MARLIN All right, we're excited. First day of school, here we go. We're ready to learn to get some knowledge. Now, what's the one thing we have to remember about the ocean? NEMO It's not safe. MARLIN That's my boy. So, first we check to see that the coast is clear. We go out and back in. And then we go out, and back in. And then one more time--out and back in. And sometimes, if you wanna do it four times-- NEMO Dad.. MARLIN All right. Come on, boy. NEMO Dad, maybe while I'm at school, I'll see a shark! MARLIN 4 I highly doubt that. NEMO Have you ever met a shark? MARLIN No, and I don't plan to. NEMO How old are sea turtles? MARLIN Sea turtles? I don't know. NEMO Sandy Plankton from next door, he said that sea turtles, said that they live to be about a hundred years old! MARLIN Well, you know what, if I ever meet a sea turtle, I'll ask him. After I'm done talking to the shark, okay? Whoa, whoa, whoa! Hold on, hold on, wait to cross. Hold my fin, hold my fin. NEMO Dad, you're not gonna freak out like you did at the petting zoo, are you? MARLIN Hey, that snail was about to charge. Hmm, I wonder where we're supposed to go. FISH KIDS Bye, mom! FISH MOM I'll pick you up after school. CRAB KID Come on, you guys. Stop it! Give it back! MARLIN Come on, we'll try over there. MARLIN Excuse me, is this where we meet his teacher? BOB Well, look who's out of the anemone. MARLIN Yes. Shocking, I know. BOB Marty, right? MARLIN Marlin. BOB Bob. TED Ted. BILL Bill. Hey, you're a clownfish. You're funny, right? Hey, tell us a joke. BOB/TED Yeah, yeah. Come on, give us a funny one. MARLIN Well, actually, that's a common misconception. Clownfish are no funnier than any other fish. BILL Aw, come on, clownie. TED Yeah, do something funny. 5 BOB Yeah! MARLIN All right, I know one joke. Um, there's a mollusk, see? And he walks up to a sea, well he doesn't walk up, he swims up. Well, actually the mollusk isn't moving. He's in one place and then the sea cucumber, well they--I mixed up. There was a mollusk and a sea cucumber. None of them were walking, so forget that I-- BOB Sheldon! Get out of Mr. Johansenn's yard, now! KIDS Whoa! MR. JOHANSSEN All right, you kids! Ooh! Uuh, where'd you go? Where'd you go? Where, where'd you go? NEMO Dad, dad...can I go play too? Can I? MARLIN I would feel better if you go play over on the sponge beds. MARLIN That's where I would play PEARL What's wrong with his fin? TAD He looks funny! SHELDON Ow! Hey, what'd I do? What'd I do? BOB Be nice. It's his first time at school. MARLIN He was born with it, kids. We call it his lucky fin. NEMO Dad. PEARL See this tentacle? It's actually shorter than all my other tentacles but you can't really tell.Especially when I twirl them like this. SHELDON I'm H2O-intolerant. [sneezes] TAD I'm obnoxious. MR. RAY [singing] Oooh, let's name the zones, the zones, the zones. Let's name the zones of the open sea. KIDS Mr. Ray! SHELDON Come on, Nemo. MARLIN Whoa, you better stay with me. MR. RAY [singing]..mesopolagic, bathyal, abyssalpelagic. All the rest are too deep for you and me to see. MR. RAY Huh, I wonder where my class has gone? KIDS 6 We're under here! MR. RAY Oh, there you are. Climb aboard, explorers. [singing] Oh, knowledge exploring is oh so lyrical, when you think thoughts that are empirical. NEMO Dad, you can go now. MR. RAY Well, hello. Who is this? NEMO I'm Nemo. MR. RAY Well, Nemo, all new explorers must answer a science question. NEMO Okay. MR. RAY You live in what kind of home? NEMO An anemo-none. A nemenem-menome-nememen-nenemone-- MR. RAY Okay, okay, don't hurt yourself. Welcome aboard, explorers! MARLIN Just so you know, he's got a little fin. I find if he's having trouble swimming, let him take a break. Ten, fifteen minutes. NEMO Dad, it's time for you to go now. MR. RAY Don't worry. We're gonna stay together as a group. Okay, class, optical orbits up front. And remember, we keep our supraesophogeal ganglion to ourselves...that means you, Jimmy. JIMMY Aw, man! MR. RAY [singing] MARLIN Bye, Nemo! NEMO Bye, dad! MARLIN Bye, son! Be safe. BOB Hey, you're doing pretty well for a first timer. MARLIN Well, you can't hold onto them forever, can you? BILL Yeah, I had a tough time when my oldest went out at the drop off. MARLIN They just gotta grow up--the drop off?! They're going to the drop off?! Wh-what are you, insane?! Why don't we fry 'em up now and serve them with chips!? BOB Hey, Marty. Calm down. MARLIN Don't tell me to be calm, pony boy! BOB 'Pony boy'? 7 BILL You know for a clownfish, he really isn't that funny. TED Pity. ====================================================================================== MR. RAY [singing] Oh, let's name the species, the species, the species. Let's name the species that live in thesea. NEMO Whoa. MR. RAY [singing] There's porifera, coelenterata, hydrozoa, scyphozoa, anthozoa, ctenophora, bryozoas, three! Gastropoda, arthropoda, echinoderma, and some fish like you and me. Come on, sing with me. Oh...! MR. RAY Just the girls this time. [singing] Oh, seaweed is cool. Seaweed is fun. It makes it's food with the rays of the sun... MR. RAY Okay, the drop off. All right, kids, feel free to explore but stay close. [gasps] Stromalitic cyanobacteria! Gather. An entire ecosystem contained in one infinitesimal speck. There are as many protein pairs contained in this... TAD Come on, let's go. MR. RAY Come on, sing with me! [singing] There's porifera, coelentera, hydrozoa, scyphozoa, anthozoa, ctenophora, bryozoas, three! NEMO Hey guys, wait up! Whoa. TAD Cool. TAD Saved your life! PEARL Aw, you guys made me ink. NEMO What's that? TAD I know what that is. Oh, oh! Sandy Plankton saw one. He called, he said it was called a...a butt. NEMO Whoa. PEARL Wow. That's a pretty big butt. SHELDON Oh, look at me. I'm gonna go touch the butt. [sneezes] Whoa! SHELDON Oh yeah? Let's see you get closer. PEARL Okay. Beat that. TAD Come on, Nemo. How far can you go? NEMO Uh, my dad says it's not safe. 8 MARLIN Nemo, no! NEMO Dad? MARLIN You were about to swim into open water! NEMO No, I wasn't go out--but dad! MARLIN It was a good thing I was here. If I hadn't showed up, I don't know-- PEARL Sir, he wasn't gonna go. TAD Yeah, he was too afraid. NEMO No, I wasn't. MARLIN This does not concern you, kids. And you're lucky I don't tell your parents you were out there. You know you can't swim well. NEMO I can swim fine, dad, okay? MARLIN No, it's not okay. You shouldn't be anywhere near here. Okay, I was right. You'll start school in a year or two. NEMO No, dad! Just because you're scared of the ocean-- MARLIN Clearly, you're not ready. And you're not coming back until you are. You think you can do these things but you just can't, Nemo! NEMO I hate you. MR. RAY There's--nothing to see. Gather, uh, over there. Excuse me, is there anything I can do? I am a scientist, sir. Is there any problem? MARLIN I'm sorry. I didn't mean to interrupt things. He isn't a good swimmer and it's a little too soon for him to be out here unsupervised. MR. RAY Well, I can assure you, he's quite safe with me. MARLINLook, I'm sure he is. But you have a large class and he can get lost from sight if you're not looking. I'm not saying you're not looking-- FISH KID Oh my gosh! Nemo's swimming out to sea! MARLIN Nemo! What do you think you're doing? You're gonna get stuck out there and I'll have to get you before another fish does! Get back here! I said get back here, now! Stop! You take one move, mister. Don't youdare! If you put one fin on that boat..are you listening to me? Don't touch the bo--Nemo! TAD [whispering] He touched the butt. MARLIN You paddle your little tail back here, Nemo. That's right. You are in big trouble, young man. Do you hear me? Big...big-- 9 NEMO Aaaah! Daddy! Help me! MARLIN I'm coming, Nemo! KIDS Aaaah! MR. RAY Get under me, kids! NEMO Ah! Oh no! Dad! Daddy! MARLIN Oh! Nemo! Unh! Nemo! Nemo, no! Nemo! Nemo! Nemo! No! No! Aah! Nemo! Nemo! DIVER Whoa! Hold on. MARLIN Oh no. No, no. It's gone, it's gone. No, no, it can't be gone. No, no! Nemo! Nemo! Nemo! No! Nemo! Nemo! No! No, please, no! No, no! MARLIN Has anybody seen a boat!? Please! A white boat! They took my son! My son! Help me, please! DORY Look out! MARLIN Waaaah! MARLIN Ooh, ooh... DORY Ohh. Oh, oh. Sorry! I didn't see you. Sir, are you okay? MARLIN He's gone, he's gone.. DORY There, there. It's all right. MARLIN He's gone. DORY It'll be okay. MARLIN No, no. They took him away. I have to find the boat. DORY Hey, I've seen a boat. MARLIN You have? DORY It passed by not too long ago. MARLIN A white one? DORY Hi. I'm Dory. MARLIN Where!? Which way!? DORY Oh, oh, oh! It-it went, um, this way! And it went this way! Follow me! MARLIN 10 Thank you! Thank you, thank you so much! DORY No problem. MARLIN Hey! Wait! DORY Will you quit it? MARLIN What? DORY I'm trying to swim here. What, ocean ain't big enough for you? MARLIN Huh? DORY You got a problem, buddy? Huh? Huh? Do 'ya? Do 'ya? Do 'ya? You want a piece of me? Yeah, oooh, I'm scared now. Whaat!? MARLIN Wait a minute.. DORY Stop following me, okay!? MARLIN What? You're showing me which way the boat went! DORY A boat? Hey, I've seen a boat. It passed by not too long ago. It went this way, it went this way. Follow me! MARLIN Wait a minute, wait a minute! What is going on? You already told me which way the boat was going! DORY I did? Oh dear... MARLIN If this is some kind of practical joke, it's not funny! And I know funny..I'm a clownfish! DORY No, it's not. I know it's not. I'm so sorry. See, I suffer from short-term memory loss. MARLIN Short-term memory loss..I don't believe this! DORY No, it's true. I forget things almost instantly. It runs in my family..or at least I think it does. Hmmm..where are they? Can I help you? MARLIN Something's wrong with you, really. You're wasting my time. I have to find my son. [gasps] BRUCE Hello. DORY Well, hi! BRUCE Name's Bruce. It's all right, I understand. Why trust a shark, right? So, what's a couple of bites like you doing out so late, eh? MARLIN Nothing. We're not doing anything. We're not even out. BRUCE Great! Then how'd you morsels like to come to a little get-together I'm havin'? DORY 11 You mean like a party? BRUCE Yeah, yeah, that's right--a party! What do you say? DORY Ooh, I love parties! Parties are fun! MARLIN Parties are fun, and it's tempting but-- BRUCE Oh, come on, I insist. MARLIN O-okay..that's all that matters. DORY Hey, look--balloons! It is a party! BRUCE Ha ha ha! Mind your distance, though. Those balloons can be a bit dodgy. You wouldn't want one of them to pop. BRUCE Anchor! Chum! ANCHOR There you are, Bruce, finally! BRUCE We got company. ANCHOR It's about time, mate. CHUM We've already gone through all the snacks and I'm still starvin'! ANCHOR We almost had a feeding frenzy. CHUM Come on, let's get this over with. ====================================================================================== BRUCE Right, then. The meeting has officially come to order. Let us all say the pledge.. BRUCE/ANCHOR/CHUM 'I am a nice shark, not a mindless eating machine. If I am to change this image, I must first change myself. Fish are friends, not food'. ANCHOR Except stinkin' dolphins. CHUM Dolphins! Yeah, they think they're sooo cute! 'Hey, look at me. I'm a flippin' little dolphin! Let me flip for 'ya! Ain't I a somethin'!' BRUCE Right, then. Today's meeting is step 5, 'BRING A FISH FRIEND'. Now do you all have your friends? ANCHOR Got mine. DORY Hey there! BRUCE How 'bout you, Chum? CHUM Oh, um, I seem to have misplaced my uh, friend. 12 BRUCE That's all right, Chum. I had a feeling this would be a difficult step, you can help yourself to one of my friends. CHUM Oh, thanks, mate. A little chum for Chum, eh? BRUCE I'll start the testimonies. Hello, my name is Bruce. ANCHOR/CHUM Hello, Bruce. BRUCE It has been three weeks since my last fish, on my honor, or may I be chopped up and made into soup. CHUM You're an inspiration to all of us. ANCHOR Amen. BRUCE Right, then. Who's next? DORY Ooh! Pick me! Pick me! BRUCE Yes, the little Sheila down the front. DORY Woo-hoo! BRUCE Come on up here. DORY Hi. I'm Dory. BRUCE/ANCHOR/CHUM Hello, Dory. DORY And, uh, well, I don't think I've ever eaten a fish. CHUM Hey, that's incredible. BRUCE Good on 'ya, mate! DORY Whew! I'm glad I got that off my chest. BRUCE All right, anyone else? Hello, how 'bout you, mate? What's your problem? MARLIN Me? I don't have a problem. BRUCE Oh. Okay.. BRUCE/ANCHOR/CHUM Denial. BRUCE Just start with your name. MARLIN Okay. Uh, hello. My name is Marlin. I'm a clownfish-- CHUM A clownfish? Really?! 13 BRUCE Go on, tell us a joke! CHUM Ooh! I love jokes! MARLIN Actually I do know one that's pretty good. There was this mollusk and he walks up to a sea cucumber. Normally, they don't talk, sea cucumbers, but in a joke, everyone talks. So the sea mollusk says to the cucumber... NEMO Daddy! MARLIN Nemo! CHUM Nemo! Ha ha ha! Nemo! I don't get it. BRUCE For a clownfish, he's not that funny. MARLIN No, no, no, no. He's my son. He was taken by these divers. DORY Oh my, you poor fish. CHUM Humans. Think they own everything. ANCHOR Probably American. BRUCE Now there is a father looking for his little boy. MARLIN Ugh! What do these markings mean? BRUCE I never knew my father! [sobs] CHUM Aw, come here. ANCHOR Group hug. CHUM We're all mates here, mate. MARLIN I can't read human. DORY Well then we gotta find a fish who can read this. Hey, look. Sharks! MARLIN No, no, no, Dory! DORY Guys, guys! MARLIN No, Dory! DORY That's mine! Give it to me! Gimme! Oww! MARLIN Oh, I'm sorry. Are you okay? DORY Ow, ow, ow. 14 MARLIN I'm so sorry. DORY You really clocked me there. Am I bleeding? MARLIN Ohh... DORY Ow, ow, ow. BRUCE Dory, are you oka--oohh. Oohh, that's good. ANCHOR/CHUM Intervention! BRUCE Just a bite! ANCHOR Hold it together, mate! CHUM Remember, Bruce, fish are friends, not food! BRUCE FOOD! MARLIN Dory, look out! BRUCE I'm havin' fish tonight! CHUM Remember the steps, mate! BRUCE Just one bite! BRUCE G'day! MARLIN/DORY Aaaaaaaah! BRUCE Arrrr! MARLIN There's no way out! There's got to be a way to escape! DORY Who is it? MARLIN Dory, help me find a way out! DORY Sorry, you'll have to come back later. We're trying to escape. MARLIN There's gotta be a way out! DORY Look, here's something! 'ESSS-CA-PE'! I wonder what that means. It's funny, it's spelled just like the word 'escape'. MARLIN Let's go! BRUCE Here's Brucey! MARLIN 15 Wait a minute..you can read?! DORY I can read? That's right, I can read! MARLIN Well, then here. Read this now! ANCHOR He really doesn't mean it, y'know! He never even knew his father! CHUM Don't fall off the wagon! MARLIN Oh no, it's blocked! ANCHOR No, Bruce. Focus! CHUM Sorry about--this, mate! ANCHOR He's really--a nice guy! MARLIN I need to get that mask! DORY You want that mask? Okay. MARLIN No, no, no, no, no, no! MARLIN Quick grab the mask! ANCHOR Oh no. Bruce? BRUCE What? [gasps] Swim away! Swim away! DORY Aw, is the party over? PELICAN Nice. ====================================================================================== NEMO Dad? Daddy? DENTIST Barbara? BARBARA Uh-huh? DENTIST Prep for his anterior crown, would you, please? And I'm going to need a few cotton rolls. BARBARA Okay. DENTIST Hello, little fella! NEMO Aah! DENTIST Heh heh heh! Beauty, isn't he? I found that guy struggling for life out on the reef and I saved him. So, has that novocaine kicked in yet? 16 PATIENT I think so. We're ready to roll. BUBBLES Bubbles! [muttering] My bubbles. PEACH He likes bubbles. NEMO Aah! Ohh! No! Uhh! JACQUES Bonjour. NEMO Aah! BLOAT Heh heh! Slow down, little fella. There's nothing to worry about. DEB Oh, he's scared to death. NEMO I wanna go home. Do you know where my dad is? PEACH Honey, your dad's probably back at the pet store. NEMO Pet store? BLOAT Yeah, you know, like I'm from Bob's Fish Mart. GURGLE Pet Palace. BUBBLES Fish-O-Rama. DEB Mail order. PEACH Ebay. GURGLE So which one is it? NEMO I'm from the ocean. GURGLE Ah, the ocean. The ocean! Aaah! He hasn't been decontaminated yet! Jacques! JACQUES Oui. GURGLE Clean him! JACQUES Oui. GURGLE Ocean! JACQUES Ooh, la mer. Bon. Voila. He is clean. BUBBLES Wow. The big blue. What's it like? NEMO Big...and blue? 17 BUBBLES I knew it. DEB Kid, if there's anything you need, just ask your auntie Deb, that's me. Or if I'm not around, you can always talk to my sister Flo. Hi,how are you? Don't listen to anything my sister says, she's nuts! Ha ha ha ha! PEACH [muffled] We got a live one! BLOAT Can't hear you, Peach. PEACH I said we got a live one. GURGLE Yes! BLOAT Oh boy, oh boy, oh boy, oh boy! DEB What do we got? PEACH Root canal, and by the looks of those x-rays it's not gonna be pretty. PATIENT Owwwwwwwww! BLOAT Rubber dam and clamp installed? PEACH Yep. GURGLE What did he use to open? PEACH Gator-Glidden drill. He seems to be favoring that one lately. DEB I can't see, Flo. PATIENT You're getting a little too--aaaaah!!! PEACH Now he's doing the Schilder technique. BLOAT Oooh, he's using a Hedstrom file. GURGLE That's not a Hedstrom file. That's a K-Flex. BLOAT It's got a teardrop cross-section. Clearly a Hedstrom. GURGLE No, no. K-Flex. BLOAT Hedstrom! GURGLE K-Flex! BLOAT Hedstro--! [inflates] There I go. A little help over here. DEB I'll go deflate him. 18 DENTIST All right, go ahead and rinse. GURGLE Ugh! The human mouth is a disgusting place. PEACH Hey, Nigel. NIGEL What did I miss? Am I late? PEACH Root canal and it's a doozy. NIGEL Root canal, eh? What did he use to open? PEACH Gator-Glidden drill. NIGEL He seems to be favoring that one. Hope he doesn't get surplus sealer at the portal terminus... hello. NEMO [gasps] NIGEL Who's this? DEB New guy. Ha ha ha! GURGLE The dentist took him off the reef. NIGEL An outie. From my neck of the woods, eh? Sorry if I ever took a snap at you. Fish gotta swim, birds gotta eat. [gasps] DENTIST Hey! No, no, no, no! They're not your fish. They're my fish. Come on, go! Go on, shoo! Oh, the picture broke. This here's Darla. She's my niece. She's going to be eight next week. Hey, little fella. Say hello to your new mummy. She'll be here Friday to pick you up. You're her present. Shh, shh, shh! It's our little secret. Well, Mr. Tucker, while that sets up I'm going to see a man about a wallaby. BLOAT Oh, Darla. NEMO What? What's wrong with her? GURGLE She wouldn't stop shaking the bag. BUBBLES Poor Chuckles. DEB He was her present last year. BLOAT Hitched a ride on the porcelain express. PEACH She's a fish killer. NEMO I can't go with that girl! I have to get back to my dad! Aaah! Daddy! Help me! GURGLE Oh, he's stuck! GILL 19 Nobody touch him! Nobody touch him. NEMO Can you help me? GILL No. You got yourself in there, you can get yourself out. PEACH Gill.. GILL I just wanna see him do it, okay? Calm down. Alternate wiggling your fins and your tail. NEMO I can't. I have a bad fin. GILL Never stopped me. GILL Just think about what you need to do. BLOAT Come on. GILL Perfect. BUBBLES Yay! GURGLE You did it! DEB Good squirming! Ha ha ha! PEACH Wow. From the ocean. Just like you, Gill. GILL Yeah. PEACH I've seen that look before. What are you thinking about? GILL I'm thinking, tonight, we give the kid a proper reception. BLOAT So kid, you got a name or what? NEMO Nemo. I'm Nemo. ====================================================================================== MARLIN Nemo. Nemo. [mutters] DORY Are you gonna eat that? Careful with that hammer... MARLIN Huh? No, no! What does it say? Dory! DORY Sea monkey has my money... MARLIN Wake up! Get up! Come on! Come on! DORY Yes, I'm a natural blue... MARLIN 20 Get up! DORY Look out! Sharks eat fish! Aaaaaah! MARLIN/DORY AAAAAAAAAAHHH!!! DORY Wow. Dusty. MARLIN [gasps] The mask! Where's the mask? No! No, not the mask! Get it! Get the mask! Get the mask! Get it! DORY [singing] Hoo doot doo doot doot doo doot. Whoo-hoo! La la la la la la. Just keeps going on, doesn't it? Echo! Echo! Hey, what are you doing? MARLIN It's gone. I've lost the mask. DORY Did you drop it? MARLIN You dropped it! That was my only chance of finding my son, now it's gone. DORY Hey, Mr. Grumpy Gills. When life gets you down, you know what you gotta do? MARLIN I don't wanna know what you gotta do when life gets you down. DORY [singing] Just keep swimming. Just keep swimming, swimming, swimming. What do we do? We swim, swim. MARLIN Dory, no singing. DORY [singing] Ho ho ho ho ho ho! I love to swim! When you want to swim.. MARLIN See, I'm going to get stuck now with that song now it's in my head! DORY Sorry. MARLIN Dory, do you see anything? DORY Aaah! Something's got me! MARLIN That was me. I'm sorry. DORY [gasps] Who was that? MARLIN Who could it be? It's me! DORY Are..are you my conscience? MARLIN Yeah, yeah. I'm your conscience. We haven't spoken for a while. How are you? DORY Hmm, can't complain. MARLIN Yeah? Good. Now, Dory. I want you to tell me..do you see anything? 21 DORY I see..I see a light. MARLIN A light. DORY Yeah. Over there. Hey, conscience. Am I dead? MARLIN No, I see it too. What is it? DORY It's so pretty. MARLIN I'm feeling...happy. Which is a big deal for me. DORY I want to touch it. Oh! MARLIN Hey, come back. Come on back here. DORY [singing] I'm gonna get you. I'm gonna get you. I'm gonna swim with you. MARLIN I'm gonna get you. I'm gonna be your best friend...good feeling's gone. MARLIN I can't see! I don't know where I'm going! DORY Haah! MARLIN The mask! DORY What mask? DORY Okay, I can't see a thing. MARLIN Oh, gee! DORY Hey, look! A mask! MARLIN Read it! DORY I'm sorry, but if you could just bring it a little closer, I kind of need the light. That's great, keep it right there. MARLIN Just read it! DORY Okay, okay. Mr. Bossy. Uh, 'P'. Okay, 'P'. 'Shh-eer...Sher--P. Sher--P. Shirley? P.--'. Oh! The first line's 'P. Sherman'! MARLIN P. Sherman doesn't make any sense! DORY Okay, second line. '42'. MARLIN Don't eat me! Don't eat me! Aaaah! DORY Light, please. 'Walla--Walla--Walla-beee'... 22 MARLIN Waah! Waaah! Waaaah! DORY The second line's '42 Wallaby Way'! MARLIN That's great! Speed read! Take a guess! No pressure! No problem! There's a lot of pressure! Pressure! Take a guess now with pressure! DORY 'Sydney'. It's 'Sydney'! MARLIN Duck! DORY Aaah! MARLIN I'm dead, I'm dead, I'm dead, I'm dead, I'm dead, I died, I'm dead. MARLIN Whoo-hoo! [singing] We did it, we did it! Oh yeah, yeah, yeah! No eating here tonight, whoo! BOTH [singing] Eating here tonight! MARLIN Dory. DORY [singing] No, no, no eating here tonight. You on a diet-- MARLIN Dory! What did the mask say? DORY 'P. Sherman, 42 Wallaby Way, Sydney'. [gasps] I remember what it said! I usually forget things, but I remembered it this time! MARLIN Whoa, whoa, wait! Where is that? DORY I don't know. But who cares? I remembered! MARLIN/DORY Aaah! DORY P. Sherman, 42 Wallaby Way, Sydney. I remembered it again! ====================================================================================== JACQUES Psst. Nemo. NEMO Mmmm... JACQUES Nemo. NEMO Huh? JACQUES Suivez-moi. Follow me. BLOAT/BUBBLES/GURGLE [chanting] Ha! Ho! Hwa! Hwee! Ha! Ho! Ho! Ho! Ha! Ho! Hwa! Hwee! Ha! Ho! Ho! Ho! Ha! Ho! Hwa! Hwee! Ha! Ho! Ho! Ho! Hahoo! Wahoo! Yahoo! Ho! Ha! Ho! Wahee! Ha! Ho! Ho! Ho! Hoo! GILL State your name. 23 NEMO Nemo. GILL Brother Bloat, proceed. BLOAT Nemo! Newcomer of orange and white, you have been called forth to the summit of Mount Wannahockaloogie to join with us in the fraternal bonds of tankhood. NEMO Huh? PEACH We want you in our club, kid. NEMO Really? BLOAT If you are able to swim through..THE RING OF FIRE! [whispers to Jacques] Turn on the Ring of Fire! The Ring of Fire, you said you could do it--THE RING OF FIRE! BUBBLES Bubbles! Bubbles! Let me--oww! BLOAT/BUBBLES/GURGLE [chanting] PEACH Isn't there another way? He's just a boy! JACQUES [wailing] GILL From this moment on, you will now be known as Sharkbait. BLOAT/BUBBLES/GURGLE Sharkbait! Ooh ha ha! GILL Welcome, brother Sharkbait! BLOAT/BUBBLES/GURGLE Sharkbait! Ooh ha ha! GILL Enough with the Sharkbait. GURGLE Sharkbait! Ooh..ba-ba-doo. GILL Okay, Sharkbait's one of us now, agreed? BLOAT/BUBBLES/GURGLE Agreed! GILL We can't send him off to his death. Darla's coming in 5 days, so what are we gonna do? I'll tell you what we're gonna do: we're gonna get him outta here. We're gonna help him escape. NEMO Escape? Really? GILL We're all gonna escape! GURGLE Gill, please, not another one of your escape plans. DEB Sorry, but they, they just, they never work. 24 BLOAT Yeah. Why should this be any different? GILL 'Cause we've got him. NEMO Me? GILL You see that filter? NEMO Yeah? GILL You're the only one who can get in and out of that thing. What we need you to do is take a pebble inside and jam the gears. You do that and this tank's gonna get filthier and filthier by the minute. Pretty soon, the dentist'll have to clean the tank himself. And when he does, he'll take us out of the tank, put us in the individual baggies, then we roll ourselves down the counter, out of the window, off the awning, into the bushes, across the street and into the harbor! It's foolproof! Who's with me? BLOAT Aye! JACQUES Aye! DEB Aye! BUBBLES Aye! GURGLE I think your nuts. GILL/NEMO [sighs] GURGLE No offense, kid, but, um..you're not the best swimmer. GILL He's fine, he can do this. So Sharkbait, what do you think? NEMO Let's do it. ====================================================================================== DORY I'm going to P. Sherman, 42 Wallaby Way, Sydney. Where are you going? I'm going to P. Sherman, 42 Wallaby Way, Sydney. If you're askin' where I'm goin'. I'll tell you that's where I'm going. It's P. Sherman, 42 Wallaby Way, Sydney. Where? I'm sorry, I didn't hear you. P. Sherman, 42 Wallaby Way... MARLIN Excuse me. Ex-excuse me, um, hi. Do you know how to get to--hello? W-w-w-wait! Can you tell me--hey! Hold it! Wait a minute! I'm trying to talk to you. Okay, fellas, come back here. Please, one quick question. I need to aaaaand they're gone again. [sighs] DORY P. Sherman 42 Wallaby Way, Sydney. Why do I have to tell you over and over again? I'll tell you again. I don't get tired of it-- MARLIN Okay, all right. DORY Huh? MARLIN Here's the thing. DORY 25 Uh-huh. MARLIN Y'know, I just, I-I think it's best if I just, if I just, carry on from here by..by myself. DORY Okay. MARLIN Y'know, alone. DORY Uh-huh. MARLIN Without, without..well, I mean, not without you. I mean, it's just that I don't want you... with me. DORY Huh? MARLIN I don't wanna hurt your feelings.. DORY You want me to leave? MARLIN Well, I mean not..yes, yeah. It's just that you know I-I just can't afford anymore delays and you're one of those fish that cause delays. And sometimes it's a good thing. There's a whole group of fish. They're..'delay fish'. DORY You mean..[whimper]you mean you don't..like me? [sobs] MARLIN No, of course I like you. It's because I like you I don't wanna be with you. It's a complicated emotion. Oh, don't cry. I like you. MOONFISH LEADER Hey, you! Lady, is this guy botherin' you? DORY Um, I don't remember. Were you? MARLIN No, no, no, no, no. We're just, we're..hey, do you guys know how I can get to-- MOONFISH LEADER Look, pal. We're talkin' to the lady, not you. Hey-hey, you like impressions? DORY Mm-mmm-mmmm. MOONFISH LEADER Okay. Just like in rehearsals, gentlemen. So, what are we? Take a guess. DORY Oh, oh, I've seen one of those. MOONFISH LEADER I'm a fish with a nose like a sword. DORY Wait, wait, um.. MARLIN It's a swordfish. MOONFISH LEADER Hey, clown boy! Let the lady guess. Where's the butter? DORY Oh-oh-oh! It's on the tip of my tongue. MARLIN [coughs up answer]Lobster. 26 MOONFISH LEADER Saw that. MARLIN What? MOONFISH LEADER Lots of legs, lives in the ocean. DORY Clam! MOONFISH LEADER Close enough. [singing] Oh, it's a whale of a tale, I'll tell you lad, a whale of a tale. DORY Oh, they're good. MARLIN Will somebody please give me directions? MOONFISH LEADER [impersonating Marlin] Will somebody please give me directions? DORY Ha ha ha ha ha! MARLIN I'm serious. MOONFISH LEADER Blah-blah-blah! Me-me-blah! Blah-blah-blah-blah-me-me-me! MARLIN Thank you. DORY Oh dear. Hey, hey come back! Hey, what's the matter? MARLIN What's the matter? While they're doing their silly little impressions, I am miles from home, with a fish that can't even remember her own name. DORY Boy, bet that's frustrating. MARLIN Yeah. Meanwhile my son is out there. DORY You're son Chico? MARLIN Nemo. DORY Right. Got it. MARLIN But it doesn't matter, 'cause no fish in this entire ocean is gonna help me. DORY Well, I'm helping you. Wait right here. Hey, guys. MOONFISH LEADER What, is he bothering you again? DORY No, no, he's a good guy. Go easy on him, he's lost his son, Fabio. Any of you heard of P. Sherman, 42 Wallaby Way, Sydney? MOONFISH LEADER Sydney? Oh sure. Why, Ted here's got relatives in Sydney. Don't you, Ted? MOONFISH TED Sure do. 27 DORY Oh, hey! They know Sydney! MARLIN [gasps] DORY You wouldn't know how to get there, would you? MOONFISH LEADER What you wanna do is follow the EAC, that's the East Australian Current. Big current, can't miss it, it's in..that direction. And then you gotta follow that for about, I don't know, what do you guys think? About three leagues? And that little baby's gonna put you right past Sydney. MOONFISH SCHOOL TA-DAA! MARLIN Great! That's great! Dory, you did it! DORY Oh, please. I'm just your little helper. Helping along, that's me. MARLIN Well, listen fellas, thank you. MOONFISH LEADER Don't mention it. And, uh, loosen up. Okay, buddy? DORY Oh, you guys. You really nailed him. Bye. MOONFISH LEADER Oh, hey ma'am, one more thing. DORY Yes. MOONFISH LEADER When you come to this trench, swim through it, not over it. DORY Trench, through it, not over it. I'll remember. Hey, hey! Hey! Hey! Hey, wait up, partner. Hold on. Wait! Wait-wait! I got, I gotta tell you something..whoa. Nice trench. Hello! Okay, let's go. MARLIN Bad trench, bad trench. Come on, we're gonna swim over this thing. DORY Whoa, whoa, partner. Little red flag goin' up. Somethin's telling me we should swim through it, not over it. MARLIN Are you even looking at this thing? It's got death written all over it. DORY I'm sorry, but I really, really, really think we should swim through. MARLIN And I'm really, really done talking about this. Over we go. DORY Come on, trust me on this. MARLIN Trust you? DORY Yes, trust. It's what friends do. MARLIN Look! Something shiny! DORY 28 Where? MARLIN Oh, it just swam over the trench. Come on, we'll follow it. DORY Okay. DORY Boy, sure is clear up here. MARLIN Exactly. And look at that, there's the current. We should be there in no time. DORY Hey, little guy. MARLIN You wanted to go through the trench. DORY I shall call him Squishy and he shall be mine and he shall be my Squishy. Come here, Squishy. Come here, little Squishy. [Baby talk]---oww! MARLIN Dory! That's a jellyfish! DORY Bad Squishy! Bad Squishy! MARLIN Shoo! Shoo, shoo! Get away! Come here, let me see. DORY Don't touch it! Don't touch it! MARLIN I'm not gonna touch it. I just wanna look. DORY Heeey, how come it didn't sting you? MARLIN It did. It's just that.. DORY Ow! Ow, oww! MARLIN ..hold still. I live in this anemone and I'm, I'm, I'm used to these kind of stings. Come here. DORY Ow, ow! Oww! MARLIN It doesn't look bad, you're gonna be fine. But now we know, don't we? DORY Yeah. MARLIN That we don't wanna touch these again. Let's be thankful this time it was just a little one.[gasps] MARLIN/DORY Aaaah! MARLIN Don't move! This is bad, Dory. DORY Hey, watch this! Boing! Boing! MARLIN [gasps] Dory! 29 DORY Boing-boing-boing! [singing] You can't catch me! MARLIN Dory! Don't bounce on the tops! They will..not sting you. The tops don't sting you, that's it! DORY Ooh! Two in a row, beat that. MARLIN Dory! All right, listen to me. I have an idea, a game. DORY A game? MARLIN A game. DORY A game? MARLIN Yes. DORY Aah! I love games! Pick me! MARLIN All right, here's the game. Um, whoever can hop the fastest out of these jellyfish, wins. DORY Okay! MARLIN Rules, rules, rules! DORY Okay! MARLIN You can't touch the tentacles, only the tops. DORY Something about tentacles, got it. On your mark, get set, go! MARLIN W-wait! Wait! Not something about them, it's all about them! Wait! DORY Weeee! MARLIN Dory! DORY Gotta go faster if you wanna win! MARLIN [gasps] Dory! DORY Boing! Boing! Boing-boing-boing-boing! MARLIN Wait a minute--whoa! Dory! DORY Weeee! MARLIN So, we're cheating death now. That's what we're doin'. We're havin' fun at the same time. I can do this, just be careful. DORY Yeah, careful I don't make you cry when I win! 30 MARLIN Oh, I don't think so! DORY Ha ha ha ha! Whooo! Give it up, old man. You can't fight evolution, I was built for speed. MARLIN The question is, Dory, are you hungry? DORY Huh? Hungry? MARLIN Yeah, 'cause you're about to eat my bubbles! Duck to the left! Right there! The clownfish is the winner! Woohoo! We did it! We're gonna...Dory? Oh no. Dory! Dory! Dory! [gasps] Dory! Uggghhh! DORY Ugh...am I disqualified? MARLIN No, you're doing fine! You're, you're actually winning! But you gotta stay awake. Uh, where does P. Sherman live? DORY P..Sherman..Wallaby Way...Sydney... MARLIN That's it! Oww! Ow! Stay awake! Stay awake! Ow! Stay awake! Stay--awake! DORY Awake...P..Sherman.. MARLIN Awake... DORY ..42 Wallaby Way... MARLIN Awake...wake up...Nemo... ====================================================================================== GILL You miss your dad, don't you, Sharkbait? NEMO Yeah. GILL Well, you're lucky to have someone out there who's lookin' for you. NEMO He's not looking for me. He's scared of the ocean. GILL Peach, any movement? PEACH He's had at least four cups of coffee, it's gotta be soon. GILL Keep on him. GILL My first escape, landed on dental tools. I was aimin' for the toilet. NEMO Toilet? GILL All drains lead to the ocean, kid. NEMO Wow. How many times have you tried to get out? 31 GILL Aah, I've lost count. Fish aren't meant to be in a box, kid. It does things to 'ya. BUBBLES Bubbles! Bubbles, bubbles, bubbles--- PEACH Potty break! Potty break! He just grabbed the Reader's Digest! We have 4.2 minutes. GILL That's your cue, Sharkbait. BLOAT You can do it, kid. GILL Okay, you gotta be quick. Once you get in, you swim down to the bottom of the chamber and I'll talk you through the rest. NEMO Okay. GILL Go on, it'll be a piece of kelp. NEMO [takes a deep breath] GILL Nicely done! Can you hear me? NEMO Yeah. GILL Here comes the pebble. Now, do you see a small opening? NEMO Uh-huh. GILL Okay, inside it you'll see a rotating fan. Very carefully, wedge that pebble into the fan to stop it turning. NEMO Aaah! GILL Careful, Sharkbait. NEMO I can't do it! PEACH Gill, this isn't a good idea. GILL He'll be fine. Try again. NEMO Okay. GILL That's it, Sharkbait. Nice and steady. NEMO I got it! I got it! PEACH [sigh] BLOAT He did it! GURGLE Whew! 32 GILL That's great, kid! Now, swim up the tube and out. NEMO Oh no! Gill! Gill! GILL Sharkbait! BLOAT Oh my gosh! GILL Get 'im outta there! Get 'im outta there! BUBBLES Help him! GURGLE What do we do!? What do we do!? PEACH Oh no! GILL Stay calm, kid! Just don't panic! NEMO Help me! GILL Sharkbait! Grab hold of this! NEMO No! No! GILL Feed me more! GURGLE That's it! GILL Come on, Sharkbait! Grab it! NEMO I got it! GILL Pull! PEACH Gill, don't make him go back in there. GILL No. We're done. ====================================================================================== CRUSH Dude. MARLIN Ooh... CRUSH Dude. Focus, dude. Dude. MARLIN Ooooh... CRUSH Oh, he lives! Hey, dude! MARLIN Ooooh..what happened? 33 CRUSH Oh, saw the whole thing, dude. First you were like, 'whoa'! And then we were all like, 'whoa'! And then you were like, 'whoa'. MARLIN What're you talking about? CRUSH You, mini-man. Takin' on the jellies. You got serious thrill issues, dude. MARLIN Ooh. CRUSH Awesome. MARLIN Ooh..ooh, my stomach. Ooooh.. CRUSH Oh, man. No hurlin' on the shell, dude, okay, just waxed it. MARLIN So Mr. Turtle... CRUSH Whoa, dude. Mr. Turtle is my father. Name's Crush. MARLIN Crush? Really? Okay Crush, listen I need to get to the East Australian Current. EAC? CRUSH Ha ha ha, dude, ha ha, you're ridin' it, dude! Check it out! CRUSH Okay, grab shell, dude! MARLIN Grabbing--waaaaaaaaaaaaaaaaah!!! Aaaaaaaaaaaah!!! Aaaaaaaaaaaah!!! Whooooooaaaa!!! CRUSH Ha ha! Righteous! Righteous! Yeah! MARLIN Stop! CRUSH So, what brings you on this fine day to the EAC? MARLIN Well, Dory and I need to get to Sydney. [gasps] Dory! Dory! Is she all right!? CRUSH Oh. Oh, Little Blue. She is sub-level, dude. MARLIN Dory, Dory! Dory! DORY Hmm-mmm.... MARLIN Oh, Dory. I-I-I'm so sorry. This is all my fault, it's my fault... DORY ..29, 30! Ready or not, here I come! There you are! Catch me if you can! Ha ha! Ha ha ha ha! MARLIN Huh? SQUIRT Whoa! MARLIN [gasps] Oh my goodnes! 34 CRUSH Whoa. Kill the motor, dude. Let us see what Squirt does flying solo. SQUIRT Whoa! Whoa! That was so cool! Hey dad, did you see that? Did you see me? Did you see what I did? CRUSH You so totally rock, Squirt! So give me some fin..noggin.. CRUSH/SQUIRT ..dude! CRUSH Oh, intro. Jellyman, Offspring. Offspring, Jellyman. SQUIRT Jellies? Sweet. CRUSH Totally. MARLIN Well, apparently, I must've done something you all like. Heh, uh, dudes. SQUIRT You rock, dude. MARLIN Ow. CRUSH Curl away, my son. Aw, it's awesome, Jellyman. Little dudes are just eggs, leave 'em on the beach to hatch, then coo-coo-ca-choo, they find their way back to the big 'ol blue. MARLIN All by themselves? CRUSH Yeah. MARLIN But-but-but dude, how do you know when they're ready? CRUSH Well, you never really know. But when they'll know, you'll know, you know? Ha. DORY Hey! Look, everybody! SQUIRT I know that dude. It's the Jellyman. DORY Well, go on, jump on him. TURTLE KIDS Turtle pile! MARLIN W-w-wai-wait-- TURTLE KID 1 Are you funny? TURTLE KID 2 Where's your shell? MARLIN Hold on, I need to breath-- TURTLE KID 3 Are you running away? TURTLE KID 4 Did you really cross the jellyfish forest? 35 TURTLE KID 5 Did they sting you? MARLIN One at a time! TURTLE KID 6 Mr. Fish, did you die? DORY Sorry. I was a little vague on the details. SQUIRT So where are you going? MARLIN Well, you see my son was taken. My son was taken away from me. TURTLE KIDS [gasp] DORY No way. SQUIRT What happened? MARLIN No, no, no, kids. I don't wanna talk about it. TURTLE KIDS Awww! Please? SQUIRT Pleeeease? MARLIN [sighs] Well, okay. I live on this reef, a long long way from here. DORY Oh, boy. This is gonna be good, I can tell. MARLIN And my son, Nemo, see he was mad at me. And maybe he wouldn't have done it if I hadn't been so tough on him, I don't know. Anyway, he swam out in the open water to this boat and when he was out there, these divers appeared and I tried to stop them but the boat was too fast. So we swam out in the ocean to follow them... TURTLE KID They couldn't stop them. And then Nemo's dad, he swims out to the ocean and they bump into.. SMALL FISH ..three ferocious sharks! He scares away the sharks by blowin' them up! BIG FISH Golly, that's amazing! SMALL FISH And then dives thousands of.. LOBSTER ..feet straight down into the dark. It's like wicked dark down there, you can see a thing. How's it goin', Bob? And the only thing that they can see down there.. SWORDFISH ..is the light from this big horrible creature with razor sharp teeth. Nice parry, old man. And then he has to blast his way... DOLPHIN So, these two little fish have been..searching the ocean for days. On the East Australian Current. FEMALE BIRD Which means that he may be on his way here right now. That should put them in Sydney.. MALE BIRD 1 ..Harbor in a matter of days. I mean, it sounds like this guy's gonna stop at.. 36 MALE BIRD 2 ..nothing until he finds his son. I sure hope he makes it. MALE BIRD 3 That's one dedicated father if you ask me. GULLS Mine! Mine! Mine! Mine! Mine! Mine! Mine! Mine! Mine! NIGEL Oh, would you just shut up! You're rats with wings! PELICAN ..bloke's been lookin' for his boy Nemo. NIGEL Nemo? PELICAN He was taken off the reef by divers and this.. NIGEL There, take it! You happy! GULLS Mine! Mine! Mine! Mine! NIGEL Hey, hey, hey! Say that again! You said something about Nemo. What was it? GULLS Mine! Mine! Mine! CRAB Whooooooaaa..watcha! GULL Mine! PELICAN Last I heard, he's headin' towards the harbor. NIGEL Ho ho! Brilliant! ====================================================================================== NEMO [sighs] DEB Is he doing okay? GURGLE I don't know, but whatever you do, don't mention D-A-R.. NEMO It's okay, I know who you're talking about. NEMO Gill? Gill? GILL Hey, Sharkbait. NEMO I'm sorry I couldn't stop the-- GILL No, I'm the one who should be sorry. I was so ready to get out, so ready to taste that ocean. I was willing to put you in harm's way to get there. Nothing should be worth that. I'm sorry I couldn't get you back to your father, kid. NIGEL All right! Hey, hey, hey, hey--! 37 DENTIST What the!? PATIENT AAAAAAAAAH!!! Oooooh... DENTIST Well, uh, that's one way to pull a tooth. He he he he he! Huh, darn kids. Well, good thing I pulled the right one, eh, prime minister? He he he he! NIGEL Hey, hey. Psst! PEACH Oh, Nigel. You just missed an extraction. NIGEL Ooh! Has he loosened the periodontal ligament yet--oh, what I'm talkin' about!? Nemo! Where's Nemo? I gotta speak with him. NEMO What? What is it? NIGEL Your dad's been fighting the entire ocean looking for you. NEMO My father? Really? GILL Really? NIGEL Oh yeah. He's travelled hundreds of miles. He's been battling sharks and jellyfish and all sorts of-- NEMO Sharks? That can't be him. NIGEL Are you sure? What was his name? Some sort of sportfish or something: tuna, uh, trout.. NEMO Marlin? NIGEL That's it! Marlin! The little clownfish from the reef. NEMO It's my dad! He took on a shark! NIGEL I heard he took on three. DEB/BLOAT/GURGLE Three!? GILL Three sharks!? BLOAT That's gotta be forty eight hundred teeth! NIGEL You see, kid, after you were taken by diver Dan over there, your dad followed the boat you were on like a maniac. NEMO Really? NIGEL He's swimming and he's swimming and he's giving it all he's got and then three gigantic sharks capture him and he blows them up! And then dives thousands of feet and gets chased by a monster with huge teeth! He ties this demon to a rock and what does he get for a reward? He gets to battle an entire jellyfish forest! And now he's riding with a bunch of sea turtles on the East Australian Current and the word is he's headed this way right now, to Sydney! 38 BLOAT Wow! Ha ha ha! DEB Oh, what a good daddy! GILL He was lookin' for you after all, Sharkbait. GILL [gasps] GURGLE He's swimming to the filter! GILL [gasps] Sharkbait! BLOAT Not again! GILL Sharkbait! DEB No! GURGLE You've got your whole life ahead of you! BLOAT Oh no! GILL We'll help you, kid! BLOAT Gotta get him out! DEB Gimme that thing! DEB Get him outta there! GURGLE Come on, kid! Grab the end! ALL [gasps] DEB Sharkbait! BLOAT Sharkbait! Are you okay!? GURGLE No! GILL Can you hear me, Sharkbait!? Nemo! Can you hear me!? NEMO Yeah, I can hear you. GILL Sharkbait, you did it! GURGLE Sharkbait, you're--covered with germs! Aaaaaaah!!! GILL That took guts, kid. GILL 39 All right, gang. We have less than 48 hours before Darla gets here. This tank'll get plenty dirty in that time but we have to help it along any way we can. Jacques! JACQUES Oui! GILL No cleaning. JACQUES I shall resist. GILL Everybody else, be as gross as possible. Think dirty thoughts. We're gonna make this tank so filthy, the dentist'll have to clean it. BLOAT [belch] GILL Good work. NEMO Ha ha ha ha! ====================================================================================== CRUSH All right, we're here, dudes! Get ready! Your exit's comin' up, man! MARLIN Where!? I don't see it! DORY Right there! I see it! I see it! MARLIN You mean the swirling vortex of terror!? CRUSH That's it, dude! MARLIN Of course it is. CRUSH Okay, first: find your exit buddy! CRUSH Do you have your exit buddy? DORY Yes! CRUSH Okay, Squirt here will now give you a rundown of proper exiting technique! SQUIRT Good afternoon, we're gonna have a great jump today! Okay, crank a hard cutback as you hit the wall! There's a screaming bottom turn, so watch out! Remember: rip it, roll it and punch it! MARLIN It's like he's trying to speak to me, I know it! You know, you're really cute! But I don't know what you're saying! Say the first thing again! CRUSH Okay, Jellyman! Go, go, go, go, go, go! MARLIN/DORY Aaaaaaaaaah!!! Weeeeeeeeeeee!!! Whoooooooooooaaaaa!!! Aaaaaaaaaaah!!! Woohoooo!!! Whoooooaaa!!! DORY Whoooo! MARLIN 40 Ha ha ha ha! That was..fun! Ha ha! I actually enjoyed that! DORY Hey, look! Turtles! CRUSH Ha ha! Most excellent! Now, turn your fishy tails 'round and swim straight on through to Sydney! No worries, man! MARLIN No worries! Thank you, dude Crush! TURTLE KIDS Bye! Bye, Jellyman! CRUSH You tell your little dude I said 'hi', okay? SQUIRT See you later, dudes! DORY Bye, everyone! MARLIN Oh, Nemo would've loved this. Hey, ooh! Hey, Crush! Crush, I forgot! How old are you? CRUSH Hundred and fifty, dude! And still young! Rock on! MARLIN Hundred and fifty! Hundred and fifty, I gotta remember that. DORY Whoa. We goin' in there? MARLIN Yup. DORY P. Sherman, 42 Wallaby Way, Sydney? MARLIN Yup. We're gonna just swim straight. DORY [singing] Just keep swimming, just keep swimming. MARLIN Dory? ====================================================================================== MARLIN Boy, this is taking a while. DORY Hey, how about we play a game? MARLIN Okay. DORY Uh, okay. I'm thinking of something, uh, orange. And it's small.. MARLIN It's me. DORY Right. Okay.. DORY ..orange, and uh, small.. MARLIN It's me. 41 DORY All righty, Mr. Smarty Pants. DORY ..orange and small, and white stripes.. MARLIN Me. And the next one's just a guess: me. DORY Okay, that's just scary. MARLIN W-w-wait, I have definitely seen this floating speck before. That means we've passed it before and that means we're going in circles and that means we're not going straight! DORY Hey. Hey! MARLIN We gotta get to the surface, come on! Let's figure it out up there. Let's go! Follow me! Wha--? DORY Whoa, whoa, whoa! Hey! Relax. Take a deep breath. Now, let's ask somebody for directions. MARLIN Oh, fine. Who do you wanna ask, the speck? There's nobody here! DORY Well, there has to be someone. It's the ocean, silly, we're not the only two in here. Let's see...okay, no one there. Uhh, nope. Nada. [gasps] There's somebody. Hey! Excuse-- MARLIN Dory! Dory! Dory! Okay, now it's my turn. I'm thinking of something dark and mysterious. It's a fish we don't know. And if we ask it directions, it could ingest us and spit out our bones! DORY What is it with men and asking for directions? MARLIN Look, I don't wanna play the gender card right now. You wanna play a card? Let's play the 'Let's Not Die' card. DORY You wanna get outta here, don't you? MARLIN Of course, I do. DORY Well then, how are we gonna do that unless we give it a shot and hope for the best? Hmmm? Hmmmm!? Come on, trust me on this. MARLIN All right. DORY Excuse me! Woohoo! Little fella? Hello. Don't be rude, say 'hi'. MARLIN Ha..hello. DORY His son Bingo.. MARLIN Nemo. DORY ..Nemo, was taken to, uh.. MARLIN Sydney. DORY 42 Sydney. Yes. And it's really, really important that we get there as fast as we can. So can you help us out? Come on, little fella. Come on. MARLIN Dory, I'm a little fella. I don't think that's a little fella. DORY Oh. Oh, oh, big fella. Big fe--whale. Okay. Maybe he only speaks whale. MOOOOO-WEEEEEEE-NEEEEED... MARLIN Uh, Dory..what're you doing? DORY TOOOOOOO-FIIIIIIND... MARLIN What're you doing? DORY HIS-SOOOOOOOOOOOON... MARLIN Are you sure you speak whale? DORY CAN-YOOOOOOOUUU-GIIIIIIIIIVE-USSSS-DIRECTIOOOOOOOONS-TOOOOOOOOO... MARLIN Dory! Heaven knows what you're saying! See, he's swimming away. DORY COOOME-BAAAAAAAAAAAAAACK! MARLIN He's not coming back. You offended him. DORY Maybe a different dialect. MOOOOOOOOOOOOOO! MOOOOOAAAAAAAAAA..! MARLIN Dory. Dory, this is not whale. You're speaking like..upset stomach. DORY Maybe I should try humpback. MARLIN No, don't try humpback. DORY WAAAAAAAAAAAAAAOOOOOOO!!! WAAAAAAAAAOOOOOO!!! MARLIN Okay, you actually sound sick. DORY Maybe louder, huh? RAAAH!!! RAAAAH!!! MARLIN Don't do that! DORY Too much orca. Didn't it sound a little orca-ish? MARLIN It doesn't sound orca! It sounds like nothing I've ever heard! DORY MOOOO..MOOOOOOOOOOOOOOO!!! MARLIN It's just as well, he might be hungry. DORY Don't worry. Whales don't eat clownfish, they eat krill. KRILL 43 Swim away! DORY Oh, look. Krill. MARLIN Move, Dory! Move! DORY Aah-aaah! Aaaaaaaaaah! ====================================================================================== GILL Look at that. Would you look at that? Filthy. Absolutely filthy. And it's all thanks to you, kid. You made it possible. Jacques, I said no cleaning! JACQUES I am ashamed. PEACH Hey, look. Scum angel. GURGLE Aah! Aaaah! Ooh-ooh! Aaaaah! BUBBLES Bubbles! I love the bubbles--! [coughs] DEB Flo! Flo! Has anybody seen Flo? Flo! PEACH Nine o' clock and cue dentist. DENTIST Hello, Barbara. Sorry I'm late. PEACH Okay. Okay, here we go. Here we go, okay. DENTIST Little Davey Reynolds. PEACH Okay. Walks to the counter, drops the keys.. GURGLE Bloat, that's disgusting! BLOAT Tastes pretty good to me. [belch] GURGLE Eww! Don't you people realize we are swimming in our own-- PEACH Shhh! Here he comes. DENTIST Crikey, what a state. Oh. Barbara, what's my earliest appointment tomorrow? BARBARA Uh, ten 'o clock, luv. DENTIST Leave it open, would you? I gotta clean the fish tank before Darla gets here. GILL He he! Did you hear that, Sharkbait? NEMO Yay! He's gonna clean the tank! He's gonna clean the tank! We're gonna be clean! GILL Are you ready to see your dad, kid? 44 NEMO Uh-huh. GILL Of course you are. Y'know, I wouldn't be surprised if he's out there in the harbor waitin' for you right now. NEMO Yeah. ====================================================================================== MARLIN Aaaaaaaaaaaah! Ooof! DORY Ha~~haaa~~haaaaaaah! Whooo! MARLIN Aaaaaaaaaaaah! DORY Here comes a big one--whooooooo! Come on, you gotta try this! MARLIN Would you just stop it!? DORY Why? What's wrong? MARLIN We're in a whale! Don't you get it!? DORY A whale? MARLIN A whale! 'Cause you had to ask for help! And now we're stuck here! DORY Wow. A whale. You know I speak whale. MARLIN No, you're insane! You can't speak whale! I have to get out! I have to find my son! I have to tell him how old sea turtles are! [sobs] DORY Woo-ho-ho-ho-ho-ho-hoo! Hey. You okay? DORY There, there. It's all right. It'll be okay. MARLIN No. No, it won't. DORY Sure it will, you'll see. MARLIN No. I promised him I'd never let anything happen to him. DORY Huh. That's a funny thing to promise. MARLIN What? DORY Well, you can't never let anything happen to him. Then nothing would ever happen to him. Not much fun for little Harpo. DORY Hmm.. MARLIN What's going on? 45 DORY I don't know. I'll ask him. MMMWWHAAAAAAAAA! HUUUWHAAAAAAAAA.. MARLIN Dory. Dory. MARLIN ..AAAAAAAAAAT'SSS-GOOIIIIIIING.. MARLIN Dory. DORY ..OOOOOOOOONNN? DORY I think he says we've stopped. MARLIN Of course, we've stopped. Just stop trying to speak whale, you're gonna make things worse. [gasps] What is that noise? Oh no. Look what you did. The water's going down! It's-it's-it's going down! DORY Really? You sure about that? MARLIN Look, it's already half-empty! DORY Hmm..I'd say it's half full. MARLIN Stop that! It's half-empty! DORY Okay, that one was a little tougher. He either said we should go to the back of the throat or he wants a root beer float. MARLIN Of course he wants us to go there! That's eating us! How do I taste, Moby!? Huh!? Do I taste good!? You tell him I'm not interested in being lunch! DORY Okay. HEEEEEEEEE-- MARLIN Stop talking to him--waaaah! DORY Aaaaaaaaaaaaaaaaaaah!!! MARLIN What is going on!? DORY I'll check! WHAAAAAAA--! MARLIN No! No more whale! You can't speak whale! DORY Yes, I can! MARLIN No, you can't! You think you could do these things but you can't, Nemo! DORY Okay. MARLIN Dory! DORY He says it's time to let go! Everything's gonna be all right! MARLIN 46 How do you know!? How do you know something bad isn't gonna happen!? DORY I don't! MARLIN/DORY AAAAAAAAAAAAAAAAAAHHH!!! AAAAAAAAAAAAAAAHHH!!! MARLIN Ha ha ha! We're alive! DORY Look! Sy-d-ney..Sydney! Uh, Sydney! Sydney again! MARLIN You were right, Dory! We made it! We're gonna find my son! MARLIN THAAAAAAAAAAAAAAANK-YOOOOOOOOOOOOOUUUU-SIIIRRRRRRRRRRRRRRRR! DORY Wow. I wish I could speak whale. MARLIN Okay. All we gotta do is find the boat that took him. DORY Right! MARLIN Come on, Dory. We can do this! ====================================================================================== PEACH [yawn] Morning. [gasps] It's morning, everyone! Today's the day! The sun is shining, the tank is clean and we are getting out of--[gasps]--the tank is clean. The tank is clean! DEB But how? GILL Boss must've installed it last night while we were sleepin'. NEMO What're we gonna do? GILL What's it say, Peach? PEACH [muffled] The AquaScum two-thousand.. GILL I can't hear you, Peach. PEACH 'The AquaScum 2003 is an all-purpose, self-cleaning maintenance free salt water purifier that is guaranteed to even extend the life of your aquarium fish'. BLOAT [inflates] Stop it! PEACH 'The AquaScum is programmed to scan your tank environment every 5 minutes'? GURGLE Scan? What does that mean? GURGLE Aaah! AQUASCUM Temperature: 82 degrees. PH balance: normal. ALL Oooooh. 47 PEACH Nice. GURGLE Ooh..ah..curse you, AquaScum! BLOAT That's it for the escape plan. It's ruined! NEMO Then what're we gonna do about-- ALL [gasps] Darla! GILL Stay down, kid! BLOAT False alarm. GURGLE My nerves can't take much more of this. BLOAT What're we gonna do when that little brat gets here? GILL I'm thinkin', I'm thinkin'. NEMO Aaah! Oh! Gill! GILL [gasps] Nemo! NEMO Help me! Help me! GILL Hold on! I'm comin'! NEMO Help me! GILL Swim down! Come on, kid! Swim down! Come on! BLOAT Everybody jump in! DEB Swim down! GILL That's it! DENTIST What the!? ALL Yay! GILL Good work! NEMO Gill! GILL [gasps] Nemo! BLOAT Sharkbait! GILL 48 Roll, kid! Lean! Lean! DENTIST Whoops. That would've been a nasty fall. NEMO Gill! Don't let me go belly up! GILL Just calm down, Nemo. NEMO Don't let me go belly up! GILL You won't go belly up, I promise. You're gonna be okay. ALL [gasps] Darla! ====================================================================================== DORY All right, do any of these boats look familiar to you? MARLIN No, but the boat has to be here somewhere! Come on, Dory, we're gonna find it. DORY I'm totally excited. [yawn] Are you excited? [yawn] MARLIN Dory, wake up, wake up. Come on. DORY [gasps] Duck! MARLIN That's not a duck. It's a--pelican! Whooooaaaaah! DORY Aaaaaaaaaaaah! MARLIN No! I didn't come this far to be breakfast! PELICAN Hey, hey, Nigel. Heh, would you look at that? NIGEL Huh? Wha-what? PELICAN Sun's barely up and already Gerald's had more than he can handle. NIGEL Yeah. Reckon somebody oughta help the poor guy. PELICANS Yeah, yeah, right. NIGEL Well, don't everybody fly off at once. NIGEL All right, Gerald, what is it? Fish got your tongue? DORY Aaaaaaaaaaaaaah!!! NIGEL Love a duck! MARLIN I gotta find my son Nemo! NIGEL 49 [gasps] Nemo? Hey, hey, hey! He's that fish! Y'know the one we were talking about! The one that's been fighting the whole ocean! Hey, I know where your son i--huh? Hey, wait! Come back! Stop! MARLIN Dory, keep going! He's crazy! NIGEL I got something to tell 'ya! GULL Mine. NIGEL Okay, don't make any sudden moves. Hop inside my mouth if you want to live. MARLIN Hop in your mouth, huh? And how does that make me live? GULL Mine. NIGEL Because I can take you to your son. MARLIN Yeah, right. NIGEL No. I know your son. He's orange, he's got a gimpy fin on one side.. MARLIN That's Nemo! GULLS Mine! Mine! Mine! Mine! Mine! Mine! DORY Aaaaaaaaaaaaaah!!! NIGEL Fasten your seatbelts! GULLS Mine! Mine! Mine! Mine! Mine! Mine! DORY Whoooooo! Woohooooo! GULLS Mine! Mine! Mine! Mine! Mine! Mine! DORY Ha-haaaa! Ha ha ha ha! MARLIN Aaaaaaaaaaaaaaaah! NIGEL Everybody hold on! MARLIN/DORY Aaaaaaaaaaaaaaaaah! GULLS Mine! Mine! Mine! Mine! Mine! Mine! ====================================================================================== BUBBLES Aaaah! Too loud! Too loud for me! DARLA [singing] Twinkle, twinkle little star. PEACH Find a happy place, find a happy place, find a happy place! 50 BARBARA Darla, you're uncle will see you now. DENTIST All right, let's see those pearly whites. DARLA RAAAH! I'm a piranha. They're in the Amazon. DENTIST And a piranha's a fish, just like your present. DARLA [giggling] I get a fishy! Fishy, fishy, fishy! DENTIST Oh no. Poor little guy. BLOAT He's dead! GILL Sharkbait! DARLA Yay! Fishy, fishy, fishy! DENTIST He he he! Must've left your present in the car, sweetie. Ha ha ha ha ha! DARLA Awwwww. DENTIST I'll go and get it. GILL [gasps] He's still alive! PEACH He's not dead! BLOAT What's happening? Why is he playing dead? GILL He's gonna get flushed down the toilet! He's gonna get outta here! DEB Yay! BLOAT He's gonna get flushed! GURGLE What a smart little guy! GILL Oh no, not the trash can! BUBBLES Nemo! No! NIGEL Hey! Hey! I found his dad! MARLIN Where's Nemo!? Where is he!? BLOAT Dentist! Dentist! GILL He's over there! MARLIN 51 What's a dentist!? What is that!? [gasps] Nigel, get in there! NIGEL I can't go in there. MARLIN Oh yes, you can! Charge! DARLA Aaaaaaaaaaaah! DENTIST What the--!? Darla, sweetie! Look out! DARLA Aaaaaaaah! DENTIST Hold still! DARLA Aaaaaaaah! DENTIST Easy! Easy! DARLA Aaaaaaaah! DENTIST Hold still! Nobody's going to hurt you! Oof! MARLIN [gasps] Nemo. DORY [gasps] Oh my goodness. DENTIST Gotcha! Keep down! MARLIN Nemo! NEMO Daddy? DENTIST Out with 'ya! And stay out! NEMO Daddy!? DARLA Fishy? Fishy! Wake up! Wake up! DEB Oh no! GILL Quick! To the top of Mt. Wannahockaloogie! DARLA Why are you sleeping!? PEACH Hurry! GILL Bloat! Ring of Fire! DARLA Fishy--aaaaaaaaaaaah! Aaaaaaaaaah! DENTIST What!? All the animals have gone mad! 52 DARLA Aaaaaaaah! Get it out! GURGLE Smack her in the head! BLOAT Go, Gill! Go! DARLA Fish in my hair! Aaaaaaaah! NEMO Gill. GILL Sharkbait. Tell your dad..I said..hi. Go get 'em. DENTIST Ooooh. [gasps] BLOAT He did it! Ha ha! DEB Yay! BUBBLES I'm so happy! GURGLE Is he gonna be okay, Gill? GILL Don't worry. All drains lead to the ocean. DARLA Fishy! NEMO Aaaaaaaaaaaaaah! Daddy! ====================================================================================== NIGEL I'm, I'm so sorry. Truly, I am. DORY Hey.. MARLIN Dory. If it wasn't for you, I never even would have made it here. So, thank you. DORY Hey! Hey, wait a minute. W-w-wait! Where are you going? MARLIN It's over, Dory. We were too late. Nemo's gone and I'm going home now. DORY No..no, you can't! Stop! Please don't go away. Please? No one's ever stuck with me for so long before. And if you leave, if you leave...I just, I remember things better with you. I do. Look, P. Sherman, 42..40..2..agh! I remember it, I do. It's there, I know it is because when I look at you, I can feel it. And I, I look at you and...I'm home. Please. I don't want them to go away. I don't wanna forget. MARLIN I'm sorry, Dory, but I do. ====================================================================================== CRAB 1 Manna from heavens! CRAB 2 Sweet nectar of life! 53 CRAB 1/CRAB 2 Hey! Hey, hey! Hey! CRAB 1 This is our spot! CRAB 2 Go on! Get outta here! CRAB 1/CRAB 2 Hey, hey! Hey! Hey, hey, hey! CRAB 1 Yeah, that's it fella! Just keep on swimmin', you got that! CRAB 2 Too right, mate! Oh, Oh! I got a live one here! NEMO Hey, have you seen my dad? CRAB 2 Gotcha! Hey! Hey! Come back here! CRAB 1 You let 'im go! CRAB 1/CRAB 2 Hey! Hey, hey, hey! NEMO Dad! Dad! Dad! DORY Aah! No! NEMO Um, excuse me. Are you all right? DORY I don't know where I am! I don't know what's going on, I think I lost somebody but I, I can't remember. NEMO It's okay, it's okay. I'm looking for someone too. Hey, we can look together. DORY I'm Dory. NEMO I'm Nemo. DORY Nemo? That's a nice name. ====================================================================================== NEMO Dad! DORY Dad! NEMO Dad! DORY Dad! Wait a minute, is it your dad or my dad? NEMO My dad. DORY Got it. Dad! NEMO Where are we, anyway? 54 DORY Dad! Dad! Oh. S-ss-syl--shi--Sydney. [gasps] 'P. Sherman, 42 Wallaby Way, Sydney'. DORY Aaaaah! Nemo! It's you! Aaaaaah! You're Nemo! NEMO [muffled] Yes! Yes! I'm Nemo! DORY Oh! You're Nemo! [gasps] You were dead! I saw you! And then I--[gasps], here you are! I found you! You're not dead! And your father--[gasps]! Your father! NEMO My father!? You know my father!? Where is he!? DORY [gasps] This way! He went this way! Quick! DORY Hey! Hey, hey! Hey! CRAB 1/CRAB 2 Hey! Hey, hey, hey! DORY Hey! Have you seen an orange fish swim by? It looks just like him! NEMO But bigger! CRAB 2 Yeah, I saw 'im, bluey! But I'm not tellin' you where he went. And there's no way you're gonna make me! GULL Mine. CRAB Huh!? Aaaah! All right! I'll talk! I'll talk! He went to the fishing grounds! Aaaaah! GULLS Mine!Mine! Mine! Mine! Mine! Mine! ====================================================================================== FISH Hey! Look out! MARLIN Sorry. Just trying to get home. NEMO Dad! Dad! MARLIN Nemo? NEMO Daddy! MARLIN Nemo? NEMO Dad! DORY Nemo's alive! MARLIN Dory? [gasps] Nemo! NEMO Daddy! 55 MARLIN Nemo! I'm coming, Nemo! NEMO Dad! MARLIN Nemo! NEMO Dad! MARLIN Oh, thank goodness! It's all right, son. It's gonna be okay. FISH Turn around! You're going the wrong way! Aaaaaaaaaaah! DORY Aaaaaaaaaaaah! Look out! MARLIN Move! Move! FISH Aaaaaaaaaaaah! DORY Help! AAAAAAAAAAAAH!!! MARLIN Dory! NEMO Come on! DORY Heeeeeeeelp!!! Help! NEMO Dory! DORY Help! Get us out! Aaaaaaaah! MARLIN No, no, no! No! Dory! NEMO Dad! I know what to do! MARLIN Nemo! No! NEMO We have to tell all the fish to swim down together! MARLIN Get out of there, now! NEMO I know this will work! MARLIN No, I am not gonna lose you again! NEMO Dad, there's no time! It's the only way we can save Dory! I can do this! MARLIN You're right. I know you can. NEMO Lucky fin! MARLIN Now go! Hurry! 56 NEMO Tell all of the fish to swim down! MARLIN Well!? You heard my son! Come on! NEMO Dory! DORY [gasps] NEMO You have to tell everybody to.. MARLIN ..swim down together! Do you understand what I'm saying to you!? Swim down! DORY Everybody swim down! NEMO Come on! You have to swim down! DORY Swim down, okay? NEMO Swim.. MARLIN down! Swim down! Swim down! Swim down! MARLIN Don't give up! Keep swimming! Just keep swimming! NEMO It's working! FISH Keep swimming! Keep swimming! Keep swimming! MARLIN Just keep swimming! Keep swimming! NEMO Come on, dad! MARLIN You're doing great, son! NEMO That's my dad! MARLIN Come on! Let's get to the bottom! Keep swimming! DORY [singing] Just keep swimming, just keep swimming. MARLIN Almost there! Keep swimming! FISH Keep swimming! Keep swimming! Keep swimming! Keep swimming! Yay! MARLIN Oww! DORY Hey! MARLIN Dory! Where's Nemo!? DORY 57 [gasps] There! MARLIN Oh no. Nemo! MARLIN Nemo? Nemo? It's okay. Daddy's here, daddy's got you. NEMO [coughs] Daddy? MARLIN Oh, thank goodness. NEMO Dad...I don't hate you. MARLIN No, no, no. I'm so sorry, Nemo. MARLIN Hey, guess what? NEMO What? MARLIN Sea turtles? I met one! And he was a hundred and fifty years old. NEMO Hundred and fifty? MARLIN Yep. NEMO 'Cause Sandy Plankton said they only live to be a hundred. MARLIN Sandy Plankton? Do you think I would cross the entire ocean and not know as much as Sandy Plankton!? NEMO Ha ha ha ha! MARLIN He was a hundred and fifty! Not one hundred! Who is this Sandy Plankton who knows everything? ====================================================================================== MARLIN Time for school! Time for school! Get up! Let's go! Go! MARLIN I'm gonna win! NEMO No, you're not! I did it! Woohoo! Ha ha ha! MARLIN Oh! My own son beats me! MR. RAY Climb aboard, explorers! MARLIN So just then, the sea cucumber looks over to the mollusk and says : 'with fronds like these, who needs anemones?'! BOB/TED/BILL Haaa-ha ha ha ha ha ha! MR. RAY Well, hello, Nemo! Who's this? NEMO Exchange student. 58 SQUIRT I'm from the EAC, dude! MR. RAY Sweet. NEMO/SQUIRT Totally. BOB But seriously, Marty, did you really do all the things you say you did? BRUCE Uh, pardon me. BOB/TED/BILL [gasps] BRUCE Hello. TED Ohh! BRUCE Don't be alarmed. ANCHOR Oh, we just wanna make sure that our newest member got home safe. DORY Thanks, guys. BRUCE Well, we'll see you next week. CHUM Keep up with the program, Dory. ANCHOR Remember: fish are friends.. DORY ..not food! Bye! MR. RAY Hold on! Here we go! Next up, knowledge! MARLIN Bye, son! Have fun! NEMO Bye, dad! Oh! Oh, Mr. Ray! Wait. I forgot something. NEMO Love you, dad. MARLIN I love you too, son. NEMO Uh, dad, you can let go now. MARLIN Sorry! Now go have an adventure! SQUIRT Goodbye! See you later, dudes! DORY Bye, Elmo! MARLIN Nemo. DORY 59 Nemo! Bye, Nemo! NEMO See you after school, Dory! Bye, dad! MARLIN Bye, son. ====================================================================================== DENTIST Barbara? BARBARA Uh-huh? DENTIST I don't understand it. Here this thing has a lifetime guarantee and it breaks! Had to clean the tank myself, take all the fish out, put 'em in bags and---where'd the fish go? GILL Come on, Peach! DEB Hurry! GILL You can do it! BLOAT Yeah, that's it! You can do it! GURGLE Just a little further! PEACH That's the shortest red light I've ever seen! BLOAT Come on, Peach! PEACH Oooh--aaaaah! ALL Yay! We did it! Ha ha ha ha ha! BLOAT Now what? ###################################################################################### # FINDING NEMO, and all related media, characters, and stories # # are copyright 2003 Walt Disney Pictures and Pixar Animation Studios. # # The transcript below contains parts of a screenplay written by Andrew Stanton, # # Bob Peterson and David Reynolds. This transcript is provided for fans' enjoyment # # and reference and does not intend copyright infringement. The entire content of # # this transcript is property of Andrew Stanton, Bob Peterson and David Reynolds, # # Walt Disney Pictures and Pixar Animation Studios. # # No claim is lain on the ownership of the words contained within this transcript # # on the part of BaD_BURN. # # # # GIVE CREDIT WHERE CREDIT IS DUE. RETAIN THIS COMMENT BLOCK. # # # # The transcript is intended for teaching /educational purposes only. It falls under # # the U.S. Code 17/Sec. 107 - Limitations on exclusive rights: 'Fair Use'. # # Notwithstanding the provisions of sections 106 and 106A, the fair use of a # # copyrighted work, including such use by reproduction in copies or phonorecords or # # by any other means specified by that section, for purposes such as criticism, # # comment, news reporting, teaching (including multiple copies for classroom use), # # scholarship, or research, is not an infringement of copyright. # ###################################################################################### 60 61 62 | 1 | 5.3% |
| MUSIC ENHANCEMENT VIA IMAGE TRANSLATION AND VOCODING Nikhil Kandpal* Oriol Nieto, Zeyu Jin University of North Carolina at Chapel Hill Computer Science Department Chapel Hill, NC, USA Adobe Research San Francisco, CA, USA ABSTRACT to develop a solution that works for polyphonic signal enhancement and reflects the unique qualities of music perception. Our approach performs enhancement on the recording' s melspectrogram representation. This is achieved by treating the melspectrogram as an image and training an image-to-image translation model similar t0 Pix2Pix [3] to transform low-quality melspectrogram into that of a high-quality signal. We hypothesize that it is easier to enhance polyphonic signals in the mel-spectrogram domain as polyphonic sources are additive and have a very small temporal span compared to waveforms Finally; to map generated high-quality mel-spectrograms to perceptually realistic waveforms we train vocoding model based on DiffWave [4|. Training this model on only the high quality samples of music performance makes it robust to the artifacts that reside in the synthetic mel-spectrogram_ We evaluate Our approach by performing listening test with 211 participants, and we show that this approach achieves a much better perceptual enhancement than several state-of-the-art techniques_ We also compare the subjective listening test scores with widely used audio quality metrics and suggest that, similar to speech enhancement, these metrics correlate poorly with human perception [1[5 : With this work; we hope to motivate both future research in music enhancement as well as music quality perceptual metrics akin to those in the speech literature [6], [1. To promote further research, audio samples generated in our experiments and source code are provided at Our project website In this paper; we refer to Pix2Pix models operating on melspectrograms as MelzMel models and vocoding applied to the music domain as musecoding. We summarize our contributions as follows: A music enhancement model leveraging recent work on conditional image synthesis and vocoding: generative process for simulating realistic low-quality music recordings from professional-quality recordings An analysis of the reliability of common audio enhancement evaluation metrics in the music domain_ 8 Consumergrade music recordings such as those captured by mo bile devices typically contain distortions in the form of background noise, reverb_ and microphone-induced EQ. This paper presents 2 deep learning approach to enhance low-quality music recordings by combining (i) an image-to-image translation model for manipulating audio in its mel-spectrogram representation and (ii) a music vocod8 ing model for mapping synthetically generated mel-spectrograms to perceptually realistic waveforms We find that this approach to music enhancement outperforms baselines which use classical methods for mel-spectrogram inversion and an end-to-end approach directly mapping noisy waveforms t0 clean waveforms. Additionally, 8 in evaluating the proposed method with listening test; we analyze the reliability of common audio enhancement evaluation metrics when used in the music domain: Index Terms _ Music Enhancement; Image-to-Image Translation, Diffusion Probabilistic Models, Vocoding 7 1. INTRODUCTION With the rise of Internet influencers and music hobbyists, large portion of music content is created with cheap and accessible recording devices in non-treated environments _ While being audible, these recordings often have degraded quality stemming from background noise, unpleasant reverb, and resonance caused by the microphone and the environment_ This prompts US t0 investigate quality enhance1 ment for music signals, transforming low-quality amateur recordings into professional ones. The main difficulty of such an endeavor is that $o many aspects of the low-quality recording setup are unknown Parameters of the recording device, such as frequency response characteristics, vary drastically across different hardware. Additionally, acoustic properties such as the size, shape, and reflectivity of the recording environment vary between different recording setups_ Finally, background noise is hard to capture and generalize, especially non-stationary 2. RELATED WORK noise. A solution that faithfully transforms low-quality recording into what it would sound like recorded professionally must implicTo our knowledge there is little prior work studying music quality itly or explicitly infer all of these aspects from the signal alone In enhancement. The work most similar to our contributions focuses on speech enhancement; end-to-end methods such as HiFi-GAN [1 and speech enhancement; conditional speech synthesis, O music source Demucs |2] achieve this by extracting the speech source from a mixseparation_ ture of sources However; music signals are often polyphonic, i. eEarly approaches to speech enhancement have used classithere can be an arbitrary number of sources to be extracted at once_ cal signal processing techniques such as Wiener filtering [8] and Moreover; the perception of music quality typically differs from that non-negative matrix factorization [9|. More recently, deep learningof speech: For example, human listeners may find reverb pleasant based methods have achieved state-of-the-art on speech enhancein music, while it is usually undesired in speech. Therefore, we aim ment_ These methods either manipulate the audio in its magnitude Work done during an internship with Adobe Research Ihttps: nkandpa2. github 10 music enhancement High-Quality Spectrogram MelzMel GAN Musecoder DDPM Low-Quality Spectrogram Conv Encoder Conv Decoder Denoising Steps High-Quality Waveform Noise Diffusion Steps Fig: 1. Model architecture of our Mel2Mel + Diffwave model. First, low-quality mel-spectrogram is enhanced by conditional GAN. The resulting synthetic mel-spectrogram is then "musecoded" into a waveform by a Denoising Diffusion Probabilistic Model (DDPM) spectrogram representation (followed by spectrogram inversion method to recreate the corresponding waveform) |10} H[2| or map directly from the low-quality waveform to a cleaned waveform 113] 2 : Methods that operate on the time~frequency domain generally produce audible artifacts due to the use of phase reconstruction algorithms like the Griffin-Lim algorithm [14] A recent work addresses this with neural-network based vocoders [15], yet its quality is not 0 par with an end-to-end approach [16p Alternatively; methods that work on the time domain typically require more training steps |1| Conditional speech synthesis techniques produce speech wave _ forms from conditioning information such as magnitude spectrograms, problem commonly known as vocoding Some state-of-theart vocoding methods involve using generative adversarial networks [17[8, or denoising diffusion probabilistic models 4/19] for generating audio. Music source separation focuses on taking a mix of multiple music stems (vocals, drums, etc. ) and separating the mix into its individual sources Some approaches to music source separation operate by masking spectrograms |20] or directly mapping the mix waveform t0 individual source waveforms [21122/_ The music enhancement problem is different than music source separation, since our goal is not only to extract all musical sources from a noisy mixture but also to reduce reverb and adjust EQ such that the listening experience is improved dataset we assume access to high-quality recordings and define generative process for simulating low-quality ones. First, we simulate the reverb and varied microphone placements of a nonprofessional recording environment by convolving the high-quality music signal with a room impulse response. Next; we apply additive background noise scaled to achieve randomly sampled SNR between 5 and 30 dB. Finally, we simulate a low-quality microphone frequency response by applying 4-band equalization with randomly sampled gains between15 and 15 dB and frequency bands from 0-200, 200-1000, 1000-4000, and 4000-8000 Hz 3. 3. Mel-Spectrogram Enhancement with Mel2Mel Our first step in music enhancement is modeling the distribution of high-quality mel-spectrograms conditioned 0n their low-quality counterparts To estimate this distribution, we use existing work on image-to-image translation with conditional adversarial networks in an approach similar to [12|. In this framework a generator and a discriminator are trained using an aligned dataset of low and high-quality recording pairs. The generator maps from low to high-quality mel-spectrograms with the objective of maximizing the discriminator's loss and minimizing the C1 distance between the generated mel-spectrogram and the ground truth high-quality mel-spectrogram. The discriminator is trained to classify whether a given mel-spectrogram is generated or comes from the true data distribution. It performs this classification on patch-wise basis, predicting a class for each patch in the input melspectrogram_ For this reason, the discriminator acts aS a learned loss function for the generator which enforces realistic local features and the /1 loss enforces global consistency with the ground truth melspectrogram_ 3. 0 METHODS 3. 1. Modeling Approach In this paper; we investigate the approach of enhancing music in its mel-spectrogram domain, as it is easier to represent complex harmonic structures and polyphonic sound sources_ We then transform the resulting mel-spectrograms to waveforms through Diffwavebased vocoder (a process that in this context could be more aptly named "musecoding"). Decoupling waveform generation from melspectrogram enhancement allows us to train a musecoder that is not only robust to noise and other artifacts, but can also be used for any generation and enhancement task without the need of retraining: Figure [I depicts a block diagram of our proposed architecture: This approach is motivated by recent advances in vocoding that generate natural-sounding speech from mel-spectrograms [4]. 3. 4. Musecoding Recent work has shown that deep learning models can generate perceptually realistic waveforms from speech mel-spectrograms _ In our experiments, we evaluate the Diffwave |4] vocoder applied to music a process that we call "musecoding" Diffwave is a denoising diffusion probabilistic model (DDPM) This class of models defines forward diffusion process which iteratively adds gaussian noise to audio waveforms from the training dataset A model is then trained to estimate the reverse transition distributions of each noising step conditioned on the mel-spectrogram of the clean audio. Sampling from this model requires sampling noise from a standard gaussian and iteratively denoising using the reverse transition probability distributions from the model_ For further discussion of DDPMs see [4] and [23|. 3. 2 Data Simulation The modeling techniques we consider in this paper require aligned pairs of highand low-quality music recordings. To construct such Model Clean MOS 1 4. 39 = 0. 05 4. 06 = 0. 06 3. 01 + 0. 09 2. 85 =0. 09 4. 3. Baselines We evaluate our approach against two separate baselines First, we pair Mel2Mel for mel-spectrogram enhancement with inverse mel-scaling and the Griffin-Lim algorithm for musecoding. Both inverse mel-scaling and Griffin-Lim require solving optimization problems [291, So we run both solvers for 100 iterations, which yields per-sample runtime comparable to that of the Diffwave musecoder: Our second baseline is an end-to-end approach for music enhancement: Namely, we use the Demucs model architecture [21 and train it using the /1 reconstruction loss on our dataset of lowand high-quality recording pairs This matches the original training objective used for this architecture on the task of music source separation. We train this model for 360 epochs with batch size 64 and learning rate 0. 0003. We find that after this number of epochs the validation loss plateaus MelzMel + Diffwave Mel2Mel + Griffin-Lim No Enhancement Table 1. Mean Opinion Scores in a human listening test. As a musecoding baseline, we also consider mel-spectrogram inversion with inverse mel-scaling and the Griffin-Lim algorithm [14]. 4. EXPERIMENT SETUP 4. 1. Dataset We train and evaluate models on the Medley-solos-DB dataset [24]_ containing 21, 572 three-second, single-instrument samples recorded in professional studios. We exclude the distorted electric guitar samples to avoid fitting our models to production effects. We use 5841 samples for training, 3494 for validation and the rest for testing: We start by downsampling Our data to 16 kHz following the setup of prior vocoding work [4[71. This sample rate has shown to be favored by most speech enhancement work [D[2 and can be pOtentially super-resolved to 48 kHz with bandwidth extension techniques [5]. Using the procedure described in Section[. 2] we generate a dataset of 'highand low-quality recording pairs. For simulation of low-quality recordings, we source room impulse responses from the DNS Challenge dataset [25_ and realistic background noise from the ACE Challenge dataset |26/. As a final step, we apply a low-cut filter to remove nearly inaudible low frequencies below 35 Hz and normalize the waveforms to have a maximum absolute value of 0. 95_ We find that this treatment helps improve our models' training stability: When evaluating, we apply the same treatment (low-cut filter at 35 Hz and normalization) before applying OUr enhancement models. 4. 4. Evaluation Metrics To evaluate the results of different enhancement models we conducted Mean Opinion Score (MOS) test with human listeners on Amazon Mechanical Turk (AMT) Additionally; we evaluate enhancement methods by computing the frequency-weighted segmental SNR (fwSSNR) 1301, multi-resolution spectrogram loss (MRS) [34| C1 spectrogram distance, and Frechet Audio Distance (FAD) [32 between enhanced and clean reference signals_ In Sec tion|5. 3 we analyze the effectiveness of these objective metrics at approximating human listener ratings in the music domain. 5. 0 RESULTS 5. 1. Mean Opinion Score Test To evaluate our proposed MelzMel Diffwave music enhancement model, we conducted an MOS test with human listeners on AMT. We used 200 audio samples from our test set, added 8 different types of simulated degradation and passed these low-quality waveforms through our method, MelzMel + Griffin-Lim; and Demucs_ The lowquality, enhanced, and ground truth high-quality samples were then presented to human listeners who were asked to give a quality score from to 5. We used the ground truth high-quality recording as high anchor and the same recording with 0 dB white noise as low anchor Each Human Intelligence Task (HIT) started with screening test in which human listeners were required to identify which one of 5 audio samples sound the same aS a reference sample. 4 out of the 5 samples are passed a small amount of effects including low pass filters, high pass filters, comb filters, and added noise. Passing the screening test was required to continue The rest of the HIT consisted of 34 tests in which were validation tests to check if listeners were paying attention_ If they failed the validation test, the entire HIT was invalidated. In the end we collected 9, 095 answers from 214 listeners. The results shown in Table 1 suggests that MelzMel with a Diffwave musecoder achieves the highest MOS with a score near that of clean audio from the dataset_ 4. 2. Model Architectures and Hyperparameters In all experiments, we compute mel-spectrograms with 128 mel bins, an FFT size of 1024, and 256 sample hop length: When training models that generate O are conditioned 0n mel-spectrograms, we use log-= scale amplitudes to reduce the range of values and to avoid positive restrictions on our models domain Or range. The Mel2Mel generator described in SectionB3] consists of 2 downsampling blocks; each containing a 2D convolutional kernel of size 3 and stride 2, instance normalization |27 | and ReLU activation functions. This is followed by 3 ResNet blocks 28] with kernel size 3 and instance normalization_ Finally, the representation is upsampled back to the original dimensionality of the input with two upsampling blocks, each containing a transposed convolutional kernel of size 3 and stride 2, instance normalization, and ReLU activation functions. The Mel2Mel discriminator is a fully convolutional model made Up of three blocks, each containing a convolutional kernel of size 4 and stride 2, instance normalization, and LeakyReLU activation function. The last layer does not have any normalization O activation function Both the generator and discriminator are trained with batch size of 64 and learning rate of 0. 0002 for 200 epochs The Diffwave model described in SectionBAuses the architecture and training objective described in /4]. We train this model for 3000 epochs using a batch size of & and a learning rate of 0. 0002 5. 2. Perturbation Ablation Study To gain insight into which perturbations are handled most effectively by each enhancement model, we perform an ablation study isolating each perturbation introduced in the low-quality signal generative process Table[contains mean opinion scores for each enhancement Model Clean Random EQ 4. 35 = 0. 06 4. 15 =0. 07 2. 98 = 0. 1 3. 39 + 0. 10 3. 99 + 0. 08 SNR 5 SNR 10 4. 27 =0. 06 4. 24 = 0. 06 3. 53 + 0. 09 3. 07 = 0. 1 2. 71 +0. 1 SNR 15 4. 46 = 0. 06 3. 96 = 0. 09 3. 18 = 0. 11 2. 85 = 0. 11 3. 04 = 0. 12 DRR 0 DRR 3 DRR 6 4. 24 =0. 07 4. 01 = 0. 08 3. 10 = 0. 08 2. 55 = 0. 10 2. 48 + 0. 11 4. 28 = 0. 04 4. 19 = 0. 06 3. 77 = 0. 06 3. 84 = 0. 06 2. 82 + 0. 07 2. 77 _ 0. 09 3. 13 = 0. 07 3. 21 + 0. 07 4. 01 = 0. 06 3. 91 = 0. 07 4. 42 =0. 06 3. 96 = 0. 08 2. 99 + 0. 10 3. 30 + 0. 10 4. 21 +0. 07 Mel2Mel + Diffwave Mel2Mel + Griffin-Lim Demucs No Enhancement Table 2. Mean Opinion Scores in a human listening test. Each column contains the ratings for single perturbation type: EQ, additive background noise at different signal-to-noise ratios (SNR), and reverb at different direct-to-reverberant ratios (DRR)_ Enhancement Metric Rank Correlation with MOS fwSSNR 0. 5 ~MRS 056 ~LI 0. 4 ~FAD 053 Model fwSSNR 9. 04 7. 61 6. 58 8. 23 6. 96 MRS 1. 40 1. 57 1. 65 1. 80 1. 89 Ll + 1350 1, 57 1. 69 1. 83 2. 16 FAD 4. 73 4. 54 3. 98 5. 54 5. 90 Independent Training Joint Fine-tuning Joint Training Sequential Training No Enhancement Table 3. Spearman rank correlation between MOS test ratings and audio enhancement metrics. Table 4. Performance of MelzMel + Diffwave enhancement models using different training schemes model applied to signals with randomly sampled EQ, additive noise with signal-to-noise ratios (SNR) of 5, 10, and 15 dB, and reverb with direct-to-reverberant ratios (DRR) of 0, 3, and 6 dB. This ablation shows that the Mel2Mel Diffwave model excels at removing noise even at SNR values as low as 5 dB and at undoing 4-band equalization simulating a non-flat microphone frequency response. Interestingly, none of the models tested perform dereverberation very well, and in fact degrade signals that contain no noise and only simulated reverb: This may be due to train-test mismatch; since all samples enhanced during training time contained some level of additive noise. This ablation also lends insight into the types of perturbations that affect human listeners"perception of music_ From the difference between the scores given to clean samples and non-enhanced samples, it is clear that additive noise impacts the listener'$ perception significantly while reverb is mostly ignored. 5. 4. Alternate Training Schemes In Section |3. 1] we motivated approaching music enhancement by training two decoupled models that separately handle melspectrogram enhancement and musecoding: Here, we investigate training schemes for these models other than independently training them on their respective tasks In addition t0 independent training, we (1) finetune the Mel2Mel generator and Diffwave musecoder jointly using the Diffwave objective, (2) train the models sequentially by first training the musecoder and then training the Mel2Mel generator with musecoder parameters frozen; and (3) train the MelzMel generator and musecoder jointly as single model using the Diffwave objective. Table shows the performance of the resulting models. In Section[3]we discussed the reliability of using these metrics for evaluating algorithms, and find that FAD is the most perceptually aligned metric when it comes to denoising Given this observation, our results suggest that joint training may yield better denoising performance than independent training: Joint training has the added benefit that only a single model is trained using a non-adversarial objective. However; this comes with the downside that the trained model cannot be split into enhancement and musecoding sub-models: Future work could focus on further exploring such training schemes. 5. 3. Perceptual Alignment of Objective Metrics The results of the MOS test also provide a mechanism t0 evaluate how well objective metrics for audio quality align with human perception in the music domain We measure fwSSNR, MRS, FAD, and 61 spectrogram distance on the same samples submitted for MOS evaluation_ We then take the mean score across all samples with given perturbation type (i. e. SNR 5, DRR 0, etc. ) and perform Spearman rank correlation with the mean scores measured in the human MOS test_ In Tablel3] we show the rank correlation for each objective metric_ We find that none of the four metrics evaluated correlate very strongly with human opinion scores, the highest achieving a rank correlation of 0. 56_ We also identify particular failure modes of these metrics AlL four metrics fail to identify robotic artifacts induced by the GriffinLim algorithm and actually rate the Mel2Mel + Griffin-Lim model as the best of all models we tested. Additionally, fwSSNR MRS, and 61 spectrogram distance all fail to identify additive noise effectively, and rate non-enhanced samples at SNR values of 10 and 15 dB as being better than any enhancement model output: FAD does not have this failure mode. 6. CONCLUSION We propose a music enhancement model that decomposes the task into mel-spectrogram enhancement and waveform synthesis from mel-spectrograms_ This model was trained using high-quality samples from a public dataset paired with low-quality samples generated by simulating artifacts that typically appear in amateur recordings_ A human MOS test shows that this model outperforms state-of-theart baselines Additionally, we found that current objective metrics for audio enhancement do not accurately reflect human perception of music. We hope this work encourages researchers to further advance the rather unexplored and yet timely topic of automatic music enhancement; either by designing more performant models 01 by proposing metrics that better align with human music perception. 7 _ REFERENCES [17] Kundan Kumar; Rithesh Kumar; Thibault de Boissiere, Lucas Gestin, Wei Zhen Teoh, Jose Sotelo, Alexandre de Brebisson Yoshua Bengio, and Aaron Courville_ "Melgan: Generative adversarial networks for conditional waveform synthesis,"2019. [18] Jaeseong You_ Dalhyun Kim; Gyuhyeon Nam, Geumbyeol Hwang, and Gyeongsu Chae, "Gan vocoder: Multi-resolution discriminator is all you need;" 2021. [19] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss Mohammad Norouzi, and William Chan, Wavegrad: Estimating gradients for waveform generation; 2020. [20] Romain Hennequin, Anis Khlif, Felix Voituret; and Manuel Moussallam_ "Spleeter: fast and efficient music source separation tool with pre-trained models Journal of Open Source Software, vol. 5, pp. 2154, 06 2020. [21] Alexandre Defossez, Nicolas Usunier; Leon Bottou, and Francis Bach; "Music source separation in the waveform domain; 2021. [22] Yi Luo and Nima Mesgarani_ "Conv-tasnet: Surpassing ideal time-frequency magnitude masking for speech separation;' IEEEIACM TASLP vol. 27, no. &, pp. 1256-1266, Aug 2019. [23] Jonathan Ho, Ajay Jain, and Pieter Abbeel, "Denoising diffusion probabilistic models;' 2020. [24] Vincent Lostanlen and Carmine-Emanuele Cella, Deep convolutional networks on the pitch spiral for musical instrument recognition;' 2017. [25] Chandan K Reddy, Harishchandra Dubey; Kazuhito Koishida, Arun Nair; Vishak Gopal_ Ross Cutler Sebastian Braun, Hannes Gamper; Robert Aichner; and Sriram Srinivasan ~Interspeech 2021 deep noise suppression challenge;' 2021 _ [26] J. Eaton, N. D. Gaubitch, A_ H. Moore, and P A. Naylor; "The ace challenge corpus description and performance evaluation,'in 2015 IEEE WASPAA, 2015, pp. 1-5_ [27] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky, Instance normalization: The missing ingredient for fast stylization, ; 2017. [28] Kaiming He, Xiangyu Zhang, Shaoqing Ren; and Jian Sun_ Deep residual learning for image recognition;' in 2016 IEEE CVPR, 2016, pp. 770-778_ [29] Yao-Yuan Yang; Moto Hira, Zhaoheng Ni, Anjali Chourdia, Artyom Astafurov, Caroline Chen, Ching-Feng Yeh; Christian Puhrsch_ David Pollack; Dmitriy Genzel, Donny Greenberg, Edward Z Yang; Jason Lian, Jay Mahadeokar; Jeff Hwang; Ji Chen, Peter Goldsborough, Prabhat Roy; Sean Narenthiran, Shinji Watanabe, Soumith Chintala, Vincent Quenneville-Belair; and Yangyang Shi, Torchaudio: Building blocks for audio and speech processing; arXiv preprint arXiv:2110. 15018, 2021. [30] YHu and Philipos C. Loizou; valuation of objective quality measures for speech enhancement_ IEEE TASLP vol: 16, pp_ 229-238,. 2008. [31] Ryuichi Yamamoto, Eunwoo Song; and Jae-Min Kim; Parallel wavegan: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram;" 2020_ [32] Kevin Kilgour; Mauricio Zuluaga; Dominik Roblek; and Matthew Sharifi, Frechet audio distance: A metric for evaluating music enhancement algorithms,"2019. 1] Jiaqi Su, Zeyu Jin, and Adam Finkelstein, Hifi-gan: Highfidelity denoising and dereverberation based o speech deep features in adversarial networks 2020. [2] Alexandre Defossez, Gabriel Synnaeve, and Yossi Adi, "Real time speech enhancement in the waveform domain, Interspeech2020, 2020. [3] Phillip Isola, Jun-Yan Zhu; Tinghui Zhou, and Alexei A Efros, ~Image-to-image translation with conditional adversarial networks_ 2018_ [4] Zhifeng Kong; Wei Ping; Jiaji Huang, Kexin Zhao, and Bryan Catanzaro, Diffwave: versatile diffusion model for audio synthesis,'2021_ [5] Jiaqi Su, Yunyun Wang. Adam Finkelstein; and Zeyu Jin Bandwidth extension is all you need, in ICASSP 2021-2021. IEEE, 2021, pp. 696-700. [6] Pranay Manocha, Zeyu Jin, Richard Zhang; and Adam Finkelstein, "Cdpam: Contrastive learning for perceptual audio similarity;" in ICASSP 2021-2021. IEEE, 2021, pp. 196-200. [7] Chandan KA Reddy, Vishak Gopal, and Ross Cutler; Dnsmos: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors in ICASSP 2021-2021. IEEE, 2021, pp. 6493-6497. [8] P Scalart and J. V. Filho, "Speech enhancement based on priori signal to noise estimation;" in 1996 IEEE ICASSP Proceedings, 1996, vol. 2, pp. 629-632 vol: 2 [9] Hideaki Kagami, Hirokazu Kameoka; and Masahiro Yukawa, "Joint separation and dereverberation of reverberant mixtures with determined multichannel non-negative matrix factorization,'in 2018 IEEE ICASSP, 2018, pp. 31-35_ [10] Kun Han, Yuxuan Wang, DeLiang Wang; William $. Woods, Ivo Merks, and Tao Zhang; Learning spectral mapping for speech dereverberation and denoising;' IEEEIACM TASLP, vol. 23, no. 6, pp. 982-992, 2015. [11] Donald S_ Williamson and DeLiang Wang; "Speech dereverberation and denoising using complex ratio masks;" in 2017 IEEE ICASSP, 2017, pp. 5590-5594. [12] Daniel Michelsanti and Zheng-Hua Tan; "Conditional generative adversarial networks for speech enhancement and noiserobust speaker verification,"Interspeech 2017, Aug 2017. [13] Santiago Pascual, Joan Serra, and Antonio Bonafonte, ~Towards generalized speech enhancement with generative adversarial networks,'2019. [14] D. Griffin and Jae Lim; ~Signal estimation from modified short-time fourier transform IEEE Transactions on Acoustics, Speech, and Signal Processing, vol_ 32, no. 2, Pp. 236 243, 1984 [15] Adam Polyak; Lior Wolf;, Yossi Adi, Ori Kabeli, and Yaniv Taigman, High fidelity speech regeneration with application to speech enhancement;' in ICASSP 2021-2021. IEEE, 2021, pp. 7143-7147 [16] Jiaqi Su, Zeyu Jin, and Adam Finkelstein; "Hifi-gan-2: studioquality speech enhancement via generative adversarial networks conditioned on acoustic features;' in 2015 IEEE WASPAA, 2021_ | 1 | 5.3% |
| Cog Video: Large-scale Pretraining for Text-to-Video Generation via Transformers Wenyi Hongi Ming Ding' Wendi Zheng Xinghan Liut Jie Tangtt tTsinghua University IBAAI {hongwy18@mails dm18@mails jietang@mail}. tsinghua. edu. cn 8 2 8 3 ] 1 Abstract Large-scale pretrained transformers have created milestones in text (GPT-3) and text-to-image (DALL-E and CogView) generation: Its application to video generation is still facing many challenges: The potential huge computation cost makes the training from scratch unaffordable; The scarcity and weak relevance of text-video datasets hinder the model understanding complex movement semantics. In this work, we present 9B-parameter transformer Cog Video, trained by inheriting a pretrained text-to-image model, Cog View2. We also propose multi-frame-rate hierarchical training strategy to better align text and video clips. As (probably) the first open-source large-scale pretrained text-to-video model, Cog' Video outperforms all publicly available models at a large margin in machine and human evaluations A man is skiing: (running an tnet beach in the late afternoon_ A couple are having dinner: Nietrooisa A lion is Idrinking water: | A girl is dancing: Anime Figure I: Samples generated by Cog' Video. The actual text inputs are in Chinese. Each sample is 4-second clip of 32 frames, and here we sample 9 frames uniformly for display purposes More samples, models and codes will be available athttps : / /github. com THUDM/CogVideo, Equal contribution_ Preprint_ Under review. L Introduction Autoregressive transformers, eg DALL-E 183 and Cog View [5], have revolutionized text-to-image generation recently. It is natural to investigate the potential of autoregressive transformers on textto-video generation: Previous works followed this basic framework [35, /9, e. g: VideoGPT [361, verifying its superiority over GAN-based methods [4J[26], but are still far from satisfactory: One common challenge is that the generated video frames tend to gradually deviate from the text prompt, making the generated characters hard to perform the desired actions. Vanilla autoregressive models might be good at synthesizing videos with regular (e. g. straightly moving cars) Or random patterns (e. g: speaking by randomly moving lips), but fail on text prompt such as ta lion is drinking water"= The main difference between the two cases is that; in the former case the first frame already provides sufficient information for the subsequent changes, while in the latter the model has to precisely understand the action "drink in order to correctly generate the desired action the lion lifts the glass to its lip, drinks and then puts down the glass. Why do the autoregressive transformers well understand the text-image relations, but struggle to understand the text-action relations in videos? We hypothesize that the datasets and the way to utilize them are the main reasons_ First; it is possible t0 collect billions of high-quality text-image pairs from Internet [18], but the text-video data are more scarce. The largest annotated text-video dataset, VATEX [31/, has only 41, 250 videos. The retrieval-based text-video pairs, e. g: HowtolOOM [16], are weakly relevant and most of them only describe the scene without the temporal information: Second, the duration of videos varies lot. Previous models split the video into many clips of a fixed number of frames for training, which destroys the alignment between the text and its temporal counterparts in the video. If a ~drinking video is split into four individual clips of "holding glass' ~lifting' ~drinking" and "putting down with the same text "drinking" the model will be confused t0 learn the accurate meaning of drinking: Present Work. Here we present a large-scale pretrained text-to-video generative model, Cog Video, which is of 9. 4 billion parameters and trained on 5. 4 million text-video pairs_ We build Cog Video based on a pretrained text-to-image model, Cog View2 [6/, in order to inherit the knowledge learned from the text-image pretraining: To ensure the alignment between text and its temporal counterparts in the video, we propose the multi-frame-rate hierarchical training: The flexibility of the textual condition makes it possible to simply prepend a piece of text describing the frame rate to the original text prompt for modeling different frame rates To keep the text-video alignment; we choose a proper frame rate description to include the complete action in each training sample. The frame rate token also controls the intensity of the changes throughout continuous frames in generation. Specifically, we train a sequential generation model and frame interpolation model The former model generates key frames according to the text, and the latter recursively fill the middle frames by varying the frame rates t0 make the video coherent: As shown in Figure[] Cog' Video can generate high-resolution (480x480) videos Human evaluation demonstrates that Cog Video outperforms all publicly available models at a large margin. Our main contributions can be concluded as follows: We present Cog' Video, which is the largest and the first open-source pretrained transformer for text-to-video generation in the general domain: Cog'Video elegantly and efficiently finetunes a pretrained text-to-image generative model for text-to-image generation, avoiding the expensive full pretraining from scratch_ We propose the multi-frame-rate hierarchical training to better align text-clip pairs, which significantly improves the generation accuracy; in particular for movements of complex semantics. This training strategy endows Cog Video with the capacity of controlling the intensity of changes during the generation. 2 Related Work 2. 1 Video Generation Video generation is a long-standing research topic. Most previous works focus on the next-frame prediction task forecasting the future frames based on the first video frame. Early works, e. g CDNA [8p and PredRNN [32], leverage deterministic methods to directly predict the next frame via CNNs or RNNs. However; these deterministic models are unable to capture the stochastic temporal patterns and synthesize coherent complex scenes. Generative models, especially Generative Adversarial Networks [10] (GANs), begin to dominate the area as they can perform unconditional o class-conditional video synthesis without the first frames. VGAN [30] is the first one t0 use GAN for video generation: It decomposes video to a static background and moving foreground, and then generates them with 2D and 3D convolutional networks respectively. TGANI19 proposes to separately generate the temporal latent variables and spatial information, and MoCoGAN [26] similarly decomposes the latent space into context and motion subspaces DIGAN [37| applies implicit neural representations for video encoding: Recently; text-to-video generation emerges as a promis Ising direction: The framework of VQVAE [28 and autoregressive transformers 29, [1] quickly becomes the mainstream method 34185, 9]. Ho et al. 11 proposes video diffusion model along with gradient method recently for text-to-video generation. The previous methods are basically trained on specific dataset; e. g: UCF-101 [22], making the trained model domain-specific. Moreover; most of these models are not publicly available 2. 2 Autoregressive Transformer Recent years have witnessed the autoregressive transformer emerging as a powerful generative model. The autoregressive models become the most prevalent framework for text generation [23] With its prominent capacity of fitting; transformer [29] gradually becomes the standard neural structure for text generation One milestone is GPT-3 [1J. In computer vision, van den Oord et al_ [28 first proposes to train VQVAE to compress the image into sequence of tokens from learned dictionary, which can be efficiently handled by autoregressive models. VQ-GAN [7 | learns a more semantic-aware dictionary for unconditional image generation: In the text-to-image generation; pretrained autoregressive transformers such as DALL-E [18] and CogView [5] have shown superiority in open-domain image generation_ Besides the pure GPT-style generation, Cog View2 [6] proposes a new language model CogLM for infilling in the image generation. Recent autoregressive transformers [17] B36, [34} B5[ have also shown their superiority in video generation. Among them, GODIVA [347 and NUWA [351 focus on the open-domain text-to-video generation. However; they simply generate frames or frame blocks one by one in chronological order; and may suffer from poor text-video alignment (Cf: s[ 3 Method In this section, we first introduce multi-frame-rate hierarchical training to better align text and video semantics in $. 4 and then illustrate an efficient method dual-channel attention t0 inherit the knowledge in pretrained text-image models for video generation in $B. 2] To overcome the large memory and time overhead caused by the large model and long sequence, we refer to Swin Attention [14] and extend it to autoregressive video generation in $B3] 3. 1 Multi-frame-rate Hierarchical Training Here we present the multi-frame-rate hierarchical training and generation. We follow the framework of VQVAE [28] and first tokenize each frame into image tokens_ Each training sample consists of 5 frames of tokens, but our training method differs in the construction of training sequences and generation process. Training: The key design is to add a frame-rate token to the text and sample frames at this frame rate to compose a fixed-length training sequence. The motivations are two folds: Directly separating the long video into clips at fixed frame rate often leads to semantic mismatching: We still use the full text but the truncated clip might oly contain incomplete action (2) The adjacent frames are usually very similar: A giant change over the previous frame will probably incur a large loss This will lead the models less inclined to explore the long-range correlation because simply copying the previous frame acts like a shortcut: Input Text: Input Frames Image Tokenizer A lion is drinking water: RWFETIBjK 2 Discretize 20*20-400 image tokens per frame 5 frames Text tokenization Frame Rate Flatten Text Frame-1 Frame-2 Frame-3 Frame-Frame-5 Transformer (Stage 1: Sequential Generation) Interpolate frames Sequence IB] FrameSequence 2 Frame-3 Frame-4 Frame Rate Text Frame-2 Frame-5 Transformer (Stage 2: Recursive Interpolation) Figure 2: Multi-frame-rate hierarchical generation framework in Cog'Video. Input sequence includes frame rate, text, frame tokens. [B] (Begin-of-image) is separator token, inherited from Cog View2. In stage 1, Ts frames are generated sequentially on condition of frame rate and text Then in stage 2, generated frames are re-input as bidirectional attention regions to recursively interpolate frames_ Frame rate can be adjusted during both stages. Bidirectional attention regions are highlighted in blue and unidirectional regions are highlighted in green Therefore; in each training sample, we want the text and the frames to match as possible. We predefined series of frame rates, and select the lowest frame rate for each text-video pair; as long as we can sample at least 5 frames at this frame rate in the video Although the above method increases the alignment of text and video, the generation at a low frame rate could be incoherent. We train another frame interpolation model t0 insert transition frames t0 the generated samples of the sequential generation model Thanks to the generality of CogLM [6], the two models can share the same structure and training process only with different attention masks_ Generation. The multi-frame-rate hierarchical generation is recursive process, illustrated in Figurep] Specifically, the generation pipeline consists of a sequential generation stage and a recursive interpolation stage: Sequentially generate Ts key frames based on low frame rate and text: The input sequence is [{Frame Rate}{Text} [B] {Framel} {Frame Ts}]. In practice, we always set Ts 5 and the minimum sampling frame rate to 1 fps. (2) Recursively interpolate frames based on the text, frame rate and known frames. In each round of interpolation, we split generated frames into multiple % ]-frame blocks overlapping at the beginning and the end, and interpolate a frame between the successive frames in each block: The input sequence is [{Frame Rate} {Text} [B] {Frame1} {Frame Ts}] where Frame 2i(i = 1, 2, 2 J) are to be autoregressively generated. By recursively halfing {Frame Rate}, we can conduct finer and finer interpolation to generate videos of many frames. The effect of CogLM. Tasks such as frame interpolation rely heavily on bidirectional information_ However; most previous works use GPT [34, B6] [5], which is unidirectional. To be aware of the bidirectional context; we adopt Cross-Modal General Language Model (CogLM) proposed in [6] which unites bidirectional context-aware mask prediction and autoregressive generation by dividing tokens into unidirectional and bidirectional attention regions While bidirectional regions can attend to all bidirectional regions, unidirectional regions can attend to all bidirectional regions and previous unidirectional regions As shown inp] (1) all frames in stage and the Znd, 4th frames in stage 2 are in the unidirectional region; (2) {Frame Rate}, {Text} and all other frames belong to the bidirectional region. In this way, bidirectional attention context is fully exploited in text and given frames without interfering with auto-regressive frame prediction. 3. 2 Dual-channel Attention Large-scale pretraining usually demands a large dataset. For the Xout open-domain text-to-video generation;, ideally we need the dataset Addition to cover sufficient text-video pairs to infer both spatial and temporal correlation between video and text: However; to collect Layer Norm high-quality text-video pairs is often difficult; expensive and timeDual-channel Attention consuming: Addition 10 A natural idea is to make use of the image data to facilitate the Attention-base Attention-plus learning of spatial semantics. Video Diffusion Model 11 and Spatial Channel) (Temporal Channel) NUWA [35| try to add text-image pairs into text-video training; which achieves better results on multiple metrics_ However; as Layer Norm for training a video-only generation model, adding image data Xin will significantly increase training costs, especially in large-scale pretraining scenarios Figure 3: The dual-channel attention block: We initialize In this paper; we propose to leverage pretrained image generation the Attention-plus the same as models instead of image data Pretrained text-to-image models, Attention-base so that the model e. g: Cog'View2 [6], already have good command of the textbehaves exactly the same as image relations. The coverage of the dataset to train these models CogView2 when it is initialized. is also larger than that of videos_ The proposed technique is dual-channel attention, where we only add a new spatial-temporal attention channel to the pretrained Cog View2 [6] at each transformer layer: All the parameters in the Cog View2 are frozen in the training, and only the parameters in the newly added attention layer (See the attentionplus in FigureB] are trainable_ We denote the original attention block in CogView2 as attention-base Here we also emphasize that directly finetuning CogView2 for text-to-video generation cannot well inherit the knowledge, because the temporal attention follows a different attention pattern and quickly ruins the pretrained weights during the initial phase of training with large gradients. Specifically, the dual-channel attention block with Sandwich-LN [5] can be computed as x =0 attention-base(LayerNorm (Tin _ + (1 a). attention-plus(LayerNorm(Tin) ), (2) Tout Tin + LayerNorm(x). The mixture factor & is a vector EUR (0, 1)&, where d is the hidden size of the input feature TinTo restrict the range of a within (0, 1), we reparameterize it as & sigmoid(a) EUR (0, 1)8 where a EUR Rd is a learnable parameter: The attention-plus block has the same shape of parameters as the normal multi-head attention block, attention-base, but differs in the procedure of computation as follows. In our training, we tried two kinds of attention, 3D local attention and 3D Swin [14] attention for attention-plus block: Here we depict the 3D local attention, and the latter is a natural replacement introduced in sectionB3 In 3D local attention, the receptive field (RF) for the token at (t, EUR, y) (where (t, EUR, y) corresponds to the coordination along time, height and width), is a 3D block with extent lt, lz, ly EUR Nt: RF(t. f, y) = {(k, i, j) Ix i] < 1, ly jl < ly;lt _ kl < lt; (k, i, j) $ Mask(t, z, y)}; (3) where Maskat. f '9) represents an attention mask for token (t,., y). In the sequential generation model (Stage 1), the Mask ensures the auto-regressive order; In the interpolation model (Stage 2), the Mask is designed as in as CogLM [6] to make the known frames visible to all the frames_ It is worth noting that two channels are fused and share the same FFN in each layer; because FFN is a module ofheavy parameters containing much vision knowledge. Due to the similarity between images and videos, bringing its knowledge to the temporal channel will facilitate video modeling: Finally, sharing FFN can reduce parameters, thus speeding up training and reducing memory overhead: 5 3. 3 Shifted Window Attention in Auto-regressive Generation To further alleviate the large time and memory overhead in the temporal channel during training and inference, we refer t0 Swin Attention [14]. The original Swin attention is only applied to non-autoregressive scenarios, We extend it t0 the autoregressive and temporal scenario by applying an auto-regressive attention mask in the shifted windows An interesting finding is that, the Swin attention provide a chance for parallel generation in faraway regions of different frames, which further accelerates the auto-regressive genera tion: The dependence of the generation of a specific token relies on Auto-regressive mask: token can only attend to previous frames Or tOt-i+1 t-i+2 kens before itself in the current frame_ Shifted window. Only tokens within Figure 4: In 3D autoregressive swin attention (winthe distance of window size in both dow size 2 X 2 as an example), the token in the red width and height dimensions can be box can only attend to (either directly or indirectly) directly attended to. the yellow O green tokens The gray tokens in the i-th frame and the token in the red box can be As shown in Figurel] we can start generating generated in parallel. parts of the tokens in the following frames before finishing the generation of all the previous frames they can work in parallel. Suppose X, Y is the height and width of each frame, and Ar, Ay are the height and width of shifted window. For two tokens at (t1, 11, 91) and (t2, *2, Y2), +1 t2, the latter cannot attend to the former either directly O indirectly if (81 T2)Y + (y1 Y2) > (t2 t1 +1)(AcY + Ay); which means that the i-th token in the t-th frame can be generated with the (i A, Y _ Ay)-th token in the (t + 1)-th frame in parallel. In this way; we can generate [z44w, tokens in parallel at most_ thus greatly enhance parallelism and accelerate inference compared to auto-regressive with standard attention which can only generate one token at a time_ Training Based on the methods above, the training details of Cog'Video are listed as follows: Model: The backbone of Cog' Video in both stages is Transformer with dual-channel attention The Transformer has 48 layers, with hidden size of 3, 072 in each attention channel, 48 attention heads and 9. 4 billion parameters in total. Among them, 6 billion parameters are fixed to CogView? s parameters, which include Position-wise Feed-Forward Networks (FFN), the spatial channel of dualchannel attention, first frame '$ positional embeddings and all image and text vocabulary embeddings. The specific implementation of Transformer structure is almost identical to CogView [5| such as using Sandwich LayerNorm and PB-Relax to stabilize training. Shifted CogLM attention window is adopted in recursive interpolation model with window size 10 x 10. Dataset: We pretrain our model on dataset of 5. 4 million captioned videos with a spatial resolution of 160 x 160 (can be upsampled to 480 X 480 by Cog View2). For the sequential generation model (Stage 1), we adjust the frame rate in each sample to accommodate the whole video, while the minimum frame rate is set t0 [ fps. For the recursive interpolation model (Stage 2), we split videos into clips of different lengths to accommodate prediction on multiple frame rates including 2, 4, 8 fps. Pretraining: The sequence lengths in both stages are 2, 065, consisting of 64 text tokens, 5 (frames) 400 (per frame) image tokens, and seperator token. Both text and images are tokenized with icetkH]The parameters are updated by Adam with max learning rate = 2 x 10 81 0. 9, 82 0. 95, weight decay = 1 10-2. See Appendix for pretraining details. ~https: '/github _ com/ THUDM_ icetk Table I: (Left) Video generation performance on UCF-1O1. Class labels are used as the text inputs means that the model is only trained on the training split of UCF-1O1. (Right) Video generation performance on Kinetics-600_ The metrics are based on the 16-frame generated videos priming on first 5 frames, following settings of Rakhimov et al 17|**means that the ground truth used in FVD testing is the reconstruction result of the tokenizer: Method IS FVD 24. 69 27. 38 28. 87 1209 32. 36 838 29. 71 655 32. 70 577 79. 28 332 VideoGPTI36 DVD-GANI4 TGANv2[20p MoCoGAN-HDI24 DIGAN[37 DIGAN[37 TATS-basel9l Cog"Video (Ours Cog' Video (Ours_ Method Latent Video Tranformerl 17 Video Transformer[33 DVD-GAN-FPI4_ TriVD-GAN-FP[15] Cog Video (Ours, Cog Video (Ours FVD(L) 224. 73 170 69. 15 25. 74 109. 23 59. 55 50. 46 626 545 5 Experiments 5. 1 Machine Evaluation Machine evaluation is conducted on two popular benchmarks for video generation; ie, UCFIOI [22] and Kinetics-600 [3]. Following Rakhimov et al. [17|, Yu et al. [37/, we use Frechet Video Distance (FVD) [27E and Inception score (IS [21] as metrics in the evaluation_ FVD is calculated based on I3D modell2| trained on Kinetics-400, and IS is based on C3D model [25_ which was first trained on the Sports-IM dataset 12 and then finetuned on the UCFIOI dataset_ Our evaluation code is the same as the official TGAN-v2 implementationz] UCF-IOI is a human action dataset consisting of 13, 320 videos annotated with 101 action classes. Due to the gaps of image style and frame rate between Cog' Video 's training set and UCF-101, we use class labels as the input text and finetune Cog'Video on the whole dataset for 10, 000 iterations with a batch size of 192. During inference, we generate samples of various classes according to the class distribution: FVD and IS are evaluated over 2, 048 and 10, 000 samples respectively, following Yu et al. [37]. Results are shown in Table[](Left). Kinetics-600 contains 600 classes of human action videos, with roughly 350, 000 train and 50, C 000 test videos in total. We use the action category as input text, and finetune Cog Video on the training set for 12, 000 iterations with batch size of 640. Following the setup of Weissenborn et al. [33], Rakhimov et al. [17], we center-crop and downsample each frame to 64X64 to measure the FVD of the model_ Results are shown in Table[@](Right). 5. 2 Human Evaluation To further evaluate Cog'Video, we invite 90 anonymous evaluators to rate for Cog' Video and other open source baselines including GAN-based model TGANv2 [20] and GPT-based model VideoGPT 36]. 30 classes in UCFIOL are randomly picked as text conditions, and several aspects are rated (See Appendix for details). For VideoGPT; we use the official unconditional pretrained model to generate samples. For TGANv2, we use the official source code to train an unconditional generation model under the same setting as that in Saito et al. [20[ To assign unconditionally generated samples into corresponding categories, we choose TSM [13 as the action recognition model for postclassification. We only keep the samples whose likelihood to a certain class is at least 80%. Results in Figure[lshow that Cog Video significantly outperforms baselines on multiple important aspects including frame texture, motion realism and semantic relevance, and achieves the top score by the overall quality It can be seen that 49. 53% evaluators choose Cog Video as the best method, and only 15. 42% and 5. 6% favor VideoGPT and TGANv2, respectively: https: Igithub com_ [pfnet research/tgan2 https : '/github _ com/ wilsonlyan/VideoGPT TGAN videoGPT Cog video IStage Cogideo Ground Truth TGe VidecGpt Cuyvicec St10" Convicen Ground Truch Lonvide? Staje videoGPt 015, 42 5. 61 TGAN 4957 CogVideo Ovrnl Scnre Frame exturz Kotion Kejiism bemantic evonce Human preference The percentage of being chosen as the best:. Overall scores (1-10) for each method: Scores (1-5) on three important aspects_ Figure 5: Human evaluation results. ~Cog Video IStage" refers to the method in ablation study, which only generates videos sequentially with the CogVideo's Stage 1 to the desired number of frames: Table 2: Ablation study on 5, 000-sample subset of Kinetics-600 s testset: FVD is evaluated on generated H-frame samples priming on 5 frames and the recovered ground-truth by the image tokenizer: The setting column indicates the difference between each method and CogVideo. Models of each setting are trained on Kinetics-600 trainset for 11, 000 iterations with a batch size of 160. Method Setting None FVD 108. 27 Cog Video 1-stage Generation( Noverlap 1-stage Generation( Noverlap 2) Initialized with Cog View2 Randomly Initialized hierarchical hierarchical 137. 13 120. 82 124. 92 166. 13 Pretrain Pretrain Cog' View2 5. 3 Ablation Study To verify the effectiveness of hierarchical multi-frame-rate generation and incorporating Cog View2, we conduct ablation studies on Kinetics-600 and UCF-1O1 datasets. We will first briefly introduce the compared methods and analyze the quantitative results in $[5. 3 Tand qualitative results in $653. 2 Hierarchical multi-frame-rate generation. In comparison with CogVideo; we finetune I-stage video generation model on Kinetics-600 from the sequential generation model in CogVideo, which generates long videos by sliding windows In each window, We generate the rest frames based on Noverlap previous known frames. Larger Noverlap means more previous frames can be utilized during the inference, but will increase time overhead: Dual-channel attention with Cog View2's weights. To highlight the effectiveness of our finetuning strategy, we additionally finetune (1) randomly initialized model, (2) model incorporating Cog' View2'$ weights but leaving the temporal channel unfixed (equivalent t0 CogVideo without pretraining O videos) on Kinetics-600 for comparison 5. 3. 1 Quantitative Evaluation Cogvideolfinetune) CogView Initialized Randomly Initialized All aforementioned models have been trained for 11, 000 5. 75 iterations with 5, 50 batch size of 160. Quantitative results are 5. 25 shown in Tablep We can see that the hierarchical method 8 is clearly superior to the 1-stage generation with differ1 4. 75 4. 50 ent Noverlap; and the model initialized with Cog'View2 s 4. 25 weights has lower FVD than the randomly initialized one: 3. 75 Figure plots the training loss curve of (1) finetuning 3. 50 Cog' Video; (2) training model from random initialization; 3, 25 (3) training model initialized with Cog View2 and partially thousand iteration fixed_ Figure 6: Training loss in ablation study: We can see that Cog View2 endows the model with 8 Randomly Initialized Input Text: Lunge Given frames: Finetuned CogVideo, hierarchical generation Finetuned CogVideo, 1-Stage Noverlap (b) Initialized with CogView2 Finetuned CogVideo, 1-Stage Noverlap Figure 7: Video samples in ablation study, which are generated priming on the class label and first 5 frames in Kinetics-600. All samples are downsampled by extracting one in every three frames for display purposes. (a) Use finetuned Cog Video to hierarchically generate samples. (b) Train a model on Kinetics-600 which is initialized as and partially fixed to Cog' View2, and hierarchically generate samples (c) Train a model on Kinetics-600 which is randomly initialized, and hierarchically generate samples. (d)(e) Use finetuned Cog Video to generate frames in stage with different Noverlap_ good initialization point from which the loss can decrease faster: Moreover; fixing part of the parameters reduces the time and memory cost: 5. 3. 2 Qualitative Evaluation Qualitative comparison is shown in Figure[ While the model trained from random initialization tends to produce irrational deformation, the model incorporating CogView2 is able to generate realistic objects, and the hierarchical generation performs better on content consistency and motion realism We also conduct human evaluation between I-stage and hierarchical video generation model under the same setting as in $6. 2 As shown in Figure/5] the hierarchical model, i. e. Cog'Video, outperforms the 1-stage model on semantic relevance, motion realism as well as texture quality: This is probably because the 1-stage model cannot estimate a proper intensity of change from the previous frames in the window, as shown in Figure[Kd)(e) 6 Conclusion We present Cog' Video, to the best of our knowledge, the largest and the first open-source pretrained transformer for text-to-video generation in general domain. CogVideo is also the first attempt to efficiently leverage the pretrained text-to-image generative model to the text-to-video generation model without hurting its image generation capacity: With the proposed multi-frame-rate hierarchical training framework; CogVideo is endowed with better understanding of text-video relations and abilities to control the intensity of changes during generation. We extend swin attention to CogLM; which achieves acceleration in both training and inference. There are still some limitations in Cog'Video, e. g restriction on the length of the input sequence still exists due to the large scale of the model and limitation of GPU memory; and we leave them for future work: Broader Impact: This paper aims to advance the open-domain text-to-video generation, which will ease the effort of short video and digital art creation The efficient training method transfers knowledge from text-to-image models to text-to-video models, which helps avoid training from scratch, and thus reduces energy consumption and carbon emission_ A negative impact is the risk of misinformation_ To alleviate it; we can train an additional classifier t0 discriminate the fakes. We believe the benefits outweigh the downsides. Acknowledgments and Disclosure of Funding We would like t0 thank Zhao Xue, Shuai Zhao, Sha Yuan for their help in data collection, Weidong Guo, Fengyu Rao, Zhaoyang Zeng; Mingkang Tang for their useful discussion, Hanxiao Qu for maintaining the machines and the computational resources supported by BAAL References T B. Brown B Mann_ N. Ryder; M Subbiah; J. Kaplan, P Dhariwal, A. Neelakantan, P Shyam; G. Sastry, A_ Askell, et al. Language models are few-shot learners. arXiv preprint arXiv: 2005. 14165, 2020. [2] J. Carreira and A Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset_ In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. [3] J. Carreira; E. Noland, A_ Banki-Horvath, C. Hillier; and A. Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808. 01340, 2018. [4] A. Clark, J. Donahue, and K. Simonyan Adversarial video generation on complex datasets. arXiv preprint arXiv: 1907. 06571, 2019. [5] M. Ding, Z Yang, W: Hong; W. Zheng; C. Zhou; D. Yin, J. Lin, X Zou; Z. Shao, H Yang; et al_ Cogview: Mastering text-to-image generation via transformers_ Advances in Neural Information Processing Systems, 34, 2021. [6] M Ding, W: Zheng; W: Hong; and J. Tang: Cogview2: Faster and better text-to-image generation via hierarchical transformers_ arXiv preprint arXiv:2204. 14217, 2022. [7] P Esser; R Rombach; and B Ommer Taming transformers for high-resolution image synthesis arXiv preprint arXiv:2012. 09841, 2020. [8] C Finn, I Goodfellow, and S. Levine. Unsupervised learning for physical interaction through video prediction. Advances in neural information processing systems; 29, 2016. [9] S. Ge_ T Hayes, H: Yang; X. Yin, G. Pang; D. Jacobs, J. -B_ Huang; and D_ Parikh_ Long video generation with time-agnostic vqgan and time-sensitive transformer: arXiv preprint arXiv:2204. 03638, 2022_ [10] I. J. Goodfellow; J. Pouget-Abadie, M. Mirza, B. Xu; D. Warde-Farley, S. Ozair, A. Courville_ and Y Bengio. Generative adversarial networks arXiv preprint arXiv:1406. 2661, 2014. J. Ho, T Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet: Video diffusion models_ arXiv preprint arXiv:2204. 03458, 2022. [12] A. Karpathy, G. Toderici, S. Shetty, T: Leung, R Sukthankar; and L Fei-Fei: Large-scale video classification with convolutional neural networks_ In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725-1732, 2014. [13] J. Lin, C. Gan, and S. Han. Tsm: Temporal shift module for efficient video understanding_ In Proceedings of the IEEEICVF International Conference on Computer Vision, pages 7083-7093, 2019. [14] Z Liu; Y Lin, Y Cao, H. Hu, Y. Wei, Z Zhang, S. Lin, and B_ Guo. Swin transformer: Hierarchical vision transformer using shifted windows_ In Proceedings of the IEEEICVF International Conference on Computer Vision, pages 10012-10022, 2021. [15] P Luc, A Clark, S_ Dieleman D. d_ L: Casas, Y Doron, A Cassirer; and K Simonyan_ Transformation-based adversarial video prediction on large-scale data_ arXiv preprint arXiv:2003. 04035, 2020. [16] A. Miech; D Zhukov; J-B. Alayrac, M. Tapaswi, I. Laptev, and J. Sivic. HowtolOOm: Learning text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 10 [17] R Rakhimov, D. Volkhonskiy, A. Artemov; D Zorin, and E. Burnaev. Latent video transformer: arXiv preprint arXiv:2006. 10704, 2020. [18] A Ramesh; M. Pavlov, G. Goh; S. Gray, C. Voss, A. Radford, M. Chen, and I. Sutskever: Zero-shot text-to-image generation: arXiv preprint arXiv:2102. 12092, 2021_ [19] M. Saito, E. Matsumoto, and S_ Saito_ Temporal generative adversarial nets with singular value clipping: In Proceedings of the IEEE international conference on computer vision, pages 2830-2839, 2017. [20] M. Saito, S. Saito, M. Koyama, and S. Kobayashi. Train sparsely, generate densely: Memoryefficient unsupervised training f high-resolution temporal gan_ International Journal of Computer Vision, 128(10). 2586-2606, 2020. [21] T Salimans I. Goodfellow, W Zaremba, V. Cheung, A. Radford, and X. Chen_ Improved techniques for training gans. In Proceedings of the 3Oth International Conference on Neural Information Processing Systems, pages 2234-2242, 2016. [22] K Soomro, A. R Zamir; and M. Shah: UcflOl: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv: 1212. 0402, 2012. [23] I Sutskever; J. Martens, and G. Hinton. Generating text with recurrent neural networks In ICML' 11, page 1017-10024, 2011. [24] Y Tian, J. Ren, M. Chai, K. Olszewski, X Peng; D. N. Metaxas, and S. Tulyakov good image generator is what you need forhigh-resolution video synthesis. arXiv preprint arXiv:2104. 15069, 2021. [25] D: Tran, L. Bourdev, R Fergus, L. Torresani, and M. Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. [26] S. Tulyakov; M-Y Liu, X Yang; and J. Kautz Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. [27] T Unterthiner; S_ van Steenkiste, K Kurach_ R. Marinier; M Michalski, and $. Gelly: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv: 1812. 01717, 2018. [28] A_ van den Oord, 0. Vinyals, and K Kavukcuoglu_ Neural discrete representation learning: In Proceedings of the 31st International Conference on Neural Information Processing Systems; pages 6309-6318, 2017. [29] As Vaswani N. Shazeer; N. Parmar; J Uszkoreit; L. Jones, A_ N. Gomez, L Kaiser; and [. Polosukhin. Attention is all you need_ arXiv preprint arXiv: 1706. 03762, 2017. [30] C. Vondrick, H. Pirsiavash; and A_ Torralba. Generating videos with scene dynamics Advances in neural information processing syStems, 29, 2016. [31] X Wang, J. Wu; J. Chen, L Li; Y-F Wang; and W. Y Wang: Vatex: large-scale, highquality multilingual dataset for video-and-language research: In Proceedings of the IEEEICVF International Conference on Computer Vision, pages 4581-4591, 2019. [32] Y. Wang; M Long, J. Wang, Z. Gao, and P S. Yu_ Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms Advances in neural information processing systems; 30, 2017. [33] D. Weissenborn; 0. Tackstrom; and J. Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv: 1906. 02634, 2019. [34] C. Wu, L. Huang, Q. Zhang; B. Li, L. Ji, F Yang; G. Sapiro, and N. Duan: Godiva: Generating open-domain videos from natural descriptions arXiv preprint arXiv:2104. 14806, 2021. [35] C. Wu; J. Liang; L. Ji, F Yang; Y Fang, D_ Jiang, and N. Duan_ Nl uwa: Visual synthesis pre-training for neural visual world creation. arXiv preprint arXiv:2111. 12417, 2021. [36] W: Yan, Y Zhang; P Abbeel, and A_ Srinivas. Videogpt: Video generation using Vq-vae and transformers arXiv preprint arXiv:2104. 10157, 2021. [37] S. Yu; J. Tack, S. Mo, H. Kim; J. Kim; J. -W. Ha, and J. Shin. Generating videos with dynamicsaware implicit generative adversarial networks arXiv preprint arXiv:2202. 10571, 2022. Attention Analysis To explore the attention mechanism of dual-channel attention; we visualize (1) the attention distribution in the temporal channel and (2) the mixture factor & controlling the ratio between the spatial and temporal channel in equation[ Figure[8]visualizes the distribution among frames and texts in sequential generation (Stage 1) with heat maps, where only 24 of 48 attention heads in 6 layers are shown for display purposes The attention patterns can be broadly classified into the following categories: Most of the attention is on the text: E. g: the attention heads in violet Most of the attention is on a certain frame. Eg: the attention heads in pink focus mainly on the previous frame; the attention heads in blue focus mainly on the first frame besides the text; the attention heads in yellow focus mostly on the frame itself: Attention is spread over several frames Eg: the attention heads in green Some attention heads exhibit a single pattern; while others may exhibit a mixture of them. Attention heads in the same layer tend to show similar patterns: In lower layers (e. g: layer 4, 12) the heads tend to allocate attention according to position, while in higher layers more attention is allocated to text (e. g. layer 44) O spread over multiple frames. One possible explanation is that there are more high-level features in higher layers such as video semantics, by which more frames and texts can interact with each other to make high-level feature analysis. It is worth noting that many heads in temporal channel do not allocate much attention to the frame itself, especially in higher layers, while attending to itself is important for inference This shows that the Cog Video performs a certain degree of decoupling in the analysis of temporal and spatial features While the spatial channel is in charge of feature analysis within the frame, the temporal channel can allocate more resources t0 explore relationships among different frames We further illustrate this perspective with FigureC which shows that features calculated by Cog View2 in the spatial channel are heavily relied on_ B Training Details Cog' Video consists of two models corresponding to two stages, i. e. sequential generation and recursive interpolation_ Both models have 7. 7 billion parameters while 6 billion of them are fixed to Cog View2, thus Cog Video has 9. 4 billion different parameters in total_ CogVideo is trained on a dataset of 5. 4 million captioned videos with spatial resolution of 160x 160 (can be upsampled to 480x480 by Cog View2) Each model is pretrained separately. The model in stage 1 is first pretrained for 76, 000 iterations on video clips with minimum frame rate of 0. 25 fps, then trained for 15, 000 iterations with minimum frame rate of 1 fps The model in stage 2 is pretrained for 78, 500 iterations with the frame rate of 2_ 4, and & fps. Both models are trained in FPI6 with a batch size of416, and optimized by Adam with max learning rate = 2 x 10-4, 81 0. 9_ 82 0. 95, weight decay = 1 x 10-2_ 12 Layer 4 Layer 12 Layer 20 Layer 28 Layer 36 Layer 44 Figure &: The attention distribution among frames and texts in sequential generation (Stage 1) Only 24of 48 attention heads in 6 layers are selected for display purposes. Each attention head is visualized with a heat map of size 5X6, where lighter color represents larger value The 5x5 block on the left indicates the sum of attention scores (after softmax between each pair of frames, and the rightmost column indicates the sum of the attention score of each frame t0 text That is to say, the grid in row i column j (j < 5) represents EreF yeF; attn_, y, and the grid in row column 6 represents ZreF, yeT attnc y, where Fi, T denotes the set of tokens in the i-th frame and text respectively; and attnz; y denotes the attention score of token x t0 y: Details about Human Evaluation In this section, we introduce more details about the human evaluation for measuring generation quality: The conduction of our human evaluation generally follows previous works including Ramesh et al. [18], Ding et al. [5] We randomly extract 30 classes from UCFIOL for video generation, using corresponding video samples in the dataset as ground truth items in the evaluation. Based on captions of selected classes, we generate video samples from models including TGANv2, VideoGPT; and our model_ Cog' Video. To further illustrate the effectiveness of hierarchical multi-frame-rate generation, we also include I-stage version of CogVideo model fine-tuned on Kinetics-600 which is described in $63] For TGANv2, we use the official source code to train an unconditional generation model under the same setting as that in Saito et al. [20]. For VideoGPT, we use the official unconditional pretrained model to generate samples To assign unconditionally generated samples into corresponding categories; we choose TSM[13/ as the action recognition model for a post-classification: We only keep the samples whose likelihood to a certain class is at least 80%. A randomly selected subset of samples is displayed in Figure [0] For each sample of the video mentioned above, we ask evaluators to give scores between and 5 ( 5 indicates the best while indicates the worst) from three aspects including frame texture, motion realism, and semantic relevance Then the evaluators are required to give a general score of quality for each sample between and 10, where higher score indicates better quality. After video samples 13 8 0. 4 12 14 32 Layer Figure 9: The scale factor & controlling the ratio between the spatial and temporal channel in equation[Jin dual-channel attention. Only a in half of the layers are shown for display reasons As & is a vector of dimension 3072, We show the mean and variance among all of its dimensions in this figure_ Skiing, {83 Biking; 95A(j4 CogVideo CogVideo Stage) 0, 'Y VideoGPT TGANv2 Groundtruth Figure 1O: A subset of human evaluation samples. The captions are randomly selected from UCF-101. The original samples are clips of 16 frames, which are downsampled to 4 frames uniformly for display purposes: from each caption are all evaluated, the evaluators are asked to select the best one from them We show snapshots of the evaluation website in Figure[ Throughout the process of human evaluation, we invited nearly 100 anonymous evaluators, while 90 of them completed the whole evaluation and were counted in the final results. None of the questions in the evaluation have any time limit We offer each evaluator 75 RMB as a reward for the evaluation_ Results of the human evaluation, including the average score and standard deviation for each group, have already been introduced in Figure[lin the main body: As ground truth samples take an absolute predominance in the best selection question; we have removed the part of ground truth samples in the selection pie plot for clearer model comparison. 14 Coayilee TAo CogVideo Task 0 Arudhaurick UFtnilsta EZEpEmi AAAIHm, T^Sin@0 6 @rrpRuaezix Be-%eil) jitmpistmtejiare s41 ~Wnt4a, TBM, Tratir5-Wnirbtatiq) Mnmzrab (4A246 Q@R) #0ade9ziybeias(g (1-@ipekraitStEz? 74* 5-6FDZ+0 #6E#haSI#8M2) Kaazrjad (WetTKAMiM) #Mariemt swoaatwo (1-Jfijizmu4, 1868X, 05542*T+* 10+07123180*4R*RSMw Ditmtlar erimRMiRZ 10-GARS) Figure 1: Snapshots of the evaluation website. 15 | 1 | 5.3% |
| Hierarchical Text-Conditional Image Generation with CLIP Latents Aditya Ramesh OpenAI aramesh@openai _ com Prafulla Dhariwal OpenAI prafullaCopenai com Alex NicholOpenAI alexCopenai. com 8 2 2 3 7 1 Casey Chu OpenAI caseyCopenai com Mark Chen OpenAI markCopenai com Abstract Contrastive models like CLIP have been shown to learn robust representations of images that capture both semantics and style. To leverage these representations for image generation; we propose a two-stage model: prior that generates CLIP image embedding given a text caption, and decoder that generates an image conditioned on the image embedding: We show that explicitly generating image representations improves image diversity with minimal loss in photorealism and caption similarity Our decoders conditioned on image representations can also produce variations of an image that preserve both its semantics and style, while varying the non-essential details absent from the image representation Moreover_ the joint embedding space of CLIP enables languageguided image manipulations in zero-shot fashion. We use diffusion models for the decoder and experiment with both autoregressive and diffusion models for the prior; finding that the latter are computationally more efficient and produce higher-quality samples. L Introduction Recent progress in computer vision has been driven by scaling models o large datasets of captioned images collected from the internet [10, [44, 60, [39, B1/[16]. Within this framework, CLIP [39] has emerged as successful representation learner for images CLIP embeddings have number of desirable properties: they are robust to image distribution shift; have impressive zero-shot capabilities, and have been fine-tuned to achieve state-of-the-art results o a wide variety of vision and language tasks [45] Concurrently, diffusion models [46][48, /25/ have emerged as promising generative modeling framework, pushing the state-of-the-art on image and video generation tasks [11 26, 24]. To achieve best results, diffusion models leverage a guidance technique 11 241 which improves sample fidelity (for images, photorealism) at the cost of sample diversity: In this work; we combine these two approaches for the problem of text-conditional image generation We first train a diffusion decoder to invert the CLIP image encoder. Our inverter is non-deterministic and can produce multiple images corresponding to given image embedding: The presence of an encoder and its approximate inverse (the decoder allows for capabilities beyond text-to-image translation. As in GAN inversion [62, [55] encoding and decoding an input image produces semantically similar output images (FigureB} We can also interpolate between input images by inverting interpolations of their image embeddings (Figure[H: However; one notable advantage of using the CLIP latent space is the ability to semantically modify images by moving in the direction of any encoded text vector (FigureE); whereas discovering these directions in GAN latent space involves Equal contribution Vibrant portrait painting of Salvador Dali with a robotic half face shiba inu wearing beret and black turtleneck a close up of a handpalm with leaves growing from it espresso machine that makes coffee from human souls, artstation panda mad scientist mixing sparkling chemicals artstation 1 2 corgi '$ head depicted as an explosion of a nebula propaganda poster depicting cat dressed french emperor napoleon holding a piece of cheese dolphin in an astronaut suit on saturn; artstation teddy bear on a skateboard in times square Figure 1: Selected 1024 X 1024 samples from production version of our model. CLIP objective img encoder "a corgi playing a flame text throwing encoder trumpet" decoder prior Figure 2: A high-level overview of unCLIP. Above the dotted line, we depict the CLIP training process, through which we learn a joint representation space for text and images. Below the dotted line, we depict our text-to-image generation process: CLIP text embedding is first fed to an autoregressive Or diffusion prior to produce an image embedding, and then this embedding is used to condition diffusion decoder which produces a final image. Note that the CLIP model is frozen during training of the prior and decoder luck and diligent manual examination: Furthermore, encoding and decoding images also provides US with a tool for observing which features of the image are recognized or disregarded by CLIP To obtain a full generative model of images, we combine the CLIP image embedding decoder with a prior model, which generates possible CLIP image embeddings from given text caption. We compare our text-to-image system with other systems such as DALL-E [40] and GLIDE [35], finding that our samples are comparable in quality to GLIDE; but with greater diversity in Our generations We also develop methods for training diffusion priors in latent space, and show that they achieve comparable performance to autoregressive priors, while being more compute-efficient We refer to ur full text-conditional image generation stack as unCLIP, since it generates images by inverting the CLIP image encoder: 2 Method Our 'training dataset consists of pairs (c, y) of images EUR and their corresponding captions y. Given an image EUR, let Zi and Zt be its CLIP image and text embeddings, respectively: We design our generative stack to produce images from captions using two components: A prior P(zily) that produces CLIP image embeddings z; conditioned on captions yA decoder P(clzi, y) that produces images conditioned on CLIP image embeddings zi (and optionally text captions y) The decoder allows US to invert images given their CLIP image embeddings, while the prior allows us t0 learn generative model of the image embeddings themselves. Stacking these two components yields a generative model P(zly) of images z given captions y: P(cly) = P(w, zily) = P(wlzi, y)P(zily). The first equality holds because Zi is a deterministic function of EUR. The second equality holds because of the chain rule _ Thus, we can sample from the true conditional distribution P(zly) by first sampling Zi using the 3 prior, and then sampling EUR using the decoder: In the following sections, we describe our decoder and prior stacks. For training details and hyperparameters, refer to Appendix[ 2. 1 Decoder We use diffusion models [25, [48 to produce images conditioned on CLIP image embeddings (and optionally text captions). Specifically, we modify the architecture described in Nichol et al 2021) by projecting and adding CLIP embeddings to the existing timestep embedding, and by projecting CLIP embeddings into four extra tokens of context that are concatenated t0 the sequence of outputs from the GLIDE text encoder: We retained the text conditioning pathway present in the original GLIDE model, hypothesizing that it could allow the diffusion model to learn aspects of natural language that CLIP fails to capture (e. g variable binding), but find that it offers little help in this regard (Section[}. While we can sample from the conditional distribution of the decoder directly, past work using diffusion models shows using guidance on the conditioning information [11J[24JB5] improves sample quality a lot: We enable classifier-free guidance [24 by randomly setting the CLIP embeddings to zero (Or learned embedding) 10% of the time, and randomly dropping the text caption 50% of the time during training: To generate high resolution images, we train tWo diffusion upsampler models [34, 443]: one to upsample images from 64 X 64t0 256 x 256 resolution, and another to further upsample those to 1024 1024 resolution_ To improve the robustness of our upsamplers, we slightly corrupt the conditioning images during training: For the first upsampling stage, we use gaussian blur [43/, and for the second we use a more diverse BSR degradation [42 [59j. To reduce training compute and improve numerical stability, we follow Rombach et al] [42 and train on random crops of 'images that are one-fourth the target size. We use oly spatial convolutions in the model (i. e,, no attention layers) and at inference time directly apply the model at the target resolution, observing that it readily generalizes to the higher resolution We found no benefit from conditioning the upsamplers on the caption, and use unconditional ADMNets with no guidance 2. 2 Prior While a decoder can invert CLIP image embeddings zi to produce images EUR, we need a prior model that produces Zi from captions y to enable image generations from text captions. We explore two different model classes for the prior model: Autoregressive (AR) prior: the CLIP image embedding zi is converted into sequence of discrete codes and predicted autoregressively conditioned on the caption y. Diffusion prior: The continuous vector Zi is directly modelled using a Gaussian diffusion model conditioned on the caption y: In addition to the caption, we can condition the prior on the CLIP text embedding 2t since it is deterministic function of the caption. To improve sample quality we also enable sampling using classifier-free guidance for both the AR and diffusion prior; by randomly dropping this text conditioning information 1O9 of the time during training: To train and sample from the AR prior more efficiently, we first reduce the dimensionality of the CLIP image embeddings zi by applying Principal Component Analysis (PCA) [377 In particular; we find that the rank of the CLIP representation space is drastically reduced when training CLIP with SAM [15] while slightly improving evaluation metrics We are able to preserve nearly all of the informatiorZby retaining only 319 principal components out of the original 1, 024. After applying PCA, we order the principal components by decreasing eigenvalue magnitude, quantize each of the 319 dimensions into 1, 024 discrete buckets, and 2Le., less than 1% average mean-squared error in reconstructing the image representations 7 Figure 3: Variations of an input image by encoding with CLIP and then decoding with a diffusion model_ The variations preserve both semantic information like presence of a clock in the painting and the overlapping strokes in the logo, as well as stylistic elements like the surrealism in the painting and the color gradients in the logo, while varying the non-essential details predict the resulting sequence using a Transformer [53[ model with a causal attention mask. This results in threefold reduction in the number of tokens predicted during inference, and improves training stability We condition the AR prior on the text caption and the CLIP text embedding by encoding them as a prefix to the sequence Additionally, we prepend a token indicating the (quantized) dot product between the text embedding and image embedding, Zi Zt. This allows uS t0 condition the model on higher dot product, since higher text-image dot products correspond to captions which better describe the image. In practice, we find it beneficial to sample the dot product from the top half of the distributiong] For the diffusion prior; we train decoder-only Transformer with a causal attention mask on sequence consisting 0f, in order: the encoded text, the CLIP text embedding; an embedding for the diffusion timestep, the noised CLIP image embedding, and final embedding whose output from the Transformer is used to predict the unnoised CLIP image embedding We choose not to condition the diffusion prior on Zi Zt like in the AR prior; instead_ we improve quality during sampling time by generating two samples of zi and selecting the one with a higher dot product with Zt_ Instead of using the EUR-prediction formulation from Ho et al. ] [251, we find it better t0 train our model to predict the unnoised Zi directly, and use a mean-squared error loss on this prediction: Lprior E, ~[1, 7]+O~q [Ifo(-{ t, y) 2ll2] 'We swept over percentiles S0%, 709, 8S9, 959 and found S0% to be optimal in all experiments. 5 0 Figure 4: Variations between two images by interpolating their CLIP image embedding and then decoding with a diffusion model. We fix the decoder seed across each row. The intermediate variations naturally blend the content and style from both input images. 3 Image Manipulations Our approach allows us to encode any given image EUR into a bipartite latent representation (Zi; TT) that is sufficient for the decoder to produce an accurate reconstruction. The latent Zi describes the aspects of the image that are recognized by CLIP while the latent TT encodes all of the residual information necessary for the decoder t0 reconstruct x. The former is obtained by simply encoding the image with the CLIP image encoder: The latter is obtained by applying DDIM inversion (Appendix F in [11]) to x using the decoder; while conditioning 0n Zi. We describe three different kinds of manipulations that are enabled by this bipartite representation_ 3. 1 Variations Given an image EUR, we can produce related images that share the same essential content but vary in other apects, such as shape and orientation (FigureB]. To do this, we apply the decoder to the bipartite representation Zi, TT) using DDIM with n 0 for sampling: With n = 0, the decoder becomes deterministic and will reconstruct the given image EUR. Larger values of introduce stochasticity into successive sampling steps, resulting in variations that are perceptually "centered" around the original image EUR. As n increases, these variations tell us what information was captured in the CLIP image embedding (and thus is preserved across samples), and what was lost and thus changes across the samples). photo of a cat an anime drawing of a super saiyan cat; artstation photo of a victorian house photo of a modern house photo of an adult lion photo of lion cub photo of a landscape in winter a photo of a landscape in fall Figure 5: Text diffs applied to images by interpolating between their CLIP image embeddings and a normalised difference of the CLIP text embeddings produced from the two descriptions. We also perform DDIM inversion to perfectly reconstruct the input image in the first column, and fix the decoder DDIM noise across each rOW. 3. 2 Interpolations Itis also possible to blend tWo images 11 and T2 for variations (Figure[4Hh; traversing all of the concepts in CLIP' $ embedding space that occur between them To do this, we rotate between their CLIP embeddings Zi1 and #i2 using spherical interpolation, yielding intermediate CLIP representations Zio slerp(zi1Zi2 0) as 0 is varied from 0 to 1. There are two options for producing the intermediate DDIM latents along the trajectory: The first option involves interpolating between their DDIM inverted latents TT1 and TT2 (by setting TTe slerp(TT; TTz 0)), which yields a single trajectory whose endpoints reconstruct T1 and T2 The second option involves fixing the DDIM latent to a randomly-sampled value for all interpolates in the trajectory: This results in an infinite number of trajectories between T1 and. 2, though the endpoints of these trajectories will generally no longer coincide with the original images. We use this approach in Figurel 3. 3 Text Diffs key advantage of using CLIP compared to other models for image representations is that it embeds images and text to the same latent space, thus allowing uS to apply language-guided image manipulations (i. e. _ text diffs), which we show in Figure[5] To modify the image to reflect a new text description y, we first obtain its CLIP text embedding 2t, as well as the CLIP text embedding #t0 of caption describing the current imagq We then compute a text diff vector 2d norm( Zt 2t0 from these by taking their difference and Instead of a description of the current image, we also experimented with using dummy caption like photo for the baseline, Or removing it altogether: These also worked well_ Pizz) Feize Ap A Ppck Oc @ Pabl Piot Piz Prpo Ppite PPEale PMAz Granny Smith: 100% iPod: 0% Pizza: 0% Granny Smith: 0. 02% iPod: 99. 98% Pizza: 09 Granny Smith: 94. 33% iPod: 0% Pizza: 5. 669 Figure 6: Variations of images featuring typographic attacks [20] paired with the CLIP model'$ predicted probabilities across three labels. Surprisingly, the decoder still recovers Granny Smith apples even when the predicted probability for this label is near O%. We also find that our CLIP model is slightly less susceptible to the "pizza' attack than the models investigated in [20]. normalizing: Now; we can rotate between the image CLIP embedding Zi and the text diff vector 2d using spherical interpolation, yielding intermediate CLIP representations 2o slerp(zi, 2d, 0 ), where 0 is increased linearly from 0 to a maximum value that is typically in [0. 25, 0. 50]. We produce the final outputs by decoding the interpolates 20, fixing the base DDIM noise t0 TT throughout the entire trajectory: Probing the CLIP Latent Space Our decoder model provides unique opportunity to explore CLIP latent space by allowing US to directly visualize what the CLIP image encoder is seeing: As an example use case, we can revisit cases where CLIP makes incorrect predictions, such as typographic attacks [201. In these adversarial images, a piece of text is overlayed on top of an object; which causes CLIP to predict the object described by the text rather than the object depicted in the image. This piece of text essentially hides the original object in terms of output probabilities In Figure[] we show an example of this attack from [20], wherein an apple can be misclassified as an iPod_ Surprisingly, we find that our decoder still generates pictures of apples with high probability even though the predicted probability of "Granny Smith' is near zero. Even more notable, the model never produces pictures of iPods, despite the very high relative predicted probability of this caption_ iPod Piza PIzA LLC Fpxae 1zlri Figure 7: Visualization Of reconstructions of CLIP latents from progressively more PCA dimensions (20, 30, 40, 80, 120, 160, 200, 320 dimensions ), with the original source image on the far right: The lower dimensions preserve coarse-= grained semantic information, whereas the higher dimensions encode finer-grained details about the exact form of the objects in the scene: PCA reconstructions offer another tool for probing the structure of the CLIP latent space. In Figure[] we take the CLIP image embeddings of a handful of source images and reconstruct them with progressively more PCA dimensions, and then visualize the reconstructed image embeddings using our decoder with DDIM on fixed seed_ This allows us t0 see what semantic information the different dimensions encode. We observe that the early PCA dimensions preserve coarse-grained semantic information such as what types of objects are in the scene, whereas the later PCA dimensions encode finergrained detail such as the shapes and exact form of the objects. For example, in the first scene, the earlier dimensions seem to encode that there is food and perhaps a container present; whereas the later dimensions encode tomatoes and bottle specifically Figure[] also serves as visualization of what the AR prior is modeling, since the AR prior is trained to explicitly predict these principal components in this order: 5 Text-to-Image Generation 5. 1 Importance of the Prior Although we train a prior to generate CLIP image embeddings from captions, the prior is not strictly necessary for caption-to-image generation: For instance, our decoder can condition on both CLIP image embeddings and captions, but the CLIP image embedding is dropped 59 of the time during training in order to enable classifier-free guidance. Therefore, at sampling time, we can condition on only the caption, although this underperforms model trained fully in this way (this model is GLIDE, and we do a thorough comparison with GLIDE in Sections[ 2]and[3}. Another possibility is to feed the decoder the CLIP text embedding as if it were an image embedding, as previously observed [61J[54]. The first two rows of Figure[8]depicts samples obtained in these two ways; the third row depicts samples obtained with prior: Conditioning the decoder on just the caption is clearly worst; but conditioning On text embeddings zero-shot does produce reasonable results_ Building on this observation, another approach would be to train the decoder to condition on CLIP text embeddings [9] instead of CLIP image embeddings (although we would lose the capabilities mentioned in Section[) To quantify the effectiveness of these alternate approaches, we train two models: small decoder conditioned on CLIP text embeddings, and a small unCLIP stack (diffusion prior and decoder). We then compare samples from the text-embedding decoder; samples from the UnCLIP stack, and samples obtained from feeding text J 1 F 1 1 ~A group of baseball an oil painting of a hedgehog using a ~A motorcycle parked in a 'This wire metal rack players is crowded at corgi wearing a calculator" parking space next to holds several pairs of the mound paty hat" another motorcycle. shoes and sandals' Figure &: Samples using different conditioning signals for the same decoder: In the first IOW, We pass the text caption to the decoder; and pass a zero vector for the CLIP embedding: In the second row, we pass both the text caption and the CLIP text embedding of the caption. In the third IOW, we pass the text and CLIP image embedding generated by an autoregressive prior for the given caption_ Note that this decoder is only trained to do the text-to-image generation task (without the CLIP image representation) S% of the time. embeddings to the unCLIP decoder zero-shot; sweeping across guidance scales for all models_ We find that these approaches respectively score FIDs of 9. 16, 7. 99, and 16. 55 on a test set, suggesting the unCLIP approach is best: We also run human evaluations comparing the first two settings, sweeping over sampling hyperparameters for each using our human evaluation proxy model (Appendix[A) We find that humans prefer the full unCLIP stack 57. 0% = 3. 19 of the time for photorealism and 53. 19 E 3. 1% of the time for caption similarity: Given the importance of the prior; it is worth evaluating different approaches for training it. We compare both the AR and diffusion priors throughout our experiments_ In all cases Sections[. 2155. 4 and[5), we find that the diffusion prior outperforms the AR prior for comparable model size and reduced training compute. 5. 2 Human Evaluations We observe in FigurefJthat unCLIP is capable of synthesizing complex, realistic images. While we can compare sample quality to past models using FID, it is not always aligned with human judgment To better gauge the generation capabilities of our system; we conduct systematic human evaluations comparing UnCLIP to GLIDE for photorealism, caption similarity, and sample diversity We follow the protocol of Ramesh et al ] Nichol et al. 40, B5_ for the first two evaluations: for photorealism, users are presented with pairs of images and must choose which looks more photorealistic; for caption 10 3 unCLIP GLIDE Figure 9: Samples when increasing guidance scale for both unCLIP and GLIDE, using the prompt; A green vase filled with red roses sitting on top of table."For UnCLIP, we fix the latent vectors sampled from the prior; and only vary the guidance scale of the decoder: For both models, we fix the diffusion noise seed for each column. Samples from unCLIP improve in quality (more realistic lighting and shadows) but do not change in content aS we increase guidance scale, preser= ving semantic diversity even at high decoder guidance scales unCLIP Prior Photorealism Caption Similarity Diversity 47. 19 = 3. 1% 41. 19 E 3. 0% 62. 6% = 3. 0% 48. 99 = 3. 1% 45. 39 = 3. 0% 70. 59 1 2. 8% AR Diffusion Table 1: Human evaluations comparing UnCLIP to GLIDE We compare to both the AR and diffusion prior for UnCLIP: Reported figures are 95% confidence intervals of the probability that the UnCLIP model specified by the row beats GLIDE: Sampling hyperparameters for all models were wept to optimize an automated proxy for human photorealism evaluations. similarity, users are additionally prompted with a caption, and must choose which image better matches the caption In both evaluations, there is a third "Not sure" option For diversity, we propose a new evaluation protocol in which humans are presented with two 4 X 4 grids of samples and must choose which is more diverse (with a third option; *Not sure") For this evaluation; we produce sample grids using 1, 000 captions from the MS-COCO validation set, and always compare sample grids for the same caption. Before running human comparisons, we swept over sampling hyperparameters for each model using a CLIP linear probe trained to be a proxy for human photorealism evaluations (Appendix[}: These hyperparameters are fixed across all three types of evaluation: We present our results in Table/l In general, the diffusion prior performs better than the AR prior in pairwise comparisons against GLIDE. We find that humans still slightly prefer GLIDE to unCLIP in terms of photorealism, but the gap is very small. Even with similar photorealism, unCLIP is strongly preferred over GLIDE in terms of diversity highlighting one of its benefits. 3 80% 3 70% 1 60% unCLIP is better 50% 2 GLIDE is better 3 40% 30% in terms of photorealism [ + in terms of caption similarity 20% in terms of diversity 1. 0 1. 5 2. 0 2. 5 3. 0 GLIDE guidance scale Figure 10: When comparing UnCLIP (with our best sampling settings) to various settings of guidance scale for GLIDE; unCLIP was preferred by human evaluators on at least one axis among photorealism, caption similarity, and diversity for each comparison: At the higher guidance scales used to generate photorealistic images, UnCLIP yields greater diversity for comparable photorealism and caption similarity 18 216 8 14 12 GLIDE unCLIP (AR) unCLIP (Diffusion) 10 1. 0 1. 5 2. 0 2. 5 3. 0 Guidance Scale 3. 5 4. 0 Figure ]l: FID versus guidance scale for UnCLIP and GLIDE. FOr the UnCLIP priors, we swept over sampling hyperparameters and fixed to the settings with the best minimum FID. 5. 3 Improved Diversity-Fidelity Trade-off with Guidance Compared to GLIDE, we qualitatively observe that unCLIP is able to generate more diverse images while leveraging the guidance technique to improve sample quality To understand why, consider Figure [9]where we increase guidance scale for both GLIDE and unCLIP. For GLIDE, the semantics camera angle, color; size) converge aS we increase guidance scale, whereas for unCLIP the semantic information of the scene is frozen in the CLIP image embedding and therefore does not collapse when guiding the decoder: In Section/5. 2 we observed that unCLIP achieves similar photorealism as GLIDE while maintaining more diversity, but that its caption matching capabilities were slightly worse. It is natural to ask whether GLIDE'$ guidance scale can be lowered to obtain the same diversity level as unCLIP while maintaining better caption 12 Model FID Zero-shot FID Zero-shot FID (filt) AttnGAN Xu et al. 2017 DM-GAN Lhu et al. 2019 DF-GAN Tao et al_ 2020= DM-GAN + CL Ye et al 2021 XMC-GAN Zhang et al. |2021 LAFITE Zhou et al_ 2021 Make-A-Scene Gatn et al_ 2022 DALL-E Ramesh et al. 120217 LAFITE Zhou et al. 12021 GLIDE Nichol et al 2021 Make-A-Scene Gatni et al_ 2022 unCLIP AR prior) unCLIP Diffusion prior) 35. 49 32. 64 21. 42 20. 79 9. 33 8. 12 7. 55 28 26. 94 12. 24 12. 89 11. 84 11. 08 10. 87 10. 63 10. 39 Table 2: Comparison of FID on MS-COCO 256 X 256. We use guidance scale 1. 25 for the decoder for both the AR and diffusion prior; and achieve the best results using the diffusion prior: matching: In Figure[o we conduct a more careful study of this question by performing human evaluations across several GLIDE guidance scales. We find that GLIDE at guidance scale 2. 0 is very close to the photorealism and caption similarity of unCLIP, while still producing less diverse samples. Finally, in Figure[ we compute MS-COCO zero-shot FID [23p while sweeping over guidance scale for both unCLIP and GLIDE, finding that guidance hurts the FID of unCLIP much less so than for GLIDE. In this evaluation; We fix the guidance scale of the unCLIP prior and only vary the guidance scale of the decoder: This is another indication that guidance hurts the diversity of GLIDE much more than unCLIP; since FID heavily penalizes non-diverse generations. 5. 4 Comparison on MS-COCO In the text-conditional image generation literature, it has become standard practice t0 evaluate FID on the MS-COCO [28_ validation set We present results on this benchmark in Tablel] Like GLIDE and DALL-E, unCLIP is not directly trained on the MS-COCO training set; but can still generalize to the validation set zero-shot. We find that, compared to these other zero-shot models, unCLIP achieves a new state-of-the-art FID of 10. 39 when sampling with the diffusion prior: In Figure[2] we visually compare unCLIP to various recent text-conditional image generation models on several captions from MS-COCO. We find that; like the other methods, unCLIP produces realistic scenes that capture the text prompts. 5. 5 Aesthetic Quality Comparison We additionally perform automated aesthetic quality evaluations comparing UnCLIP to GLIDE: Our goal with this evaluation is t0 assess how well each model produces artistic illustrations and photographs. To this end, we generated 512 'artistic" captions using GPT-3 [4] by prompting it with captions for existing artwork (both real and Al generated) Next; we trained CLIP linear probe to predict human aesthetic judgments using the AVA dataset [33| (AppendixAA For each model and set of sampling hyperparameters, We produce four images for each prompt, and report the mean predicted aesthetic judgment over the full batch of 2048 images In Figure[3] we present results 0n our aesthetic quality evaluation: We find that guidance improves aesthetic quality for both GLIDE and unCLIP. For unCLIP we only guide the decoder (we found that guiding the prior hurt results). We also plot the aesthetic quality against Recalf5] since guidance typically induces trade-off SRecall is computed with respect to the training dataset: 13 1 2 7 8 1 8 1 8 a green train is coming down the tracks a group of skiers are preparing to ski down a mountain_ a small kitchen with a lOw ceiling' a group of elephants walking in muddy water: living area with a television and table" Figure 12: Random image samples on MS-COCO prompts_ 14 4. 85 0. 600 0. 575 L 4. 80 0. 550 4. 75? 0. 525 < 4. 70 GLIDE 0. 500 GLIDE 1 4. 65 unCLIP (AR) 0. 475 unCLIP (AR) unCLIP (diffusion) unCLIP (diffusion) 4. 60 0. 450 1. 0 1. 5 2. 0 2. 5 3. 0 3. 5 4. 0 4. 60 4. 65 4. 70 4. 75 4. 80 4. 85 guidance scale mean AVA prediction Figure 13: Aesthetic quality evaluations comparing GLIDE and unCLIP using 512 auto-generated artistic prompts. We find that both models benefit from guidance, but unCLIP does not sacrifice recall for aesthetic quality: between fidelity and diversity: Interestingly; we find that guiding UnCLIP does not decrease Recall while still improving aesthetic quality according to this metric. Related Work Synthetic image generation is a well studied problem, and most popular techniques for unconditional image generation have also been applied to the text-conditional setting Many previous works have trained GANs [21] on publicly available image captioning datasets to produce text-conditional image samples 56, 63_ 49, [58/57]. Other works have adapted the VQ-VAE approach [52] to text-conditional image generation by training autoregressive transformers on sequences of text tokens followed by image tokens 40, [12[1|. Finally, some works have applied diffusion models to the problem, training either continuous 351 or discrete 221 diffusion models with auxiliary text encoders to handle textual input: Previous works have leveraged hierarchical generative processes to create high-quality synthetic images_ Razavi et al. 41 trains multi-layer discrete autoencoder; allowing them to first sample coarse-grained latent codes and then use this as conditioning information when sampling higher-resolution latent codes. Child [Vahdat and Kautz [5J[50] generate images using VAEs with a hierarchy of latent codes that increase progressively with resolution. Concurrently with our work, Gafni et al. ][17| conditions a generative image model on segmentation masks, allowing for a generative process that first samples a semantic map of an image and then conditions the generated image O this information: The computational benefits of using diffusion to model latent space has been noted by previous works Preechakul et al. 38 propose an autoencoder framework where diffusion models are used to render latent variables as images, and a second diffusion model is used to generate these latents (similar to our diffusion prior) [Vahdat et al. ][51] use a score-based model for the latent space of a VAE, whileRombach et al. 42] use diffusion models on the latents obtained from VQGAN [14] like autoencoder: Since its release, CLIP [39_ has been used extensively to steer generative image models towards text prompts_ Galatolo et al. Patashnik et al. Murdock Gal et al. ] [19, 16, B2/, [18/ guide GANs using gradients from CLIP model. For diffusion models, Dhariwal and Nichol 11| introduced classifier guidance as a way to use gradients from classifier trained on noised images to steer the model towards higher quality generations. Nichol et al:. 35 train a CLIP model on noised images and guide text-conditional diffusion model while Crowson Crowson [71/8 use an unnoised CLIP model to guide unconditional or class-conditional diffusion models _ Ho and Salimans 24] introduced classifier-free guidance and showed that one can perform guidance 15 unCLIP GLIDE Figure 14: Samples from unCLIP and GLIDE for the prompt a red cube on top of a blue cube implictly from the predictions of the model with and without the conditioning information, thus removing the need for a classifier: Nichol et al. [35 | showed classifier-free guidance works more favorably than CLIP guidance for text conditional image generation: Several previous works have trained generative image models that are directly conditioned on CLIP embeddings Zhou et al:] 61] condition GAN models on randomly perturbed CLIP image embeddings, finding that these models can generalize to CLIP text embeddings to produce text-conditional images Crowson [9] trained diffusion models conditioned on CLIP text embeddings, allowing for direct text-conditional image generation: Wang et all] 54] train an autoregressive generative model conditioned o CLIP image embeddings, finding that it generalizes to CLIP text embeddings well enough to allow for text-conditional image synthesis Bordes et al ] [3] train diffusion models conditioned on image representations from contrastive models While the diffusion models themselves cannot generate images unconditionally, the authors experimented with a simple approach for two-stage image generation by employing Kernel Density Estimation to sample image representations By feeding these generated representations to the diffusion model, they can generate imas ges end-to-end in a way similar t0 our proposed technique. However; our work differs from this in two ways: first, we use multimodal contrastive representations rather than image-only representations; second, we employ much more powerful generative models for the first stage of the generation hierarchy, and these generative models are conditioned on text: 7 Limitations and Risks Although conditioning image generation on CLIP embeddings improves diversity, this choice does come with certain limitations In particular; UnCLIP is worse at binding attributes t0 objects than a corresponding GLIDE model_ In Figure[4 we find that unCLIP struggles more than GLIDE with prompt where it must bind two separate objects (cubes) to two separate attributes (colors)_ We hypothesize that this occurs because the CLIP embedding itself does not explicitly bind attributes to objects, and find that reconstructions from the decoder often mix up attributes and objects, as shown in Figure[[5} A similar and likely related issue is that unCLIP 16 Figure 15: Reconstructions from the decoder for difficult binding problems_ We find that the reconstructions mix up objects and attributes In the first two examples, the model mixes up the color of two objects In the rightmost example, the model does not reliably reconstruct the relative size of two objects Deinp Lerpt: Diep PONEELH Deep Figure 16: Samples from unCLIP for the prompt; "A sign that says deep learning:' struggles at producing coherent text; as illustrated in Figure[6} it is possible that the CLIP embedding does not precisely encode spelling information of rendered text This issue is likely made worse because the BPE encoding we use obscures the spelling of the words in caption from the model, so the model needs to have independently seen each token written out in the training images in order to learn to render it: We also note that our stack still has hard time producing details in complex scenes (Figure[7. We hypothesize that this is a limitation of our decoder hierarchy producing an image at a base resolution of 64 X 64 and then upsampling it Training our unCLIP decoder at a higher base resolution should be able to alleviate this, at the cost of additional training and inference compute_ As discussed in the GLIDE paper; image generation models carry risks related to deceptive and otherwise harmful content: unCLIP' $ performance improvements also raise the risk profile over GLIDE. As the technology matures, it leaves fewer traces and indicators that outputs are Al-generated, making it easier to mistake generated images for authentic ones and vice versa. More research is also needed on how the change in architecture changes how the model learns biases in training data. 17 Deep A high quality photo of a dog playing in a green field next to a lake. b) high quality photo of Times Square_ Figure 17: unCLIP samples show low levels of detail for some complex scenes The risks of these models should be assessed in relation to the particular deployment context; which includes training data, guardrails in place, the deployment space, and who will have access: A_ preliminary analysis of these issues in the context of the DALLE 2 Preview platform (the first deployment of an unCLIP model), can be found in Mishkin et al. [30]. 8 Acknowledgements We 'd like to thank Jong Wook Kim, Hyeonwoo Noh, Alec Radford, Pranav Shyam, and Ilya Sutskever for helpful discussions and contributions t0 our work: We'd also like to thank Yunxin Jiao for creating several figures used in the paper: We are grateful to the Acceleration and Supercomputing teams at OpenAL for their work o software and hardware infrastructure this project used: 18 References [1] Armen Aghajanyan, Bernie Huang; Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, and Luke Zettlemoyer: CM3: A Causal Masked Multimodal Model of the Internet. arXiv:2201. 07520, 2022 [2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang: Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models_ CoRR, abs/2201. 06503, 2022. URL https / [arxiv_ org abs_ 2201. 06503 [3] Florian Bordes, Randall Balestriero, and Pascal Vincent High Fidelity Visualization of What Your Self-Supervised Representation Knows About: arXiv:2112. 09164 2021. [4] Tom B. Brown, Benjamin Mann, Nick Ryder; Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel HerbertVoss, Gretchen Krueger; Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler; Jeffrey Wu; Clemens Winter; Christopher Hesse, Mark Chen, Eric Sigler; Mateusz Litwin, Scott Gray; Benjamin Chess, Jack Clark, Christopher Berner; Sam McCandlish, Alec Radford, Ilya Sutskever; and Dario Amodei. Language Models are Few-Shot Learners arXiv:2005. 14165 2020. [5] Rewon Child. Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images arXiv:2011. 10650, 2021. [6] Katherine Crowson_ AVA Linear Probe. https [twitter com/RiversHaveW_ Jings /status 14723461867281735687s-20&t=T HRr3GwSHRGj QaMDtRe3a 2021. [7] Katherine Crowson: CLIP guided diffusion HQ 256x256. https I/colab. research. google _ com drive/12a_Wrfi2_gwwluN3VvMTwVMz9TfqctNj 2021. [8] Katherine Crowson. CLIP Guided Diffusion 512x512, Secondary Model Method. https I [twitter com/RiversHavew Jings, status 1462859669454536711/ 2021. [9] Katherine Crowson V-diffusion_ https: github. com_ crowsonkb/v-diffusion-pytorch, 2021. [10] Karan Desai and Justin Johnson: VirTex: Learning Visual Representations from Textual Annotations arXiv:2006. 06666, 2020. [11] Prafulla Dhariwal and Alex Nichol. arXiv:2105. 05233, 2021. Diffusion Models Beat GANs on Image Synthesis. [12] Ming Ding, Zhuoyi Yang; Wenyi Hong; Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang: Cog View: Mastering Text-to-Image Generation via Transformers_ arXiv:2105. 13290, 2021. [13] Alexey Dosovitskiy, Lucas Beyer; Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner; Mostafa Dehghani, Matthias Minderer; Georg Heigold, Sylvain Gelly, Jakob Uszkoreit; and Neil Houlsby: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale: arXiv:2010. 11929, 2020. [14] Patrick Esser; Robin Rombach, and Bjorn Ommer: Taming Transformers for High-Resolution Image Synthesis arXiv: 2012. 09841 2020. [15] Pierre Foret, Ariel Kleiner; Hossein Mobahi, and Behnam Neyshabur: Sharpness-Aware Minimization for Efficiently Improving Generalization arXiv:2010. 01412 2020. 19 [16] Andreas Fiirst, Elisabeth Rumetshofer; Viet Thuong Tran, Hubert Ramsauer; Fei Tang; Johannes Lehner; D P Kreil, Michael K Kopp; Giinter Klambauer; Angela Bitto-Nemling, and Sepp Hochreiter: CLOOB: Modern Hopfield Networks with InfoLOOB Outperform CLIP, 2022. URL https / /openreview net / forum? id-qw674LIPfQE [17] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman Make-AScene: Scene-Based Text-to-Image Generation with Human Priors arXiv:2203. 13131 2022. [18] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or: StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators_ arXiv: 2108. 00946, 2021. [19] Federico A. Galatolo, Mario G. C. A. Cimino, and Gigliola Vaglini. Generating images from caption and vice versa via CLIP-Guided Generative Latent Space Search: arXiv:2102. 01645, 2021. [20] Gabriel Goh; Nick Cammarata Chelsea Voss t, Shan Carter; Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah: Multimodal Neurons in Artificial Neural Networks_ Distill, 2021. doi: 10. 23915/distill. 00030. https:/ Ildistill-pub/202l/multimodal-neurons [21] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu; David Warde-Farley, Sherjil Ozair; Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks arXiv: 1406. 2661 2014. [22] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang; Dongdong Chen, Lu Yuan, and Baining Guo. Vector Quantized Diffusion Model for Text-to-Image Synthesis. arXiv:2 111. 14822 2021. [23] Martin Heusel, Hubert Ramsauer; Thomas Unterthiner; Bernhard Nessler; and Sepp Hochreiter: GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. Advances in Neural Information Processing Systems 30 NIPS 2017), 2017. [24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In NeurIPS 202 1 Workshop on Deep Generative Models and Downstream Applications, 2021. URL https '/openreview net_ forum? id-qw8AKxfYbI [25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. arXiv:2006. 11239, 2020. [26] Jonathan Ho, Chitwan Saharia, William Chan, David J. Fleet; Mohammad Norouzi, and Tim Salimans_ Cascaded Diffusion Models for High Fidelity Image Generation. arXiv:2106. 15282, 2021. [27] Diederik P Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv:1412. 6980 2014 [28] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev; Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C Lawrence Zitnick, and Piotr Dollar: Microsoft COCO: Common Objects in Context. arXiv:1405. 0312 2014. [29] Ilya Loshchilov and Frank Hutter: Decoupled Weight Decay Regularization: arXiv: 1711. 05101 2017. [30] Pamela Mishkin, Lama Ahmad, Miles Brundage, Gretchen Krueger; and Girish Sastry. DALLE 2 Preview Risks and Limitations 2022_ URL https github. com/openai/dalle-2-preview/ blob/main/system-card. md [31] Norman Mu; Alexander Kirillov, David Wagner; and Saining Xie. SLIP: Self-supervision meets Language-Image Pre-training. arXiv:2 112. 12750, 2021. [32] Ryan Murdock The Big Sleep. https: / /twitter com_ advadnoun status_ 1351038053033406468, 2021. 20 [33] Naila Murray, Luca Marchesotti, and Florent Perronnin: AVA: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408-2415, 2012. doi: 10. 1109/CVPR. 2012. 6247954_ [34] Alex Nichol and Prafulla Dhariwal Improved Denoising Diffusion Probabilistic Models: arXiv:2102. 09672, 2021. [35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever; and Mark Chen_ GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv:2112. 10741 2021. [36] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or; and Dani Lischinski. StyleCLIP: TextDriven Manipulation of StyleGAN Imagery. arXiv:2103. 17249, 2021. [37] Karl Pearson. LIII. On lines and planes of closest fit to systems of points in space, November 1901. URL https I Idoi org 10. 1080/14786440109462720 [38] Konpat Preechakul, Nattanat Chatthee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion Autoencoders: Toward Meaningful and Decodable Representation. arXiv:2111. 15640, 2021_ [39] Alec Radford Jong Wook Kim, Chris Hallacy, Aditya Ramesh; Gabriel Goh; Sandhini Agarwal, Girish Sastry; Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever: Learning Transferable Visual Models From Natural Language Supervision: arXiv:2103. 00020 2021. [40] Aditya Ramesh, Mikhail Pavlov; Gabriel Goh, Scott Gray; Chelsea Voss, Alec Radford, Mark Chen; and Ilya Sutskever Zero-Shot Text-to-Image Generation. arXiv:2102. 12092, 2021. [41] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating Diverse High-Fidelity Images with VQ-VAE-2. arXiv:1906. 00446, 2019. [42] Robin Rombach; Andreas Blattmann, Dominik Lorenz, Patrick Esser; and Bjorn Ommer: HighResolution Image Synthesis with Latent Diffusion Models. arXiv:2112. 10752, 2021. [43] Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi _ Image Super-Resolution via Iterative Refinement arXiv:arXiv:2104. 07636, 2021. [44] Mert Bulent Sariyildiz, Julien Perez. and Diane Larlus. Learning Visual Representations with Caption Annotations_ arXiv:2008. 01392, 2020. [45] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach; Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer How Much Can CLIP Benefit Vision-and-Language Tasks? arXiv:2107. 06383, 2021. [46] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep Unsupervised Learning using Nonequilibrium Thermodynamics arXiv: 1503. 03585, 2015. [47] Jiaming Song; Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models arXiv: 2010. 02502, 2020. [48] Yang Song and Stefano Ermon. Improved Techniques for Training Score-Based Generative Models arXiv:2006. 09011 2020. [49] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing; Fei Wu; and Bingkun Bao. DF-GAN: Deep Fusion Generative Adversarial Networks for Text-to-Image Synthesis arXiv:2008. 05865, 2020. [50] Arash Vahdat and Jan Kautz. NVAE: A Deep Hierarchical Variational Autoencoder: arXiv:2007. 03898, 2020. 21 [51] Arash Vahdat, Karsten Kreis, and Jan Kautz_ Score-based Generative Modeling in Latent Space. In Neural Information Processing Systems (NeurIPS), 2021. [52] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu: Neural Discrete Representation Learning: arXiv: 1711. 00937, 2017. [53] Ashish Vaswani, Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones, Aidan N. Gomez, Lukasz Kaiser; and Illia Polosukhin: Attention Is All You Need_ arXiv:1706. 03762, 2017. [54] Zihao Wang; Wei Liu, Qian He, Xinglong Wu, and Zili Yi. CLIP-GEN: Language-Free Training of a Text-to-Image Generator with CLIP. arXiv:2203. 00386, 2022. [55] Weihao Xia, Yulun Zhang: Yujiu Yang; Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang: GAN Inversion: A Survey arXiv:2101. 05278, 2021. [56] Tao Xu, Pengchuan Zhang; Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He_ AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks arXiv: 1711. 10485, 2017. [57] Hui Ye, Xiulong Yang, Martin Takac, Rajshekhar Sunderraman, and Shihao Ji. Improving Text-to-Image Synthesis Using Contrastive Learning: arXiv:2107. 024232021. [58] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang: Cross-Modal Contrastive Learning for Text-to-Image Generation: arXiv:2101. 04702 2021. [59] Kai Zhang; Jingyun Liang; Luc Van Gool, and Radu Timofte. Designing a Practical Degradation Model for Deep Blind Image Super-Resolution. 2021 IEEEICVF International Conference on Computer Vision (ICCV), Oct 2021. doi: 10. 1109/iccv48922. 2021. 00475. URL http: / /dx doi org 10. 1109/ ICCV48922. 2021. 00475 [60] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P Langlotz. Contrastive Learning of Medical Visual Representations from Paired Images and Text arXiv:2010. 00747, 2020. [61] Yufan Zhou; Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer; Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. LAFITE: Towards Language-Free Training for Text-to-Image Generation. arXiv:2111. 13792, 2021. [62] Jun-Yan Zhu, Philipp Krahenbihl, Eli Shechtman, and Alexei A. Efros Generative Visual Manipulation on the Natural Image Manifold. arXiv:1609. 03552, 2016. [63] Minfeng Zhu; Pingbo Pan; Wei Chen; and Yi Yang DM-GAN: Dynamic Memory Generative Adversarial Networks for Text-to-Image Synthesis arXiv: 1904. 01310, 2019. 22 Linear Probes for Evaluations For our evaluations, we leverage two new linear probes on top of a CLIP ViT-L/I4 [13 model To automate aesthetic quality evaluations, we follow the procedure used by Crowson [6], training a linear regression model on images and mean ratings from the AVA dataset [33 To reduce the cost of hyperparameter sweeps before conducting human evaluations, we train logistic regression model to predict win probabilities between pairs of images To train this model, we used 15, 000 pairwise image comparisons gathered from all of Our previous human evaluations_ For each comparison i, we computed CLIP image embeddings Ti and yi for the two images in the pair: We then trained a linear model f (x) such that 1/(1 + exp (f(Ti) f(yi))) approximates the probability that a human prefers the image for Yi: This can be reduced to a logistic regression problem with inputs equal t0 Yi Ti _ B Error Bars for Human Evaluation When computing error bars for human evaluations, we use the normal approximation interval with p = 0. 95. We expect the normal approximation to be accurate for such a large sample size of n 1000. Training Details The unCLIP models used for the experiments in this paper were trained with the hyperparameters described below, unless otherwise noted_ We additionally trained production version of unCLIP using similarly sized models but with modified architectures and trained for longer; we include changes to accommodate product and safety requirements (e. g. inpainting, preventing unwanted memorization), and train On a larger dataset that is filtered for aesthetic quality and safety: We report model and training hyperparameters for the paper models in Table/3 All models were trained using Adam [27 | with corrected weight decay 29| and momentum 81 0. 9_ Our CLIP model uses ViT-H/16 [13| image encoder that consumes 256 X 256 resolution images, and has width 1280 with 32 Transformer [53| blocks_ The text encoder also follows the architecture described in Radford et al. 39 it is Transformer [53 with a causal attention mask; with width 1024 and 24 Trans former blocks. Both models are trained with learning rate 3 10 and SAM [15] with p = 0. 1, where the perturbations are applied independently by the replicas, each of which uses batch size 64_ The remaining hyperparameters are the same as those reported in Radford et al, 39]_ When training the encoder; we sample from the CLIP [39| and DALL-E 40] datasets (approximately 650M images in total) with equal probability: When training the decoder; upsamplers, and prior; we use only the DALL-E dataset [40] (approximately 250M images). Incorporating the noisier CLIP dataset while training the generative stack negatively impacted sample quality in our initial evaluations_ Our decoder architecture is the 3. 5 billion parameter GLIDE model with the same architecture and diffusion hyperparameters as in Nichol et al. ] 351_ We train with learned sigma and sample with 250 strided sampling steps as in Nichol and Dhariwal [34|We use the ADMNet architecture |11 for the upsamplers_ In the first upsampling stage, we use a cosine noising schedule, 320 channels and depth of 3 resblocks per resolution inside the ADMNet We also apply gaussian blur (kernel size 3, sigma 0. 6) as described in Saharia et al. 431In the second upsampling stage, we use a linear noising schedule, 192 channels, a depth of 2 resblocks per resolution, and train with the BSR degradation from Rombach et al. [42]. Neither upsampler uses attention. To reduce inference time, we use DDIM 47 and manually tune the number of steps, with 27 steps for 256 x 256 model, and 15 steps for the 1024 X 1024 model. 23 For the AR prior; we use Transformer text encoder with width 2048 and 24 blocks and a decoder with causal attention mask width 1664, and 24 blocks. For the diffusion prior; we use Transformer with width 2048 and 24 blocks, and sample with Analytic DPM [2] with 64 strided sampling steps. To reuse hyperparameters tuned for diffusion noise schedules on images from Dhariwal and Nichol [H1 we scale the CLIP embedding inputs by 17. 2 to match the empirical variance of RGB pixel values of ImageNet images scaled to [~1, 1]: AR prior Diffusion prior 64 64 256 256 1024 Diffusion steps 1000 1000 10oO 100o Noise schedule cosine cosine cosine linear Sampling steps 64 250 27 15 Sampling variance method analytic [2| learned 134 DDIM [47| DDIM [473 Crop fraction 0. 25 0. 25 Model size IB 3. SB 700M 300M Channels 512 320 192 Depth Channels multiple 1, 2, 3, 4 1, 2, 3, 4 1, 1, 2, 2, 4, 4 Heads channels 64 Attention resolution 32, 16, 8 Text encoder context 256 256 256 Text encoder width 2048 2048 2048 Text encoder depth 24 24 24 Text encoder heads 32 32 32 Latent decoder context 384 Latent decoder width 1664 Latent decoder depth 24 Latent decoder heads 26 Dropout 0. 1 0. 1 Weight decay 4. 0e-2 6. 0e-2 Batch size 4096 4096 2048 1024 512 Iterations IM 6OOK 8OOK IM IM Learning rate 1. 6e-4 1. le-4 1. 2e-4 1. 2e-4 1. 0e-4 Adam 82 0. 91 0. 96 0. 999 0. 999 0. 999 Adam 1. Oe-10 1. Oe-6. Oe-8 1. Oe-8 1. Oe-8 EMA decay 0. 999 0. 9999 0. 9999 0. 9999 0. 9999 Table 3: Hyperparameters for the models 24 D Random samples In Figures[18[gandpOwe show random samples from our production model for some of the prompts from Figure p} Figure 18: Random samples from unCLIP for prompt "Vibrant portrait painting of Salvador Dali with robotic half face" 25 Figure 19: Random samples from unCLIP for prompt *A close up of a handpalm with leaves growing from it' 26 Idd tandy Figure 20: Random samples from unCLIP for prompt "A teddybear on a skateboard in Times Square. 27 | 1 | 5.3% |
| medical_ data_ science Ezurich Lecture 6 Machine Learning for Health Care" (261-5120-00L) (Health) Representation Learning Gunnar Ratsch; Julia Vogt, Valentina Boeva Biomedical Informatics group, Institute for Machine Learning, Department of Computer Science DINFK Gunnar Ratsch 15. 3. 2022 medical_ data_ science Ezurich Outline for today Motivation of Latent Representations Autoencoders & Sequence-to-sequence models Transformers ICU Benchmarks Generative models VAEs (GANs) SOM-VAEs Contrastive Learning DINFK Gunnar Ratsch 15. 3. 2022 2 medical_ data_ science Ezurich Towards Comprehensive Patient Models for Precision Medicine Distill Data Embed Data Connect Data Clinical Trial Design Precision Medicine Health Records Pathology Images Genomic Data X=w( p(phenotype x, t) Drugs Predict Treatment Mobile Health = argmax p(survival X, DINFK Gunnar Ratsch 15. 3. 2022 3 medical_ data_ science Elzurich What is a computational representation of a patient? Data modalities, redundancies Health state, relevant parameters Temporal aspects Ascertainment biases Interpretability What can we learn from a computational representation of a patient? Expectations Limitations How can we measure how good a representation is? DINFK Gunnar Ratsch 15. 3. 2022 4 medical_ data_ science Ezurich Computational patient representations Summary of health state of patient one of n possible discrete states a vector describing different physiological aspects Based of heterogeneous data patient state may be represented in lab test and a doctor's note there may be different ways to assess a physiological state Dimensionality reduction" how many dimensions do we need to represent a patient? Expectations: representation faithfully represents patient health state is predictive of future health states or diseases Can it reproduce the original data? Limitations: Loss of some information only relevant for small number of patients Curse of dimensionality & sample sizes Measures of quality low dimensionality interpretability prediction accuracy DINFK Gunnar Ratsch 15. 3. 2022 5 medical_ data_ science _ Ezurich Unsupervised Learning of Health States Unsupervised _Learning Patient Time Series Patient Representation Hospital Data Warehouse d Raw Patient Dataset Medications Diagnoses Clinical Descriptors Procedures Lab Tests t-T+1 Supervised Learning Deep Patient Dataset Patients Features Two major cases: 1. Latent spaces 2. Discrete states Drug Targeting Patient Similarity Disease Clinical Trial Prediction Recruitment Personalized Prescription DINFK Gunnar Ratsch 15. 3. 2022 6 Miotto et al_Scientific Reports_2016 medical_ data_ science F Ezurich SOM-VAE: INTERPRETABLE DISCRETE REPRESENTATION LEARNING ON TIME SERIES Associated Manuscripts Vincent Fortuin, Matthias Hiiser & Francesco Locatello Department of Computer Science, ETH Ziirich Universitatsstrasse 6, 8092 Ziirich, Switzerland {fortuin, mhueser, locatelf} @inf. ethz ch Heiko Strathmann Gatsby Unit; University College London 25 Howland Street, London WIT 4JG, United Kingdom heiko strathmann@gmail com Improving Clinical Predictions through Unsupervised Time Series Representation Learning Gunnar Ratsch Department of Computer Science, ETH Ziirich Universitatsstrasse 6. 8092 Ziirich, Switzerland raetscheinf. ethz ch Xinrui Lyul, Matthias Hiiser', Stephanie L. Hyland', George Zerveas?, Gunnar Ratschl Biomedical Informatics Group; Dept. of Computer Science, ETH Zirich 2 AI Lab, Center for Biomedical Informatics, Brown University ABSTRACT Abstract High-dimensional time series are common in many domains_ Since human cognition is not optimized to work well in high-dimensional spaces, these areas could benefit from interpretable low-dimensional representations. However; most representation learning algorithms for time series data are difficult to interpret: This is due to non-intuitive mappings from data features to salient properties of the representation and non-smoothness over time To address this problem; we propose a new representation learning framework building on ideas from interpretable discrete dimensionality reduction and deep generative modeling: This framework allows uS to learn discrete representations of time series, which give rise to smooth and interpretable embeddings with superior clustering performance. We introduce a new way to overcome the non-differentiability in discrete representation learning and present a gradient-based version of the traditional self-organizing map algorithm that is more performant than the original. Furthermore, to allow for a probabilistic interpretation of our method, we integrate a Markov model in the representation space. This model uncovers the temporal transition structure, improves clustering performance even further and provides additional explanatory insights as well as a natural representation of uncertainty: We evaluate our model in terms of clustering performance and interpretability on static (Fashion-JMNIST data, a time series of linearly interpolated (Fashion-JMNIST images, a chaotic Lorenz attractor system with two macro states, as well as on a challenging real world medical time series application on the eICU data set: Our learned representations compare favorably with competitor methods and facilitate downstream tasks On the real world data. In this work, we investigate unsupervised representation learning On medical time series, which bears the promise of leveraging copious amounts of existing unlabeled data in order to eventually assist clinical decision making: By evaluating on the prediction of clinically relevant outcomes, we show that in practical setting, unsupervised representation learning can offer clear performance benefits over endto-end supervised architectures. We experiment with using sequence-to-sequence (Seq2Seq) models in two different ways, as an autoencoder and as a forecaster; and show that the best performance is achieved by a forecasting Seq2Seq model with an integrated attention mechanism; proposed here for the first time in the setting of unsupervised learning for medical time series. DINFK Gunnar Ratsch 15. 3. 2022 medical_ data_ science _ Ezurich A natural choice: autoencoders Given a signal Xi we want to learn f and 9 f : Rd 7 Rm 9 : Rm 7 Rd Ingvs f(xi) = ei g(ei) = xi VMFUt that minimizes the reconstruction error L = Ilxi xill? Ilxi 94 ( fo(xi)) |2 Lucodr Ucuda f is the encoder function; and 9 is the decoder function; and ei is the representation: f and 9 are jointly optimized (gradient descent) (see Intro ML or Computational Intelligence Lab for extensive introductions) DINFK Gunnar Ratsch 15. 3. 2022 8 medical_ data_ science _ Ezurich Sequential modeling of health states Sequential autoencoder Autoencoder (AE) Most common in unsupervised representation learning The decoder reconstructs the input featureslsignals Sequence-to-sequence AE (S2S-AE) Encoder/decoder structure: sequential NN (e. g-, RNNILSTMIGRU) Input: multivariate time series up to time t Drawbacks: the representation only encodes data from thepast Sequential forecasters (with attention) Sequence-to-sequence forecaster (S2S-F) Same structure as S2S-AE The decoder predicts future_time series from time t+1 Focus on relevant past to predict future S2S-F with Attention (S2S-F-A) Attention helps focus more on past information that is most predictive of the future. Lyu; X, Huser; M: Hyland, S. L, Zerveas, G. and Ratsch, G., 2018. Improving Clinical Predictions through Unsupervised Time Series DINFK Representation Learning: Machine Learning for Health (ML4H) Workshop NeurIPS 2018 Gunnar Ratsch 15. 3. 2022 9 medical_ data_ science _ Ezurich Overviewllntuitions Latent health state Observed data time relevant event DINFK Gunnar Ratsch 15. 3. 2022 10 L Forecaster medical_ data_ science _ Ezurich Review: Recurrent Neural Networks & LSTMs & GRUs AII Recurrent Neural Networks have the form of a chain of repeating modules of neural network. ht ho__ h1 hz_ Recurrent Neural Network ht A A A A Xo X1 Xz Xt See more detailed introductions to RNNs, LSTMs, GRUs, for instance, in class "Deep Learning" _ DINFK Neural Network Layer Pointwise Operation Vector Transfer Pictures from Colah's blog Gunnar Ratsch 15. 3. 2022 11 Concatenate Copy medical_ data_ science _ Ezurich Review: Recurrent Neural Networks & LSTMs & RNN AII Recurrent Neural Networks have the form of a chain of repeating modules of neural network. nt: ht Recurrent Neural Network Single tanh Layer X directly modulates h A tanh A Xt t+ DINFK Neural Network Layer Pointwise Operation Vector Transfer Pictures from Colah's blog Gunnar Ratsch 15. 3. 2022 12 Concatenate Copy medical_ data_ science _ Ezurich Review: Recurrent Neural Networks & LSTMs & RNN AII Recurrent Neural Networks have the form of a chain of repeating modules of neural network. nt: ht Recurrent Neural Network Single tanh Layer X directly modulates h A tanh A Long-Short Term Memory (LSTM) ht Gated Recurrent Unit (GRU) ht 1 The first gate determines what the hidden state forgets ("forget gate"): 2_ The second gate decides which values we'Il update ("input gate") ) 3 The third gate creates a vector of new candidate values. 4_ The last gate produces an output: The forget and input gates ht_) combine into a single "update gate' + another changes tanh tanh 0 tanh simpler than LSTM xt Xt DINFK Neural Network Layer Pointwise Operation Vector Transfer Pictures from Colah's blog Gunnar Ratsch 15. 3. 2022 13 Concatenate Copy medical_ data_ science _ Ezurich Seq2Seq models as autoencoders et Xt Linear layer Linear layer Linear layer Linear layer ht LSTM LSTM LSTM LSTM LSTM LSTM X(t_4) Xlt_A)+1 Xt (t-4) Xt-1 Encoder Decoder DINFK Gunnar Ratsch 15. 3. 2022 14 X(t_^) X(t-A)+l Xt-1 ht_^ ht-4+1 ht_1 medical_ data_ science _ Ezurich Seq2Seq models as autoencoders Encoder and decoder function fe (xi (t-4+1) (t-4+2) (t) (t) xi Xi = ei 9d (eft)) = x(t-4+1), x6t-4+2) x(t) Loss function ZAO (t-j) L(Xt) Ilxi x(t-j)112 Ti L( Xi) L(Xt) t=4 The decoder reconstructs the inputs, i. e., the historical time series of the patient Known from and used extensively in Natural Language Processing DINFK Sutskever et al_NIPS 2014 Gunnar Ratsch 15. 3. 2022 15 medical_ data_ science _ Ezurich Seq2Sea models as forecasters et Linear layer Linear layer Linear layer Linear layer LSTM LSTM LSTM LSTM LSTM LSTM X(t_4) X(t-4)+l Xt Encoder Decoder Teacher Forcing: training procedure, that is used for RNN (NLP; generation tasks): Instead of using the predicted values at time-step t, we passed the ground-truth values for this time-step. Pros: speed-up the training: Cons: limited models Best way: use the combination of teacher-forced values and predicted values: DINFK Gunnar Ratsch 15. 3. 2022 16 More ways, now to use teacher forcing: 1) Deep_Learning Book; chapter 10. 2) Professor Forcing; NIPS 2016. Xt-1 Xt-2 Xt+A-1 Xt-A ht+1 ht+2 ht+A-1 ht+A Xt+A-1 Xt+1 medical_ data_ science _ 1 Ezurich Introduction to the Attention Mechanism Seq2Seq encoder compresses the information into a context vector of a fixed length ~ incapability of remembering long "sentences" Task: Given source sequence X = (11, CTr predict a target sequence y = (y1, YTy _ Encoder maps the input to the hidden states (h1, hTz ) with the help of RNN (bottom) Yt-1 Yt St-1l St Decoder for each i has hidden state 8i = f(si-1, Yi-1, Ci) f-RNN (top) The context vector for the output yi is computed: weighted sum of hidden encoder states (middle) Tc exp (eij_ Ci Qijhj where Qij Tz k=-1 exp (eik) j=1 Ot, T 0t, 3 h1 hz ht The set of Qij are weights defining how much of each source hidden state should be considered tor each output h1 hz h3 ht a(8i-1, hj) = v4 tanh (Wa8i-1 + Uahj) alignment score, eij (image shows bidirectional network also works for unidirectional ones) how well the inputs around position j and the output at position i match: DINFK Gunnar Ratsch 15. 3. 2022 17 Bahdanau; Dzmitry, Kyunghyun Cho, and Yoshua Bengio. "Neural machine_translation_by _jointly learning to_align and translate: at, 1 0t, 2, nzl medical_ data_ science _ Ezurich Seq2Seq forecasters with(attention Attention model Linear layer Linear layer Linear layer Linear layer et-^ et-4+1 et LSTM LSTM LSTM LSTM LSTM LSTM X(t-^) X(t-4)+1 [Xt+1, Ct+1] [Rt+4-1, Ct+4-1] Encoder Decoder The same idea: at timer EUR {t+1,., t+T} during decoding the objective is to produce a context vector Ct a weighted combination of the hidden states of the encoder: DINFK Gunnar Ratsch 15. 3. 2022 18 Ct+l Xt+1 Xt+2 Xt+A-1 Xt+A ht+1 ht+2 ht+A-1 ht+A medical_ data_ science _ Ezurich Experimental Evaluation Setting Patient Patient Time Series Representation Hospital Data Warehouse d Raw Patient Dataset t-T+1 Medications Diagnoses Clinical Descriptors Procedures Lab Tests Unsupervised learning of representations: name nonlinear temporal decoder output attention Deep Patient Dataset Patients PCA past Features AE past S2S-AE past Supervised Learning Task: Healthy discharge (within 24h) S2S-F future S2S-F-A future DINFK Miotto et al_Scientific Reports_2016 Gunnar Ratsch 15. 3. 2022 19 medical_ data_ science _ Ezurich Experimental evaluation Data: Multivariate ICU time series (d=94) from the Philips eICU_dataset (vital signsllab test results, ~20'000 patients) Length of encodedlpredicted time series: 12 h (resolution: 1 samplelh) Embedding dimension: 94 (compression rate: 12. 1) Supervised learning method: LSTM with one layer: 24h Discharge 0. 45 24h Discharge AUPRC AUROC 0. 40 1 0. 35 PCA rep AE rep. S2S-AE rep. S2S-F rep_ S2S-F-A rep 0. 436 = 0. 01 0. 811 = 0. 004 0. 824 = 0. 002 0. 824 E 0. 003 0. 825 + 0. 003* 0. 825 + 0. 003* Supervised (LSTM-3) Supervised (LSTM-1) LSTM-1 + PCA rep. LSTM-1 + AE rep. LSTM-1 + S2S-AE rep. LSTM-1 + S2S-F rep. LSTM-1 + S2S-F-A rep. 0. 471 = 0. 005 0. 474 + 0. 006 0. 477 + 0. 006* 0. 48 + 0. 007 0. 30 0. 25 1% 5% 10% 25% % of labeled data 50% 100% Lyu; X,, Hiser; M:, Hyland, S. L, Zerveas, G. and Ratsch, G. 2018. Improving Clinical Predictions through Unsupervised Time Series DINFK Representation Learning: Machine Learning for Health (MLAH) Workshop NeurIPS 2018 Gunnar Ratsch 15. 3. 2022 20 medical_ data_ science _ Ezurich Discussion (latent space) S2S models reduce input time-series to low dimensional embeddings and still achieve better performance than using the raw features. S2S-F-A outperforms the others when sufficient data is available. When labeled data is limited, 9 deep unsupervised representation shallow supervised learning can outperform deep supervised learning: Ker Science; Unl Sqctxl Xinrui Lyu Matthias Huser et al. Jsse DINFK Let Scetce Gunnar Ratsch 15. 3. 2022 21 medical_ data_ science _ Ezurich Introduction to the Transformer model Transformer is the encoder-decoder architecture with self-attention: The key component of the Encoder is Self-Attention mechanism. Layer normalization Input data: X e Rnxd Encoder overview: Q-queries, K keys, V values: linear projections of the input data X, d dimension: # QKT Latent Representation softmax )V Residual 1 Va connections Add & Normalize Feed Forward Feed Forward Projection layer / Add & Normalize Self-Attention Attention weights Positional embeddings POscoDMG Self attention Positional Embeddings is used to learn the order in the sequence. Residual connections mitigates vanishing gradients. Layer normalization helps with numerical stability: POSITIONAL ENCODING sin if i = 2k PEpi coS if i = 2k+1 EMBEDDINGS X1 Pictures from The Ilustrated transformer_Jay Alammar DINFK Vaswani, Ashish, et al. "Attention is allyou need"Advances in neural information processing systems 30 (2017). Gunnar Ratsch 15. 3. 2022 22 medical_ data_ science _ Elzurich Monitoring Patient State in Intensive Care PRESCRIBING TREATMENT ICU PATIENT EHR CLINICIAN OUR WORK DATA RECORDING PREDICTING PATIENT EVOLUTION ML MODEL DINFK Yeche, Hugo; Kuznetsova Rita at al, HiRID-ICU-Benchmark A Comprehensive Machine Learning Benchmark on High-resolution ICU Data, 35th Conference on Neural Information Processing Systems (NeurIPS 2021) Track on Datasets and Benchmarks_ Gunnar Ratsch 15. 3. 2022 23 medical_ data_ science _ Ezurich Current Limitations of Existing EHR Datasets EXISTING ICU DATABASE AnsterdamUMC Downstream Tasks Pipeline Reproducibility No preprocessing leads to unfair comparison across work due to irreproducible: MIMIC-III eICU Few Database contain labels When label is provided, task are not clinically relevant (eg: Mortality) Different label definition across work HiRID Data splitting Data processing Label definition Use of external data Our work tackles both issues for HiRID database DINFK https:Ilgithub comlratschlab/HIRID-ICU-Benchmark Gunnar Ratsch 15. 3. 2022 24 medical_ data_ science _ Ezurich Define a Variety of Clinically Relevant Tasks X Circulatory Failure Respiratory Failure Kidney Function Length of Stay Mortality Phenotyping Predict whether a patient is going to experience a circulatory failure in the next 12 hours Predict whether a patient is going to experience a respiratory failure in the next 12 hours Predict a patient average urine production in the next 2 hours Predict at a given time the remaining length of stay of a patient in the ICU Predict whether a patient is going to expire in the ICU after 24 hours of stay: Predict a patient admission group (APACHE) after 24 hours of stay DINFK Gunnar Ratsch 15. 3. 2022 25 medical_ data_ science _ Ezurich Benchmarking SOTA Machine Learning Approaches Task ICU Mortality AUPRC (1) AUROC 60. 3 + 1. 6 90. 0 = 0. 4 60. 0 = 0. 9 90. 3 = 0. 2 60. 2 + 1. 1 89. 7 = 0. 4 61. 0 = 0. 8 90. 8 = 0. 2 Patient Phenotyping B-Accuracy 39. 2 +2. 1 39. + 1. 2 41. 6 +23 42. 7 +1. 4 Benchmark codeldata: https:Ilgithub. comlratschlab/HIRID-ICU-Benchmark Metric GRU LSTM TCN Transformer Task Circulatory failure Respiratory failure Metric AUPRC AUROC (1) AUPRC AUROC GRU 36. 8 = 0. 5 90. 7 + 0. 2 59. 2+0. 3 70. 1 =0. 2 LSTM 32. 6 + 0. 8 89. 9 + 0. 1 56. 9 +0. 3 68. 2 +0. 3 TCN 35. 8 + 0. 6 90. 5 + 0. 1 58. 9 + 0. 3 70. 0 + 0. 2 Transformer 35. 2+0. 6 90. 6 + 0. 2 59. 4 +0. 3 70. 1+0. 2 Kidney func. Remaining LOS MAE MAE 0. 49 = 0. 02 0. 50 + 0. 01 0. 50 + 0. 01 0. 48 + 0. 02 60. 6 = 0. 9 60. 7 + 1. 6 59. 8 EUR 2. 8 59. 5+2. 8 TCN Temporal Convolution Networks (cf. Image analysis lecture) Transformers explained next lecture Yeche Hugo, Kuznetsova Rita et al. DINFK Gunnar Ratsch 15. 3. 2022 26 medical_ data_ science _ Ezurich Generative Models Generative models are probabilistic models of high-dimensional data. Describe the probabilistic process of generating an observation: The emphasis is on capturing the dependence between the dimensions: Provide a way of generating new datapoints: Historically, generative modelling was considered to be a subfield of unsupervised learning: Usage of generative modelling: Representation learning; density estimation, data compression etc. Latent Variable Models is a type of Generative models_ Specify the generative process in terms of unobserved/latent variables and the transformation that maps them to the observation: Trained with maximum likelihood (usually with some approximations) Easy to incorporate prior knowledge structure into models; fast generation: Need to use approximate inference or restricted models. DINFK The Deep Learning Series Lecture Series 2020, Deep Mind Gunnar Ratsch 15. 3. 2022 27 medical_ data_ science _ Ezurich Generative Adversarial Networks Model: A neural net that maps noise vectors to observations Training: use the learning signal from a classifier trained to discriminate between samples from the model and the training data Pros Can generate very realistic images Real images Conceptually simple implementation Fast generation 1 Generator L Cons Cannot be used to compute probability of observations "Mode collapse' Models ignore regions of the data distribution Training can be unstable and requires many tricks to work well Sample { | Discriminator Sample { 0 (More details in the lecture on privacy to generate realistic artificial data) DINFK The Deep Learning Series Lecture Series 2020, Deep Mind Picture from Google_Developers Gunnar Ratsch 15. 3. 2022 28 medical_ data_ science _ Ezurich Latent Variable Models 3 Variational Autoencoder Consider dataset X = {x}i1, N consisting of N i. i. d. samples. We assume that the data are generated by some random process, involving an unobserved continuous random variable z ~ N (0, I): X~ pe (xlz), Z where pe(zlx) is unknown. The lower bound on the marginal likelihood is: LvAE [ 9(zl) log pe (xlz)dz KL(qo(z/x)llp(z)), reconstruction loss "regularizer"Iprior p(z) LvAE max 0, $ 96(z/x) & pe(xlz) is modelled by neural networks: q6(z/x) ~ N(0o, EUR 02) encoder, po (xlz) ~ N(ue, o83 decoder. 06, 02, 1o, 03 2 the output of the neural networks: DINFK DP: Kingma, M. Welling, Auto-Encoding Variational Bayes I/ International Conference of Learning Representations 2014_ X N 2 2 Encoder 9zlx) Decoder P(xlz) Data: X Reconstruction: * Neural Network Perspective [source] Gunnar Ratsch 15. 3. 2022 29 medical_ data_ science _ Ezurich Towards interpretable health state representations Idea: Use self-organizing maps to encourage interpretable neighborhood relationship in latent space and smoothness over time Renal Idysfunction Desirable properties Discretellow dimensional Smooth over time Expressive Interpretable Cardiac dysfunction Healthy, https 1 /en wikipedia org/wiki 'Self-organi zing map# /media File Somtraining vg Vincent Fortuin; Matthias Huser; Francesco Locatello, Heiko Strathmann; and Gunnar Ratsch, 2019. SOMVAE: Interpretable discrete representation learning on time series: International Conference on Learning Representations (ICLR) 2019. Gunnar Ratsch 15. 3. 2022 30 DINFK medical_ data_ science _ Ezurich SOM-VAE model for discrete time series representations input ^w encoder latent encodings Markov model decoder reconstruction t+1 2 e t 2 2 t+1 4 P(z t+1l2. 4) q self-organizing map t Ze L(xt-1 xt '22 9, &t e) LsoM-VAE (2t 2 24, 18) +yLtransitions rt_1, xt) + t Lsmoothness rt_1 2 xt) LsoM-VAE(T, Sq, Te) Lreconstruction (1, iq, Ze) + a Lcommitment (w) + 8 Lsom (x) Model is jointly optimized (gradient descent wl special handling of discrete states) DINFK Vincent Fortuin; Matthias Hiser; Francesco Locatello, Heiko Strathmann; and Gunnar Ratsch, 2019. SOMVAE: Interpretable_discrete representation learningon_time series: International Conference on Learning Representations (ICLR) 2019. Gunnar Ratsch 15. 3. 2022 31 N medical_ data_ science _ Ezurich Health state representations on 2D SOM-grid Method score 6 score 12 score 24 k-means 0. 0411 + 0. 0007 0. 0384 + 0. 0006 0. 0366 = 0. 0005 SOM-VAE 0. 0407 = 0. 0005 0. 0376 + 0. 0004 0. 0354 = 0. 0004 SOM-VAE-prob 0. 0474 + 0. 0006 0. 0444 + 0. 0006 0. 0421 + 0. 0005 Performance comparison of our method wth and_without Markov model (SOM-VAE-prob and SOM-VAE) against k-means in terms of normalized mutual information. Dynamic endpoints are the maximum of the physiology score within the next 6, 12 or 24 hours (6_hours, 12_ hours, 24 hours): Each method is used to fit 64 clusters. Shown are means and standard errors over 10 runs. DINFK Vincent Fortuin; Matthias Hiser; Francesco Locatello, Heiko Strathmann, and Gunnar Ratsch, 2019. SOMVAE: Interpretable discrete Gunnar Ratsch 15. 3. 2022 32 representation learning on _time series: International Conference on Learning Representations (ICLR) 2019. medical_ data_ science _ Ezurich Health state representations on 2D SOM-grid 12 10 (a) k-means (b) VQ-VAE (c) SOM-VAE (d) Patient trajectories Color of SOM-cells: Average dynamic APACHE score in the next 24 hours. Higher APACHE score is associated with increased severity of patient health state. vQ-VAE: Van den Oord et al., 2017 (https Ilarxiv orglabs/1711. 00937) Vincent Fortuin Vincent Fortuin, Matthias Hiser; Francesco Locatello, Heiko Strathmann, and Gunnar Ratsch, 2019. SQM-VAE: Interpretablediscrete Gunnar Ratsch 15. 3. 2022 33 representation learning on time series: International Conference on Learning Representations (ICLR) 2019. DINFK medical_ data_ science _ Ezurich Health state representations on 2D SOM-grid 0. 2 Ston Eno * Stant End: 016 0 00 0. 06 WW 0 02 Figure 3: Illustration of two example patient trajectories in the SOM grid of T-DPSOM49. One patient died (red), while the other was discharged alive from the ICU (green). Superimposed is a heatmap that displays the mean APACHE score of all time points assigned to each cluster: We observe qualitative differences in the trajectories of the dying and the surviving patient: For each time series, we also show the assigned probabilities to the discrete patient health states using a blue color shading: In this work; we will separate the representations into organ systems and manually annotate different areas of the "map" using medical knowledge in collaboration with intensive care specialists. Laura Manduchi; Matthias Huser; Julia Vogt; Gunnar Ratsch, Vincent Fortuin, DPSOM: Deep Probabilistic Clustering with Self-Organizing Maps, ACM-CHIL, 2020 DINFK medical_ data_ science _ 1 Ezurich Contrastive Learning Contrastive learning is a machine learning technique used to learn the general features ofa dataset without labels by teaching the model which data points are similar or different: [Source] Vector Representation 2, Maximize agreement Vector Representation z g() Dense Layer Dense Layer Dense Layer Dense Layer Dense Layer Dense Layer Intermediate Representation h CNN Neural Network Encoder fC) CNN Neural Network Encoder fC) Similar samples < Similar downstream task label Positive Pairs X Data Augmentation x New challenges for Contrastive Learning DINFK Gunnar Ratsch 15. 3. 2022 35 medical_ data_ science _ 1 Ezurich Contrastive Learning for Time Series Contrastive learning is a machine learning technique used to learn the general features ofa dataset without labels by teaching the model which data points are similar or different: [Source] Contrastive Learning Learn a representation of a patient state at a give time t0 t-th Patient stay Similar samples < Similar downstream task label New challenges for Contrastive Learning Hugo Yeche et al Yeche, Hugo, et al. 'Neighborhood contrastive learning applied t0 online patient monitoring: International Conference on Machine Learning. PMLR, 2021. Gunnar Ratsch DINFK 15. 3. 2022 36 medical_ data_ science _ Ezurich Challenges in Using Contrastive Learning for ICU monitoring 2N 'exp pi pvli) TT CCL log M i=1 kzi exp Pi pj/t, Alignment Normalization How to construct different views outside of computer vision [1]? Should all "negatives" be treated the same [2]? CV ICU CV ICU Diversity among samples Humanly understandable signal Complex modality Strong similarities among contiguous samples i. i. d distribution of samples Balanced multi-class downstream task multiples samples from a single patient (non ii. d) Usually imbalanced downstream task Relies on strong data augmentation Limits usage of data augmentation Clear hierarchy among negative samples Uniform Normalization [1] Tian et al: (2020) [2] Wang et al. (2020) Gunnar Ratsch 15. 3. 2022 37 DINFK medical_ data_ science _ Ezurich Preserve hierarchy in the data with Neighborhood Contrastive Loss How can infuse prior Redefine Contrastive knowledge without relying Loss through the lens of on data augmentation? Neighborhood Two samples share the same neighborhood if they share some predefined attributes: Formally Examples Supervised: n(i, k) = 1if Ti and Tk share attributes 0 else ny(i, k) = 1if yi Yk Temporal nw (i, k) 1ifli kl < W and Si Sk Wang et al:. (2020) DINFK Gunnar Ratsch 15. 3. 2022 38 medical_ data_ science _ Ezurich Neighborhood Contrastive learning Objective ND NA Aligning neighbors Discriminating neighbors 2N Zi zvli) ZkIn(i, k) = 1 2kIn(i, k) = 0 push away pull towards 2N exp m Pi p)/v) log keN(i) exp (Pi 9k/t) i=1 ~l CNA exp (Pi_ pin /t) = log NG)I M i=1 leN(i) kzi exp (pi aklt) = 3 LNCL aLNA + (1 _ a)LND DINFK Gunnar Ratsch 15. 3. 2022 39 CND medical_ data_ science _ 1 Ezurich A Unifying Framework Method U n(, CL 1. 0 0 nw SACL (Cheng et al,, 2020) 0. 0 +00 Tw CLOCS (Kiyasseh et al,, 2020) 1. 0 +00 nw SCL (Khosla et al,, 2020) 1. 0 NA ny We explore two cases NCL: Unsupervised, we call NCL(nw) where: n = nw; W EUR J0, +o [; & EUR J0, 1[ Supervised, we call NCL(ny) where:n ny ; & EUR ]0, 1[ Task Sepsis onset prediction AUROC (in %_ Metric AUPRC (in % _ Utility (x100) Linear MLP Head Linear MLP Linear MLP Seq2-Seq-AE 7. 0 + 0. 3 Seq2-Seq-AE-forecast 6. 6 = 0. 3 CL 7. 9 = 0. 4 SACL (Cheng et al,, 2020) 6. 5 =0. 3 CLOCS (Kiyasseh et al,, 2020) 7. 1 +0. 5 NCL(nw ) (Ours) 8. 2 = 0. 4 7. 8 + 0. 4 7. 3 +0. 3 95 + 0. 4 7. 6 =0. 3 7. 3 +0. 4 93 + 0. 5 77. 1 +0. 5 78. 1+0. 6 75. 8 + 0. 9 76. 9 +0. 5 78. 2 =0. 3 80. 2 + 0. 4 73. 0 = 1. 2 75. 3 +0. 8 77. 2 +05 78. 8 = 0. 4 78. 8 = 0. 3 80. 7 = 0. 3 26. 8 = 1. 0 23. 5 + 1. 5 26. 2 + 0. 8 20. 5: 2. 5 23. 0 =1. 1 27. 2 = 1. 0 27. 2 + 1. 0 23. 8 = 1. 2 29. 7 + 1. 0 24. 2+1. 1 25. 8 + 0. 9 30. 2 = 1. 0 End-to-End SCL (Khosla et al,, 2020) NCL(ny ) (Ours) 7. 6 = 0. 2 6. 7 =0. 6 10. 0 = 0. 5 8. 1 +0. 4 6. 0 + 0. 5 10. 1 + 0. 3 78. 9 + 0. 3 73. 1 +1. 7 80. 3 + 0. 4 78. 8 + 0. 4 70. 0 + 1. 9 80. 8 = 0. 2 27. 9 +0. 8 20. 2 +2. 7 32. 6 + 1. 0 27. 5 1. 0 20. 6 = 1. 7 31. 9 + 0. 9 Experiments on MIMIC-IIl and Physionet 2019 datasets. DINFK Gunnar Ratsch 15. 3. 2022 40 medical_ data_ science _ Ezurich Summary & Take Home messages Representation learning is a recently developed, powerful tool to learn integrative computational summaries of observed data. Health state data is one interesting application where we assume that the patients physiological health state can be accurately represented in vector form: Autoencoders and forecaster models learn vector representations of past data that are predictive of the past and future, respectively: Generative models are an important tool for finding additional representations and to generate realistic data Contrastive learning can improve learning representations of time series DINFK Gunnar Ratsch 41 | 1 | 5.3% |
| Other values (9) | 9 |
Length
Histogram of lengths of the category
| Value | Count | Frequency (%) |
| the | 9891 | 5.0% |
| to | 4720 | 2.4% |
| of | 4429 | 2.2% |
| and | 4392 | 2.2% |
| a | 4189 | 2.1% |
| in | 2947 | 1.5% |
| you | 2864 | 1.4% |
| is | 2433 | 1.2% |
| that | 2018 | 1.0% |
| it | 1988 | 1.0% |
| Other values (16874) | 158872 |
Most occurring characters
| Value | Count | Frequency (%) |
| 389923 | ||
| e | 96883 | 7.1% |
| t | 74815 | 5.4% |
| o | 64465 | 4.7% |
| a | 63916 | 4.7% |
| n | 57635 | 4.2% |
| i | 57030 | 4.2% |
| s | 53881 | 3.9% |
| r | 46889 | 3.4% |
| h | 40697 | 3.0% |
| Other values (91) | 426718 |
Most occurring categories
| Value | Count | Frequency (%) |
| Lowercase Letter | 793147 | |
| Space Separator | 389923 | |
| Uppercase Letter | 89266 | 6.5% |
| Other Punctuation | 39866 | 2.9% |
| Control | 37500 | 2.7% |
| Decimal Number | 11857 | 0.9% |
| Dash Punctuation | 3538 | 0.3% |
| Math Symbol | 2696 | 0.2% |
| Open Punctuation | 2282 | 0.2% |
| Close Punctuation | 2187 | 0.2% |
| Other values (5) | 590 | < 0.1% |
Most frequent character per category
Lowercase Letter
| Value | Count | Frequency (%) |
| e | 96883 | |
| t | 74815 | 9.4% |
| o | 64465 | 8.1% |
| a | 63916 | 8.1% |
| n | 57635 | 7.3% |
| i | 57030 | 7.2% |
| s | 53881 | 6.8% |
| r | 46889 | 5.9% |
| h | 40697 | 5.1% |
| l | 32839 | 4.1% |
| Other values (18) | 204097 |
Uppercase Letter
| Value | Count | Frequency (%) |
| I | 8379 | 9.4% |
| A | 7347 | 8.2% |
| T | 6818 | 7.6% |
| S | 6025 | 6.7% |
| O | 5894 | 6.6% |
| N | 5893 | 6.6% |
| E | 5609 | 6.3% |
| L | 4566 | 5.1% |
| R | 4552 | 5.1% |
| M | 3807 | 4.3% |
| Other values (16) | 30376 |
Other Punctuation
| Value | Count | Frequency (%) |
| . | 17012 | |
| , | 9760 | |
| ' | 5912 | 14.8% |
| ! | 1876 | 4.7% |
| ? | 1759 | 4.4% |
| : | 1154 | 2.9% |
| " | 1080 | 2.7% |
| ; | 546 | 1.4% |
| / | 298 | 0.7% |
| # | 235 | 0.6% |
| Other values (4) | 234 | 0.6% |
Decimal Number
| Value | Count | Frequency (%) |
| 2 | 2325 | |
| 0 | 2201 | |
| 1 | 2170 | |
| 3 | 997 | |
| 4 | 899 | 7.6% |
| 5 | 820 | 6.9% |
| 6 | 687 | 5.8% |
| 9 | 642 | 5.4% |
| 8 | 587 | 5.0% |
| 7 | 529 | 4.5% |
Math Symbol
| Value | Count | Frequency (%) |
| = | 2359 | |
| + | 174 | 6.5% |
| | | 85 | 3.2% |
| ~ | 53 | 2.0% |
| < | 14 | 0.5% |
| > | 11 | 0.4% |
Open Punctuation
| Value | Count | Frequency (%) |
| ( | 1699 | |
| [ | 550 | 24.1% |
| { | 33 | 1.4% |
Close Punctuation
| Value | Count | Frequency (%) |
| ) | 1690 | |
| ] | 455 | 20.8% |
| } | 42 | 1.9% |
Control
| Value | Count | Frequency (%) |
| 37498 | ||
| 2 | < 0.1% |
Modifier Symbol
| Value | Count | Frequency (%) |
| ` | 6 | |
| ^ | 6 |
Final Punctuation
| Value | Count | Frequency (%) |
| ’ | 6 | |
| ” | 1 | 14.3% |
Space Separator
| Value | Count | Frequency (%) |
| 389923 |
Dash Punctuation
| Value | Count | Frequency (%) |
| - | 3538 |
Connector Punctuation
| Value | Count | Frequency (%) |
| _ | 546 |
Currency Symbol
| Value | Count | Frequency (%) |
| $ | 24 |
Initial Punctuation
| Value | Count | Frequency (%) |
| “ | 1 |
Most occurring scripts
| Value | Count | Frequency (%) |
| Latin | 882413 | |
| Common | 490439 |
Most frequent character per script
Latin
| Value | Count | Frequency (%) |
| e | 96883 | 11.0% |
| t | 74815 | 8.5% |
| o | 64465 | 7.3% |
| a | 63916 | 7.2% |
| n | 57635 | 6.5% |
| i | 57030 | 6.5% |
| s | 53881 | 6.1% |
| r | 46889 | 5.3% |
| h | 40697 | 4.6% |
| l | 32839 | 3.7% |
| Other values (44) | 293363 |
Common
| Value | Count | Frequency (%) |
| 389923 | ||
| 37498 | 7.6% | |
| . | 17012 | 3.5% |
| , | 9760 | 2.0% |
| ' | 5912 | 1.2% |
| - | 3538 | 0.7% |
| = | 2359 | 0.5% |
| 2 | 2325 | 0.5% |
| 0 | 2201 | 0.4% |
| 1 | 2170 | 0.4% |
| Other values (37) | 17741 | 3.6% |
Most occurring blocks
| Value | Count | Frequency (%) |
| ASCII | 1372842 | |
| Punctuation | 8 | < 0.1% |
| None | 2 | < 0.1% |
Most frequent character per block
ASCII
| Value | Count | Frequency (%) |
| 389923 | ||
| e | 96883 | 7.1% |
| t | 74815 | 5.4% |
| o | 64465 | 4.7% |
| a | 63916 | 4.7% |
| n | 57635 | 4.2% |
| i | 57030 | 4.2% |
| s | 53881 | 3.9% |
| r | 46889 | 3.4% |
| h | 40697 | 3.0% |
| Other values (86) | 426708 |
Punctuation
| Value | Count | Frequency (%) |
| ’ | 6 | |
| “ | 1 | 12.5% |
| ” | 1 | 12.5% |
None
| Value | Count | Frequency (%) |
| ï | 1 | |
| â | 1 |
| source_doc_filename | source_doc_id | source_doc_domain | document_text | |
|---|---|---|---|---|
| source_doc_filename | 1.000 | 1.000 | 1.000 | 1.000 |
| source_doc_id | 1.000 | 1.000 | 1.000 | 1.000 |
| source_doc_domain | 1.000 | 1.000 | 1.000 | 1.000 |
| document_text | 1.000 | 1.000 | 1.000 | 1.000 |
A simple visualization of nullity by column.
Nullity matrix is a data-dense display which lets you quickly visually pick out patterns in data completion.
| source_doc_filename | source_doc_id | source_doc_domain | document_text | |
|---|---|---|---|---|
| 0 | ASRnlp_law_lecture_week_1_v_2_c_transcription_1.txt | 5e311e20-4bb | ASR | Welcome everyone! This is natural Language processing for Law and Social Science. Thanks for joining remotely today. It still is a bit up in the air how we will do the hybrid verses in person versus Zum format. This term, you hear, I'm a little stuffy today. This is true. Avoid nineteen case I Caught it from my daughter who caught it in daycare. It's very mild so I hope if any of you catch it, it's not worse than this. It's definitely manageable. You can see I'm here to teach even though I have it so it's not that bad. I'm a little congested and we might end a bit early today. I Hope that's all right, but going forward. I Would like to do the course hybrid where we have some impression meetings at least, but before text money you'll hear from me about that. Thank you for all of you who filled in the students survey. There was a broad agreement that we should have an online aspect or at least should be recorded so well. we will work with that. So I have a few slides to introduce the course and then we'll have a chance to answer any sex questions about the format. So this course is a applied Natural Language Processing. It's not a course where we will just start with different texts, data tools, or different help models and learn how to come them up in there. We care just as much about the applications of those methods in the law and in social science and this is in the news all the time. Here is an example from a recent legal product called Clarity which uses in all tools to to analyse contracts and for example terms of use to highlight different clauses that are unusual. You also hear these really exciting ideas such as the World's First Robot Lawyer I'm excited about this, I think everybody should you. I think that this technology is is improving rapidly and dramatically and there is scope for many interesting things happening in the law and inoup world. But there also is a lot of hype and we will take a skeptical view of of these strong statements such as the World's First Robot Lawyer And in turn I Think that while there is a lot of interesting about tools coming about for law and social science and other applications, I Do not think we're close to having a judge to be replaced by a contact. some other reasons to be skeptical or to be concerned about the arrival of these legal inopetuls is that they can be biased So northwest. One of the classic examples is different languages have different degrees of being rendered having notions of gender, mail and female in the language and if you translate from English such as she is a doctor he is a nurse to Turkish which does not have notions of gender pronouwns and then you translate back the gender switches so basically they they have since fixed to this in google translate but it used to be where if you to see as a doctor translated to Turkish and then translated it back it will change to him as a doctor just because similarly he as a nurse would be transformed she as a nurse and this is just because. Theories this basis this statistical correlation in language where doctors tend to be male and nurses tend to be female and statistical language models and translation systems will capture that bias. These issues are based. the language models are as the technology comes more powerful, these issues become more intense to get more intense benefits but also more more intense risks and good. It is now a few years a few years old but this is language whose hole it came out in the Tousadand nineteen that could. It was basically among many other things a fake news production engine and it could produce a lot of context appropriate prose. So you can imagine to know Twitter and email. can the news being filled with all this machine produced speech that would drown out all other speech and I think that those those concerns are still relevant. but now that got two has been out for it for three years and there's an even better version called just There that's been out for a year and we have not seen the internet employee. That means that maybe it was not as bad as we thought and so in this course we want to know. Can we take legal Gptto illegal? Get there to help judges in their work? So this is the course. It's natural Language processing for law and Social science and our engineering goals are doing these kind of two pieces. This course we're going to develop skills in applied natural language processing which will include machine analysis, interpretation, generation of documents and those could be on news articles or contracts or judicial opinions or political speeches. And we want to also take a social science approach where we do not just care about sequencing language in different ways, we care about relating it to attend data, and to understand the social forces that underlie these documents. What are their determinations and what are their outcomes and so you knowsome. Of the examples that we will frequently come back to are: what are the motivations for judges? Can we analyze judicial opinions and see what's driving their decisions? Can we look at the writings and the speeches of politicians and to the end where they're coming from And I Think this is kind of the broader agenda or the goal in this research area is Knpowders language matter in society in human relationships s and what can help do to help understand that? So what we will do. We're going to read text documents as data so there's you know many ways to do this and will go over many of them. We will use supervise learning techniques for dimension reduction, topic modeling, groups interpretation, supervise learning for text regression and text classification can be predict from a speech. Is this form a left wing politician or a right wing politician will get at World embeddings, document embeddings, a lot of exciting technologies there for producing these learned representations of language of words and concepts in geometric spaces. and towards the end will get into disclosure analytics. So this is where the linguistic side of natural language processing, cynicism, and a summarization question answering I'm checking. These are all really relevant to legal applications for example. So some course logistics or beating times that will be two even fourteen to it in an sixteen so we'll have a ten minute break in the middle going back to what I started mentioning at the beginning. These are the survey results for the the course format and there were only a handful of you who would be register if there only online and everybody else wanted some online aspect or the indifferent. The based on these surveys we will certainly have a online component. like everything in the class will be durable online. but I'm hoping that we can have some impression component as well. so there's a question in the chat I have the the chat year so I'm sorry but it in general how keep track of the chat so you can always ask questions three were asked to. We need to have knowledge about law I said are to be a good in the class. The answers no no not at all so you do not to have any knowledge of it, you need to open to learning about it. So if you have some interest in social science applications or legal applications of help it will make the class much more enjoyable. but there will be no substantive aspects of health or social science that will be tested. and so given that we will have this important online component to the class I Want to make the most of the hybrid learning. The lectures will be recorded by but in some view that contacted me about different special cases which is fine but if you can it's going to make the class more fun and and more better for everyone if everybody comes in if you're going to be absent let me or the tea now and if you have questions or comments from online you can just type them in the chat as you as the doing or you can use the the raise and front in which a no monitor so help asks are people who either know Pytha nor beeligible for this course so aim going to get to that in a little bit. But the short answer is if you've never seen Python I do not recommend taking this course, it will be too difficult. I mean you can try to stay for the first two weeks and see if it's manageable for you. but in previous versions of the class, people who were new to Python and people who had never done any text analysis it was frustrating for them. and so I do not recommend the course for anyone who's sure asked and well tell you that as some emails if you'regoing to be absent for a lecture, a email email after the tea to let her know and if you have to miss a significant number of courses the email of cause you might have to do an additional assignment. So ya so relax. If you're anyone who is kind of the new to Python or has not done any ex data and turnout sure let me know you can talk about it so avoid asks and can homework only be permitted in Python or can we draw also try to this in or sure yeah you're welcome to try it in our for me I should. I wouldbe great if if anyone is interested in converting the course materials to war that would actually be a great course project so we can arrange to get extra credit for that report. asks what's the do registration deadline? There is not an official one I'm not sure exactly. I think it's varies a bit by department but I do not have an official de registration that line. If you're going to be just for for our grading purposes, it's better to do it before the first response essay which is like six weeks in five or six weeks in because others take a long time for grading and so I would rather you deregister before that. So I would say I think by five or six weeks you should know whether you will stay or not. So smart. Asks if we attend the lecturers and only watch the videos. there will be additional projects yes, so mandatory. The live attendance is mandatory and so if you're going to just watch the videos then you have to do another assignment. but I have not decided what that is yet. Okay so yes, so this is related to newly keep track of course participation through in class activities. So young asks, do you recommend someone who also general machine her knowledge but just to experience with help. If you're actually pretty comfortable machine learning with Python then this course actually wopolity Fine. So if I think that if you're doing the first two assignments, the first two home work assignments and they're not taking you a long time to do if you can finish them within of hours then then your on track. but it mainly do not recommend it. I mean it, if you're quite new to Python then I do not recommend it if you have some machine learning experience than that's good. But as I said some text analysis or snap experiences is recommended. So we have course syllabus I've already sent by email and I be in oppose to league to it again so also asks why this course worked for people who intend buillier of it judge course at so if you took my course in the spring and you l in off if you've done if you finish the assignments of the course in the spring then this course will find freedom so there's there's a little bit of overall. So I would say that he saw in the fall course it was say in the fall course ably report judge it would be fine as a prerequisite for this course. If you've done that then this should be fine for you. So those links are a bit is going to change to it screenshare to the syllabus so I just pose I did a link to this in the home Here's my contact details. Here's area's contact details: the lecture schedule that sessions which I'll cook you a bit in a second but those are at ten a man on Fridays they're not mandatory but these will. They also be recorded and Afro will go over the example coding notebook for the week and then also go over the previous week's homework. This is our daughter's the structure, All the slides including Iardaploi today slides here there in the slides thotfolder notebooks. These are the simple notebooks for learning the material and so before before doing the assignment. You should read the notebooks so you can see. You can kind of skim though you can read through these, ensure you understand them and everything that's in the homework which is here under homework. The homeowners will follow will have similar content to what's in the notebook so you can see we fill in part of the notebook and then you have to add something in a new column text which contains the lower case, title and lead. So here's lead. here's title and so No Nights is an example and here you can just like to nowtype lower so thiswill be how to get lower case. So this is just to show you that these the notebooks and the homework are designed off for self motivated learning of all the coding as aspects of the course. so find asked how do the homers omissions work? So there's there's a homework every week and so it's like this: homework Here you download this Jupiter notebook and fill it out and finish it and then you upload it to add you that it's not working yet but it's going to be on the course model. There's a submission system or using coal ufous up load it on the website and going to be due. The homework are done on thousands but the first homework is done next. Thursday So I can not actually show you if you scroll down so everything you need is going to be highlighted here. So for example, do this week. next week the homework one is done on Thursday fin is that what you were asking? I'm going to come back to that as let me know if you have some other questions about it. So here's me. I'm going to put in the correct system still working on this but camera acts are all homework mandatory if you want it mean you lose point if you do not do them but they are. The homework are a completion grade so you know we're not grading them. We're not grading all the answers but if will check that you did like you tried to do every piece. and if you say you get full credit and basically so in terms of the grading it's thirty percent for the programming homeowners and I do not go eleven homework so mistake. three points per homework or that's thirty percent and so for every lecture will have the slides. I'm going to post the links to the recordings here so like after it today, you'll be able to get the recording for for everyone here. there's going to be a tea session on free, there will be a recording link heiress about. So unique asks what can we think, what the response essays are. Can you be a bit more specific? like do you want to see an example? Okay, well get to that. We'll get to that next week. It may be the We attributes. You do not have to do one of those for a time until a month from now. time is talking about the response essays. Whether's some information here I'll provide some example response essays from previous previous years, but it was not going to get into that into detail today because it's a more complex topic but you can read about them here. But basically what it is is reading a paper one of these papers in writing a response as I about it. Like a review here, I have a link to the bibliography of references. So these are all of like the materials that the slides are based on so you do not. Someone of these are required readings but it's worth skimming through this just to see where things come from. and if you are interested and you want to to go back to add to fill in your knowledge from the slides then you can read these the other. the other required assignment is there is going to be there required readings for example in week for you have to read one of these papers and then we'll do an inner class activity about them but it's going to be. We will form groups and do short presentations to each other about these press but I'm going to provide more information to that in week there before we do that. So the the three pieces of the assessment are the homework on the coding homework which I showed you the response essays which I mentioned or reading a paper and writing a report about it and in third there's a end of there's an end of the course assignment and its in the you So we would call them an exam but I think here you would just say it's an end of course assignment where you have a few days to do an assign. For those of you who were in my class in the fall you know this is like it's a questionbasicly a question about every like sure in some questions about the required readings and so that the end of the course assignment is one the things that we will cover in the lecture are covered their of sotthat's how the course will be assessed I'm going to cover some of that information again now just in the slides so it mentioned awards the is the first that session will be on fairly area's here as well. After do will introduce yourself sure here one man after I'm a packed student at an centre and I hope to see you in the first session. So in these is sessions it's what would expect far will go over the notebooks they code note books from the bathtub and then the last week's homework after you've submitted it and then usually there will be some time left over an area can turn the recorder off and you can ask some office hours time questions. I'm going to pose course announcements on Model and if you were registered for the course this morning you should have gotten an email about that if you did not send me a note after class and so we can try to figure out your muddle registration. it's not ready yet but I'm going to work with airfare to post it but we will have this to in a forum on model and so you can post questions there before next week's class and I'll go over them at the beginning of class or I'll just answer on the model. So I wanted to make a note about the course work load because this is not like other science and Perspectives classes like it's not much work to the extent that I've gotten negative. I mean I just want to say expectations I have got a negative feedback on the class because people thought it was too much work and so the thing is, it's actually a lot of work for me too because I have degraded the response essays. So it's actually easier if there's fewer students. So if if you're worried about the course load, then there's other classes you can take that do not take as much time, but according to it, would increase. The number of credit points at I is not the credit system is the maximum for a Science and Perspectives course, but the number of hours for most people If you satisfied the course prerequisites such as having some Phantom background and a little bit of blip background. the amount of work in this course is less than ninety hours. And so it's twelve lectures, eleven programming assignments. There required readings to response essays, and then the final assignment. So that's about sixty hours of time just actually doing these things. And so that includes three more hours. So that includes the tea sessions and then study time if you are new to pythem especially, but if you're new to help then it will take longer. So I just wanted I Want to say expectations about that beforehand? Also, if you were interested in this topic of applied Up for Social science then I would highly recommend you also sign up for the two additional credits for the course project so we'll talk about that more after class next week. So if your interested in it, just stay after it you. This is simply recommended for people who might want to do graduate research after because the previous course projects have turned into conference and journal publications. two of the projects were part of into Swiss startups as well. So if your interested in legal tracker or other entrepreneurial projects based on applied help then the course project could be interesting for you so then asked one where doing for the submission of the project. there's there's a link from the syllabus on course projects that has the rough deadlines. Basically you need you haveyouhave to pick a topic within the next month and then have an outline within the next month and then the full draft of the paper is then day until remember September first so you can work on it over are so a related system of what we've talked about already. Thank you to everybody who filled out the course survey. if you registered since I said this out, send me a note, email me a no because it send you a link to the course survey. Oab'll just send out another link so you can fill it out as well if be curious who else has joined. It's about half master students and few old students and then the rest bachelor students and mostly computer science some data science. He's actually zero people from law which is somebody asked do we need substantive law background so if we did not, we would lose all these students. So we do not require that so that two you guys are So I Already went through this a bit in the syllabus, but the required readings are indicated in the syllabus schedule. In addition, there's the bibliography of references that has additional readings if you want to complement the slides in the link related to the response essays. there's the list of applications papers for response essays, which will talk about more next to be. So I wanted to just give a quick outline of some of the books that were mentioned in the references list. Again, none of these are required readings, but for those who want a deeper understanding, I do coming these books. So Natural Language Processing with Perception is the book that accompanies the Natural Language Tooloquate, which is just this classic blip trouble with kind of more standard classical like old school machine, old school natural language wing tools. If you want to learn machine learning, this is my favorite book for earmachine learning with Physicist learn and wood courses and Monster Flow. It's more generic and's not about Inop specifically, but there are a lot of top applications and for those of you in my course in the Fall you would have already seen this and this is available on oil through the Earth Library. you should be able to get it. This is a free book if you're interested in more of the theory and guess for natural language processing more product than mathematical formalization. If you want to do graduate work research in Blip, then I really recommend this book the Your Goldberg book. I think this is available for download as well on the Earth Libraries. If you can not find it, let me know I can get you a Pdf even though it came out in to this and seventeen. It's actually a little bit out of date unfortunately, so it basically has everything up until Transformers, which as we'll talk about have kind of remade inilp. but a lot of the issues and approaches here are still quite good. Another kind of classic textbook is necessary in Martin. Its kind of more than the does not really focus on neural nets inalp and is kind of more than the older generation of help research, but it's very good for some of the more linguistics oriented in semantics oriented part of the course, so this came up a bit already. Python is a course prerequisite see here for the example notebooks and you know I'm sure many of you as I said, Python is a country register, so you should already have it set up on fairy affairs can provide some help in setting things up, but we would trust everybody. Everybody should be able to get their own another environment running. As a prerequisite to this course. these are the main piping packages that we will be using. As I mentioned in all to is this broad collection of older Inalp tools. Finish is great for topic models and award embedding. Spicy is another kind of generic tool. It's great for named in any recognition, parsing in reference resolution, things like this as well as a library of pre trade world factors. and then as it mentioned this new inilp this new neural net architecture called Transformers in particular large pre train transformer models that are trained on large corporate these have really remade how help is done and hugging base transformers as the hugging base system is the standard for that. To provide an overview on the course, here are your objectives: seventy percent if you want to learn how to use help tools and fifty their parents if you want to learn how to apply opinion tools for law and social science so this is great. We're going to do both in this course which are the followings best Matches your goals for learning in top: sixty percent want to learn it for Engineering in Software development Thirty seven percent for social science research and fifty three percent for computer science research. This is good. We're going to be doing all three of these goals are going to be covered in this course so avoid asks if we need to into processor to no no and maybe you're asking if you know like at you for the assignments. The short answer is no and you do not need any special computer equipment the yeah so we you should be able to the examples on the books and assignments. We use kind of small corporate things so you do not need you do not need any specialized competition for that. If you have problems you can use a Google collaps right and so Afro will cover that in the tea. sure you can just those Google could for everything. So why this course right now we live in this really amazing time I Think for language processing where with our lifetimes there's been these new social structures that have remade the data landscape. the Internet Social Media digit join efforts by governments and Google Books for example just as amazing amazing initiatives for digitizing tons of text at the same time as having these huge crops are. We also have had this amazing increase in computational resources as from cheap disease to efficient databases, solutions for quarrying all of those corporate and then having cups give rise to go Pus and then also tips for training these gigantic volunteers and in particular for natural language analysis. We have these really interesting tools in machine learning, a blip and casual inference for the legal and the social science applications of these tools. And for me I Think it's fair to say that at least from a research perspective a lot of as these trends are especially amplified in the law and Legal language. Political Science and Political Language Here many doors that are being opened in these old fields by these new technologies and so we care about legal and political institutions such as what judges write in their opinions, what politicians say speeches, what's written in patents or in newspaper articles about the law or in legislation and regulations. Those are all millions of lines or millions of documents of unstructured texts. and there's no way that humans could read them even if they wanted to add. So this is why bring in these tools for computers to read them is so exciting. So manual asks, could you share the response to the questionable students background acknowledging presence is up. I do not have that in the slides but if all talk about that next week. I don' think there's anything not notable from that. or do have a specific question manual right? All talk about that a bit next to be but but you do not need to worry about that. So here's an outline of the course and actually I would say let's will all just go through this and they will take them break. So I know this is like an eyefool and I made this last year, but we're actually going to follow basically the same format and justice. but you can visualize everything that we're going to learn in this chorus from this gap. And you what? what we're starting with as raw data today. and next week we go from raw data to segmented documents that are pre processed in different ways. And once you have these documents, you can only use these in some social science analysis just to say oh well, how long are the documents you know? How many bonus do they use? What's the word link to the sentence Link This public a measure of reliability or sophistication in language. The second would be dictionary accounts. and I Think if you're a example researcher, a computer scientist, the fact that you should just go and count different words and count the number of the times the word good shows up as a measure of sentiment that just seems so primitive. it's like the stoneage it. But it I think that we should consider those models cape seriously and I'll give you a good reason at the end of today why dictionary methods are are not to be ignored. And so next week we'll get into tocanization. So the different ways that documents are split up in to sendances and words and looking at part of speech things like this. Once we have documents as these end up being teprimitives for all the models that we will analyse including in gram. So that's converting tokens to phrases. Topic models that's converting documents to distributions over topics so you can imagine in illegal groups there's a crime in contracts on tutors and patterns and things. Each stations are left wing politician I Think my internet might be unstable but I'll give it a second. Can you go hear me now? Can you put it in the cash? I back market thank you So think asks do do we take a look at how he's methods roughly work or do we may learn to use them or what were weregoing to do both rateboth so we will in the notebooks in homework. In the tax sessions we're going to be learning how to do these things in Python we will implement them, but in the lectures were going to be focusing on whether's the purpose, how whatever, going to help us understand the judges or the lawyers things and's so after machine learning will get into neural nets and a particular if we'll learn about word embendings which is a way to represent language in a low dimensional space we'll get into passing, getting at syntax, the relationship between subjects and objects, agents and patients. This is getting into linguistic sides things. We will then get into Transformers and that part of the course which will lead to are ample language modeling knowledge graph's entitlement. So this is getting into asking a computer does a empty public, does sentence A empty B or unknown We do information and extra going to extract relations from a corpus and learn about what the corpus is about and towards the end will be getting into these of more global semantic aspects of summarization. Question answer, automatic claim checking, casual inference from documents, identifying casual relations in documents and a lot of this gets way past that social scientists are using right now. But I think these technologies are improving rapidly in the case of the legal domain at least the there going to be clear applications that really have not been done yet but or will be running right to the frontier in this course to take a look at this if you like will be following us as long as we go in the course. Okay so I know we've already gotten abunchab logistical questions but I wanted to take break now and give everybody a chance to ask a question and then we'll take a quick coffee break. So are there any questions at the moment that have not been covered? You can put them in a chat or the race hand function so manual ask how relevant is reinforcement learning to snap. That's a very interesting question You next's not universal, but it is certainly relevant. There are many very interesting reinforcement learning applications help for example the recent paper that cannot used as reinforcement learning to to improve the quality of summaries and I have a paper with a old student which actually came out of this course using reinforcement learning for attractive summarization. So if you understand reinforcement learning, there's a lot of interesting applications and I can provide some more resources on that. So so fun. Also, memory can note of that area. can you make it note of that? They set up environment script that must have been something from last year so we'll fix that thank you find so report access. Is it possible to do the course this semester near the party next year? Sure it the mean things change right but I'm planning to teach the course again next year so should be fine. Thank you for asking that. So think asks on right, think is asking. theatre's not a simple yes or no. answered to that question Sometimes yes he mean it depends on it mean we're not. We're not going to be like trying to implement them and see Plus Pals or something so but you knew will be talking about you know will have some review of Nstarcastic Radiant Descent. You know how volunteers learn. You know it is not as it said it and this is not a substitute for machine learning in our Pop course so we will do some. but if you need that then you had take a different course or take both of right? So we're going to take a breaknow and will return in nine minutes at there fifteen. I'm also going to turn the recorder off now So if you want to ask a question the recorder please do of really's We really started the content of the course for the remainder of today so nfuzum in on this on the course outline We get to the Corpera. so text data is a sequence of characters called documents and the set of documents is the corpus which we will call them in this course. And what makes text data challenging but also interesting is that it's structured. It's just this stream of characters and the information that we want is induced in the symbolic information in the those characters and it's mixed up with a lot of information that we do not need for any and task and so a lot of what we're going to do in this course is its information extraction or if you prefer information disposal. we're trying to get rid of the information we do not need and will be the main focus of what will do next week and a week after. And to recap something that was mentioned earlier, this course is appealing in place and we do not care about that much of the documents by themselves. What we care about is related into better data and that could even just be like the time Theatre document is published. So we might say well, how syntimate towards a impose social group, How a sintimate towards immigrants changed over time And we can. we can. make a systematic measure toward immigrants and end show that How that's evolved over the last ten years less one hundred years. And the important point there is that we have met a data on the time and so just to say that it be more specifically new might start off with some positive negative syntimate capital by and judicial opinions. And that by itself is not that interesting for a lawyer or for such a scientist. But what if we had the dynamite in opinion is by Judge J at time It and so we will often use this type of notation with these subscripts for the indexing corresponding to the meta data that the document corresponds to. And we can say you know how to sintimate very over time. Or let's say we have the information on the political party of the judge are they do. They have more negative sentiment towards defendants from groups Go. So let's say that go is a dummy variable enjoying one for cases where the defendant is an immigrant and so then we can is information in our data set to say. Well, the right wing judges. They have more negative sentiment towards immigrants for example and you can see that one you relate the text features to meet data. There's many interesting questions that one can answer and a precursor question to you. This type of analysis is what's the document. So now often have this big data set. If we might have a bunch of books on what do we do with them, we need to zoom input to some degree to find out which document, which observation is useful for our meditative variation. And so you know if if we do not want to just arbitrarily make the documents really small because they do not, they do not help us answer a research question such as you know our republican judge is more negative towards of immigrants. If we made a data seat at the sentence level for that, the sentence would both abstinence. data would be super high dimensional, but we would not have any additional variation comparing the right wing judges. did the left win judges. And so this is going to be one of our first Islands activities. what should we use as a document in these contexts? So I Want to take about five minutes six minute to answer these questions? We're going to set up them, going to set up breakout groups to put you guys into pairs and this is just a way too please pull up the slides on your own computer because you're going to lose it when we stop sharing and I'm going to put everybody into pairs and this is just a way for us to start to know each other. So you going to be. You're in a breakout room of two or three people. introduce yourself and said that your major and what you are interested in the chorus and answer these three questions together. What counters the document and were is going to open the breakout rooms for six minutes and will be back at Fifteen Twenty seven so only so handle in corporate. So we have some examples in the notebook about some breed processing to especially if our working on a project not just in this course but you knlater on in your career in life. It's good to think about from given questions or given to ask the data sets out there that have been compiled and so far for court cases like in the U is for example court listeners good but in social media there is really excellent data from Twitter and Facebook for example. for Red it is also for Wikipedia All these data that's are really useful for such a science analysis. This is not in part of the course necessarily, but you know it will be useful for you later on to learn about queueing ships running web scalpers doing processing to remove home markup and there is also the hugging face hub and hugging face system. They have a lot of amazing data sets so it's good to just kind of produce through that ski that a bit can fu access should learn or should have learnt so it would say it mean that should learn because it do not be tested in this course but it will help you to know it for you for other things. So I recommend learning it but you can do it after that. We do it in the summer so all everything that we talk about this course is kind of language agnostic I'm a native English speaker so everything is going to be focused on English. but for your course projects and things you're welcome to do things in other languages. After one of area's special teas is multilingual implants so she can help you get started with it. And there's also just really great machine translation So Elevthos asks why should we use Cyllinium is not cycle evercovtalate that it depends on what what you trying to do then I think that doing a webscraper with that you are a Lib is like a nice start. but if you keep going on up for social science for data science then we will come to a point where you will need a crime driver or something on those lines. So how can we use the quantity of text as data So you know most of this course is going to be about the semantic or conceptual or symbolic content of language. but there is also interesting things to be learned just from the service features just the statistics in the characters in documents and one thing that me and old students in my group had looked at is how the style of writing changes over the life cycle for judges. and one of odd or curious thing we found is that older judges that use shorter words but longer sentences and so whether this is like better or worse writing thing is kind of subjective but it shows that there is. It seems like there is this of biological component of writing style for judges. Another relevant debate in the case of legal quantity of text, the law is on legal complexity where this depends on what country you're from but like in the U's and Finance for example they're always complaining about the law being too complex on but using help we can actually try turn this into an empirical question and ask how complex is the law and certain bedroom which is one of the applications. Papers linked to the syllabus is about measuring the complexity across different areas of the legal code and they produce the number of words in a code title which is a section of the code and they also produce a word invite measure which is basically it's the tiny of the probability distribution across words and what they show is that Public health and the Tax code for example is very complex. It has a a lot of words in it but if you look at the codes if you look at the code titles that have high word intern so a more diverse vocabulary it's quite different and so you. The Commerce in Trade or Public Health and Welfare scores highly on both measures of complexity and so from a tax lawyer perspective. This is kind of interesting because it shows that even though the Internal Revenue Code is complex in terms the number of words, it's very repetitive so it might not be as complex as if sounds. So the next set of methods that will talk about today our dictionary based method which is counting words or patterns and so in the notebook we show some examples of using regular expressions which is this really powerful string matching system and this is going to be depending on what your research, question or engineering task is how you would want to to do this. So one theme or question in the law or judicial behaviour is do judges? Are they making decisions based on justice or morality or are they doing it based on cost and benefit in analysis or efficiency? And so you can imagine using a dictionary method to go through to what different judges say and say, do they say justice demands or do they say efficiency demands And so a dictionary based method could help you get it that. There are also general dictionaries such as Wardnet and Luck which will talk to you about in a second. One example from Economics and this is also one of the applications readings is Baker Bloom and Divas where they wanted to know what is the amount of uncertainty in the macro economy and this is like if you're a Mcroeconomister, you're going to finance a big beam of when there's more uncertainty in the markets, they're more volatile. That could be costly. And so they built this data set of newspapers and they use this kind of simple dictionary method. Does the article that the word uncertain does it have a word related to the economy and then does it have some words related to the government and then they ploddedin that and this is actually just the main result in the paper is just showing that this policy of uncertainty index based on the newspapers, it spikes during these kind of stock market shocks. So like Black Mundane or the Wars elections, nine even. This is the financial crisis to two thousands and nine the Euro Death Sealing crisis. And so you can see that these kind of intuitive moments of market uncertainty are captured by this newspaper based measure. There are are some problems with that to be sure if you're curious about using this in like financial applications is recommend to keep that all paper. another set more fundamental dictionary methods are are available in World That which is this nice python package with a database of it's a big dictionary with all the words in English and how it related to each other and so you can see for example, all all the different senses of the word bass or bass are located here. so it's like music or voices or the fish. There's many meanings of the word base or bass and the the word that it captures these synonym sets of groups of beer anonymous and has information on the anthiponyms and also parts and holes and then also all towns are organized in this categorical hierarchy and so you could have employers which are the higher word in a category and then symptoms are the members of a category and so you. There's many ways this could be used, so if you want to do definition reduction, you could replace words by their hypernem for example. And if you keep on going up Twilight of Categories word that has twenty six lexacer, gray graphic suppresses and so like action, animal partifact person relations shape things like this. I think it's pretty cool that these linguists have gone together and really tried to organize all of language in this and it's all automated now and a Python and so they did the same thing for verbs as well. So this is for bonus and this is for verbs and so you could take a corpus in, say, well, which books are mostly about perception versus possession for example. and you can imagine there's a kind of digital humanities or cultural analytical types applications for these methods. Some other general dictionaries which are relevant for us include lists of function words so thinking of words like four rather than also these get at the style features of text so most of them have you going to drop them because they're not going be very useful for topics, but they can be used to get it nonpolitical dimensions. So if you want to identify authors of texts without that being correlated with the topics, then using software is a good way to do like. or luck is it's kind of the standard licenses for linguistics and Social Psychology and the like team. They have seventy lists of word's of category relevant words including the commotion, cognition, family, positive, negative. We will see some applications using Luck later on in the course and in more recently there is these big lists of words that are initiated by people on crowd scouring platforms. So Mohammed and turns have joy and sadness, anger, fear, trust, disgust, anticipation, surprise and can warmer at all. They could fourteen thousand words along violence, arousal to dominance dimension. So these last two on kind of emotional content. Those are part of this broader set of tools in blip on sentiment analysis and this. You'll see this in many papers, but also in an industry like in advertising digital marketing, they really care about determine right for obvious reasons and we want to know for a given like a review of a movie. Is it positive, negative or neutral And the standard approach could be licensed. base research for the word good, but it's easy to break that right. Like what if somebody said though the movie was not good or the movie was not very good and so just like amends other words totally transforms the meaning. and so it means that just counting words often is not going to work very well. And the more moderate approach is to use pre trained transformer based syntimate models in area I Think is going to add an example of this into this week's notebook to show how you can download a pre trained sentiment analysis from the Hugging Faced Model hub and apply that to documents and you should be careful about these off the shelf scores through these are trained models because they might be trained on a corpus that's quite different from yours. So if you take a corpus that's like initiated tweets and then you apply it key contracts, that problem will work right and so you have to be careful about that. Check that it works in your new domain and there is also some methods which will get too in the world embeddings for making some domain specific licences so you can start off with some positive negative words and then you might find that in read it. the positive words are different than the ones in Instagram for example. So I wanted to just point out a specific issue with syntimate awareness that are based on initiated data. So if you take a syntimate analysis like from hugging bass or something and this is from from some slides by crime car I saw on his Twitter feed. so I do not know exactly which model this is but he made this where if you just take let's go get initial food versus let's go get medical food this has a very positive sentiment and this has a low sentiment. This is bit so just changing the word from relation to Mexican but just that soon as by itself the sentiments exactly the same right? but they changed it from relation to medical and the sentiment went from positive to almost negative and then this is an even kind of more striking example. If you say my name is email, you get this positive sentiment. but if you say my name is unique while you get negative sentiment and so this is like really kind of striking and important and concerning issue in all Pi systems and you want to ask the mean, Is this sentimental model racist? What's what's going on here? How did that happen? And this shows how you have to be careful about using symptomatic scores that are trained on initiated data sets because they're also going to be learning this correlated information in the data set. so that's not unsintimate, but then it will learn it and apply it when you use it in a new domain. And so this is just part of this broader problem or challenge or issue in an Apple for social science. But also this is going to be relevant to many things for products as well that we care about. some true sentiment in language. It but what we get is a systematic scorer. and the model that we trained to learn a sentimate score from language has all these confounding factors. So you nonwhite examples for medical food versus Italian food. You can kind of see how that might have happened right where initial restaurants. maybe they tend to be more of scale or like thecritics like them more. and so because the language model or the competent classifies trained on these biased data sets. that food this, let's say, food critics like Italian food, then that gets baked into the intimate model even though it has nothing to do with even though it's using these words that are syntimate neutral. So this is not easy to fix me. You know, because there's not going to be, you can not just make data set that's neutral, like every data set going to have some biases in it. And so I Think that trying to fix it is this really important and interesting area of upon research that's happening and and this is not unique to determine either. This is a universal problem that I want you guys to be thinking about Throughout this whole chorus is that we are trying to measure this true quantity in language, trying to extract it to learn about science or to learn about to solve a legal task to make a prediction and whizzl we care about. But we get this. We can only make a measurement of its indirect measurement and it's going to be confounded by other language features and sometimes non language features like where it comes from or the large age or style and supervised models are just by construction how they're built. they learn features that are correlated with the label being initiated and unsupervised models you in a topic model or world embodying there also going to learn those correlations and so you like. A classic example is like the similarity between the word banana and the word yellow is actually pretty low. but the similarity between the word banana and the word green is actually really high. And it's not because bananas are green, but it's because if a banana is yellow you we just never mention it right and so you have to be very. This is just some examples of these these problems or limitations with these language models you have to be careful about when you're interpreting their their outputs. But and this is what I mentioned at the beginning dictionary methods. They do have these other limitations But They very significantly mitigate this problem. And because the researcher is familiar with the context, they know what would the words mean and so they can always regularize out any serious surroundings. And so if I'm like trying to make a sentiment analysis and my model tells me that captain means high positive, I can easily fix that with a dictionary method, right? And so this is like a defense for dictionary methods. potentially. And I think it's why economists in particular and just other empirical social scientists. they might still use dictionary methods because of this reason. And I mean they have these other limitations which you you can not. Those are difficult to deal with. but we have to be careful with these blackbox models. Okay, so let's wrap up so the first assignment is already put on the gatehub as well as the coding example that Afro will go over on. Friday. So this came up. We explained it a little bit earlier actually, so I'm just going I'm going to stop for a second and answer elithereosthis question sorry is missed that earlier Those elyptherias asked are those general dictionaries useful any more since we can not easily measure the similarly between words. and also such dictionaries require a lot of human labor to be kept up to date. So I think that's an important point. So I mean the general dictionaries. They are built up over time by these research groups. but I think they have limitations and I mean I think they should be updated over time but meant I think of now all the methods that we'll talk about in this class in terms of dictionaries and classified things. They're all going to have kind of prose and cons and we's want to identify those problems and think of ways to address them. So the way that time the timeline will be in this course about the coding practice and the homework is that for last week to so like week one, the notebook, the coding notebook for week it is got to be reviewed in the tea session on fairy week it and the home work for a week it is going to be done on Thursday week to plus one. So for example, week one notebook is going to be reviewed this fairly homework one is going to be done next Thursday arch third and then reviewed in the week to to session. the notebook two will be reviewed in the week to that session and so on. All of this is in the syllabus in terms of the days, so you now something is confusing. We'll cut you guy some slack in the first few weeks of course, but I think it'll be self explanatory as we move forward. That's the first week we have five minutes left for any questions. |
| 1 | ASRnlp_law_lecture_week_2_v_2_c_transcription_2.txt | 016e8d29-288 | ASR | I Just want to recap quickly what is already announced to the class because we now have this beginning of room for everybody to joining persons. We will prioritise the impression teaching, but there are a number of students for various reasons who will be joining remotely. Of course, if health or other issues are relevant for you, please do not feel an obligation to coming frozen if you need to from home. That's totally fine. So to start, there is a number of questions on the module queue on a page that I'd like to answer. Imjus can I take these in the order of top of posting. So one of the questions is is it okay to copy paste from the note looks for the homework in general yes so you know you're doing it. It's good to understand. I Do not recommend copying and pasting and just blindly just copying the notebooks. It's better if you understand what you're doing, but so formal there's no formal constraint on doing that. The second question is homework emissions, so there's a number of them about what to do with the field. out, no books. The edge flow page has now been set up and you should be able to reach it from the course model if there's any questions with that. let me your after to now but there should already be submission pages for all of the reducing homewares and I think in ask the same thing. The edge flow had not been installed yet, but it should be up there now. so another question is about or these this? thank you for those questions but I think those are okay now. Okay, so could you raise your hand if you in the Computational Statistics recession only I Think with three students, we can not change it. Unfortunately the the attaches are recorded and so please just watch the recordings. And if you join at eleven at the end of the computationastatistics season, you could still be able to ask questions andsomebody asked if computer science bachelor students can count the credits as gas or minor courses and actually do not know about that So it would say ask or someone in registration for your department at experience be two of that because it was one of the guys who actually ask you only not looking that based people on Du Tesigns datchelers this force is ramble even as an agantion step as well as a go as a guest from basically depend on how you begin when you chose your when you chose to basically woman in store, future of future enron, an ex sex vacations of them that be used as a agitation scope or or compulercize nature if you show that there's a guest hop analysis over as a guest. Score Great Thank you. Relax Are there some other questions or clarifications about that? I Probably is not answer but check with your registration officials at your department and let me know if something if you need more information about that right and somebody just asked to clarify if the course projects and the district are counted differently and mean they are counted actively in terms of course credits. but I'm not sure if you can count them under different programs. from my perspective if you're department will let you then it will let you. I do not have any problem with that or right? I Thank you Dived for posting that you can coordinate this with the drink study administration and that's on the model if you want to check out. Answer: Okay I think those are all the you and a page questions. Are there any others that anybody want to bring up live to Asked: Can I do the course project in the summer or or is it does during the semester You're supposed to do it during the semester but it's not due till the end of the summer. But if you have any question about locals just set me a note role you can figure that about. Okay so this is related to the question about the homework what's due So on Thursday by midnight area will go over the homework homework one on Friday morning on the to session. it's best to submit the just the ipi and believe the Jupiter notebook few directly onto edgeof for their completion grades so we do not go through every homework and check them. But so you'll get full credit for completing the substantial completion so this is going to be targeted to combination you will just spotcheck radio, low spot check up and also a programmatic component if you have any questions about whether you are whether a specific programming assignment got for fighters or not a just a check with autism we do not like you know it it becomes available for late later you its past fails so it's just zero or one whether you is if you in our judgment and I think the homewerks are not that sophisticated so I think you can give in every section and more can try a good try even if like if there's an error that you just can not figure out just limited anyway and if there's evidence that you tried it then you get clean if there are there any questions or feed back about the first to session. does that a formate work for everyone at you mean should every person work on their own notebook? Yes yes So I mean for the assignments you can work as groups to solve it but prepare your own notebook and limit that on your does I answer your question. So for this data session on fairy app will go over as the homeortist to be ready and then also the week to notebook on tocanizatioand you can use the Qna page from on on model for questions in the tax session as well. So I mentioned this a little bit last week about the final assignment this is going to be first. It's just a kind of an exam you might say where it's just covering all the material of the course and it will be distributed a week or two after the class ends. It'll provide the exact dates with plenty of time in advance and it will just be based on the ladies and the required readings. It will be designed to be pretty quickly two hours to complete, but you'll have a few days there. Four days may be a week to complete it so you can schedule this around your other obligations at that time and we will do some review and will show you some sample questions during the class or so. Last week we started with collecting and cleaning corporate and doing a quantitative measure from corporate and also a dictionary methods. It wanted to introduce an example of a dictionary method paper from my own research to see how these dictionaries are designed to solve social science problems. So this is joint work with David So who is a student with me here as well as Art at Being At Work and warm rural at all. And this paper is motivated by the recent discourse about race in ethnicity, issues of discrimination and prejudice in equity as a policy problem and does that policy problems motivate that there of above more research about this issue And there is this anecdotal or stereotypical understanding that Economics as a social science compared to political science or Sociology has done less work on these issues and previously this was just in people's heads. But in this paper, we actually look at this eventuality by applying nil up methods to the text of scientific articles. So we built a corpus of five hundred thousand academic publications from these three disciplines, and we used a dictionary approach to identify against the relevant articles. and so please read the paper if you want to see all the details. But I Wanted to note a few highlights just as an example of the dictionary methods approaches that we started discussing last week. So first we considered all publications in the journals that J Store characterizes as comprising the disciplines of Economics, Sociology in Political Science. We also topped up this corrupt with additional articles from the Scoops data set and also the Web of Science data set. And so the goal was to get close to universal coverage of published articles from and Beaten and Sixty up until to Thunyad Twenty In that end up with a half a million in publications. So this is exemplary of this first step in social science research, which normally in an opinion class. You do not worry about this right you just take the courses given, but in order to answer questions such as how much have the different social science disciplines been spending on these issues of discrimination and so much time have they been spending on it, you might not be able to answer that immediately because the data and does not exist. This shows that building data and cleaning it's this important part of social science applications of us even before you get to the modeling part. So read this live can you like? But I just wanted to point out that our dictionary method is actually quite subtle in the sense that we we be matched a number of patterns. so we have a number of synonyms which we which we refine by looking at examples of different race inathnicity groups as well as related topics. So we did not just look for the word black because that light refers to the work to the colour right and so in addition to requiring the mention of a research group is also required a mention of of a topic of discrimination, inequality, prejudiced bias, historical issues like slavery or him cargo so that in order to yoincrease precision of the resulting dictionary based classified identify articles that are had do not just mention these relevant words but also have a related a substantive topic and shown we did invitations of articles. This turned out to help a lot because often times you'll have an article that's about you now minimum wages and you'll have like in a system abstract describing what the paper does and then it's only the last sentence that says and then we looked for heterogene eighty by white, black in hospital that's actually quite pharmaceutical and so we wanted to focus on articles that were very so typically about the issue of race and acidity and these different subtle revolutions and refinements of our dictionary base method allowed us to do that. These are the main results from the paper showing that even though Economics in the top left, economics is in blue, political scientists and green and Sociology is in red. Even though Economics has the most papers, it has the least papers about race in ethnicity issues and Sociology has the least papers but the most papers about race in activity issues and so this anecdotal since that sociology especially but also political science are paying attention to these prejudice discrimination in equality issues that been anecdotal since that kind of conventional wisdom turns out to be right at yes that's just because of coverage. actually they're not in the database so if you so probably to get higher quality data is that we should have finished around twenty and fourteen or so. But it's because it takes some time for all the journals all to their articles to get into the day to base the easing of. But but if you included everything for real it's going on Still is even speaker destroy the way the number of range related obligations who then do through thawing wanting worse or it's journal quality weighted the question. We should put populist that into the title, but it basically multiplies the number of articles by the impact factor of the journal. So basically journals, they get sides often, so they're kind of more influenced on the discipline. they count more on the graph and so this is really just. this is mainly a robustness check more than anything but we. He wanted to check that it was not that Economics might have fewer publications, but they're much more impractical. We can rule that out with the bottom left panel only that's dictionary methods. Now we're going to get tobasically the whole. Most of this lecture today is about taking documents and transforming them into restrictions, and you'll see that there's many ways to do that. Almost everything that we do in this chorus is about document representation learning, a numerical representation of plain text documents, and the process of taxation, which is nsegmenting the document like breaking it up into pieces like paragraphs or sentences or words, or where pieces or letters. That's a pivotal step in that process. So just to summarize what we're doing, we start it off with some plain text documents and what we're doing today is converting notes into tokens, and then one of the workhorse document representations in blip historically and also still today is immigrants which is basically it phrases. So we a document goes from plain text to account a frequency distribution over two or three word praises and is be going to try to keep this notation consistent throughout the course. of course some papers do it differently so it do not be hundred print. but will you use capital detail referred to the documents in capital way to refer to tensions like lists of words and capital x will be some market representation or frequency representation constructed from the tokens. So to summarize and you when you are reading newspapers and somebody asks how did they recognize the documents and why these are the country factors that would be relevant for us so they need to be informative or predictive for some task be text classification, doing a topic model, training, word embeddings, or building a language model they should be. This is one of the goals that is often not satisfied, but ideally the tokens are interpreted so the holding everything equal would be useful to be able to count the number of times each word it's mentioned rather than each letter. If you just had a representation of a document that was counting a number of times each letter like G X C L was was mentioned, you would not be able to look at that data and understanding anything about the document was about. But if you could look at the talk words or phrases in a document then that will tell you we' much more interpretable and then finally treatable. So this is becoming less and less of a problem as computational resources increase. But it's still the case that you know you have a million sentences in pose and you need to compute the paradise similarity between the sentences. Let's say I want to have a million sentences I want to know how similar are they to each other? That's a million times a million comparisons, right? So how those sentences are and presented will be really important compucationally. So there are two broad approaches to this one. if you might call in the standard or classical approach that is from pre neural nets or at least pre recurrent natural nets blip where documents were represented by counts and by the longer sequence longer and sequence information in the document was removed. And then there's a more recent approach which is to maintain the sequential representation documents so you can imagine in Tea. In the first approach, you take a playtext document and you get some counts over different words. In the second approach, you get a list just the origin, all words you take in the whole document as data rather than accounts or clear over vocabulary. This is a no kind of Opstrak, but we will see examples. Here's a askematic for the first kind of of recognizing pipeline, so this would be a way to do this in Python in ineltik. The first line reads in raw hotel text from your website. You can clean out the hotel using the number of approaches such as a beautiful soup. take some snip it of the data tokens would be Taking the raw text and splitting it on the space is what it normally means. You get a list of words rather than strings and then for example you could say you do lower that will give you the lower case that will start doing so. preprocessing, putting everything in the lower case for example. and then the vocabulary at the end is just the set of its heat of unique words. In the purpose in this process of tocanization, you can think of it as building a degree. arguably most of the time it's kind, the space of possible representations that a document can be mapped into. The second kind of tocanization as used in transformer models. This new generation of help. Most of the time they used what you would called subway tocanization and the form A practical standpoint for this type of recognizing this standard type, you probably want to use spicy or Ginsim to do that. Those are currently standard tools that have like these standard recipes for doing that or Psychic Learned Tide Factorizer Back to for this type of toconization you want to use the Hugging Face toconizer. So I think after you are ready introduced that hugging base system in the tax session. So we will be. For those of you who are doing transformer based help models using context sensitive help, context sensitive embeddings, then the hugging Face Stadium is something that you will want to invest in in Learn Houdworks. So for example for Bit which is that is guess to' kind of one I just be the work horse for short document in all using Transformers rather than a vocabulary that includes all words, they have a vocabulary that includes subwords. and so for example you a word in the best vocabulary. it could be three words or four words. So here you it says playing could be reduced to play an wing. And so that's why it's called forward Tocanization because it would take the two world pieces and treat them as two separate words in the vocabulary. We'll come back to this. So before we're getting to this token representation or either list or or counts over tokens, the first set of choices that you'd want to make in developing good data is preprocessing the text. So there are many steps to this process potentially. And as it mentioned, there's a number of recidities such as a physicist learns stiff facterizer or Chinsim preprocessing or or some of the space functions that will make a lot of these decisions for you. But it's important to think about these because before example, whether you move, capitalisation or punctuation could make a big difference for your downstream outputs depending on your task. So the first usually is taking full documents and splitting them into smaller pieces so you might want to take the one that often uses letting into sentences and as a mention of a you have this task that's doing pairwise comparisons between a million sentences you have had to get the sentences in the document first right and so spicy I think would be a standard sentence solution. so when you input a document into space it will immediately do the splitting and it will adjust for nperiods at the end of more or messes or abbreviations on us a porch things like this and you'll get informative set of sentences till work. A lot better than just splitting on periods or full spots. In terms of splitting paragraphs and documents, there actually is no standard solution to that because that's going to be a quite domain specific how paragraphs are split. If you are using hotel there will usually be in be the pop tag or the bar tag and if you're like in digital documents then you'll have to have a costumed solution. I will tell you that for most digitized documents like over a line that ends with the period or it full stop orquestion market explanation point that's almost always the end of a paragraph so it do not be perfect but you can use that as a short cut to split paragraphs using our data. Just align it into the period and you know part of this breaking up process. Preprocessing is the idea. Threats is something that will repeatedly come back to is that unstructured text date has lots of noises, is a lot of extra information that is not useful and so we need to develop our pre processing and featurization steps to extract important information and exclude the irrelament. So of course theatres many papers about this, but the dining supporting paper, for example. They undertake a number of systematic investigations of different pre processing choices and especially for unsupervised learning. So remaining a topic model or clustering things like this, the preprocessing makes a big difference. Fortunately, this paper also shows that for supervise learning, classic machine learning, classified fiction, and regression, the pre processing choices do not matter as much as long as you have a sufficiently rich representation of documents. Text class birds will work well, so are choice is whether to remove capitalization and so usually the capitalized and non capitalized version. This is everything that going talk about is mostly about English, but you can imagine that there is going to be similar or parallel issues in garden or other languages. The capitalization issue is think even more nsomekind. It's like more interesting in garden, but in English of course knows are only capitalized at the beginning of a sentence. Risk in titles and things like that and so often times you increase the size of the feature space by treating capital letters differently and it's usually better to remove the there are so many exceptions to this rule. So in the legal context you can think of the First Amendment having a capital effort. Capital A Three is deteriorating about American law. So the First Amendment refers to Freedom of Speech and religion and things like that. And if you read the phrase of the First Amendment without capitalization, you know that they're talking by about you no specific law or specific contract or something but have the capital to capital as they're talking about the Bill of Rights to the U's Constitution is and so that will be an example. For legal documents, including capitalization could be pretty important. Also you of course if you're doing linguistic initiation of documents like part of speech tagging statistics, passing semetric role labeling, capitalization is really important for that. What causes what you might often have is the source documents are not capitalized correctly. So in the case of contracts for example, you'll often have sections of contracts that are all caps that all capital letters just to lie, highlight them. And and that's a problem because like things are a part of speech tagging and a synthetic parsimple break on on text like that. so you might need to do some custom checking. This is a nice example where punctuation is quite important. So I got these from Chairs Bail's slides. In general the rules about punctuation whether you should keep them or not. It's kind of the same as capitalization where usually they're not very informative if you have immigrant representation of your documents. but if you going to use your data for some linguistic invitations such as sentence splitting or part of speech tagging things like this then you need to keep the punctuation and information in your doctonants similar for numbers. Most of the time you should remove numerical information from the document just because if you are breaking up your document into matrix accounts is just counting that. How many times the number one mention you not going to be very useful or especially if you're counting like the number of times the word the number nine hundred, nine, two thousand and one hundred is mentioned that do not be very informative and no would say replaced numbers with a special character like a hashtag would you be a decent approach for language models like get two gptthree bright numbers are just treated the same as any other sea are and so they'll be part of the subdued tocanizer and will get to language models in weak nine. but this these big language models like God there actually can not solve mat problem. So if you give us some tiny pleas to plus two equals for two plus seven equals in nine things like this and give it a new problem by will actually often times provide the the correct inverse. This is particularly why we're doing this kids right now because this amazing technology of language models is transforming how lawyers and social scientists can use language as data and on for practice. But the this really exciting and intriguing finding that language bottles can solve math problems. It means that they're starting to have this kind of conceptual representation of language under the hood, but it still does not work very well. So it's easy to find math albums that get there does not can not solve. And so this is an active area of research and there are many projects to do for having language models understand numbers. And as a side note to that, having language models that can do things like fact checking for example, scientific claim verification and having them understand numbers is going to be a really important piece of that. You can see how this is practically important. Dropping software is a similar type question as punctuation and capitalization. There's a standard list of words that show up all the time in English, but do not really mean that much by themselves. On the other hand, it's easy again easy to think of counter examples in the word not is often troops a stop word. But if we're reading legal documents and we're trying to say you know what is the judge deciding having the phrase not guilty is inappropriate to include right and then just more generally in the law and in other technical all domains. Oftentimes specific phrases or means are going to have an independent and pivotal meaning beyond the words that are composing them. So like the classic example in America in law is beyond a reasonable doubt. This is like a very specific evidentiary standard in criminal cases. and the words beyond reasonable doubt by themselves. If you counter those, those would not be that informative and with all deliberate speed. That's another procedural phrase that's very important in the U's law. And so you can imagine that even though those those phrases has contained stockworks, right? So if you drop stockwords before doing in gardens, these would not come out in your future representation to one option here you're practically speaking would be to drop software when they appear by themselves so you do not care how many times the word a shows up. But if they show as part of these phrases beyond a reasonable doubt, you would keep them in the future representation. Another way to you refine the corpus to make it more informative is limiting or limiting. and so rather than include every word that shows up on its own, it could help to drop surfaces. So in English this would be to consigned consigned, consigning consigned with all four of those things are talking about the specific word route consigned and so applying a swimmer will remove all all those suffixes following a rule based algorithm porter sometimes I think we pretty well in most cases and there is something called limiting as well which spicy will do for you which rather than just split take off into the sentence, it will actually look it up in a dictionary and give you the world route only. So this was going to be designed for the hybrid zoom chat but lets lets just do this in person instead. So just to start practicing with these issues of of non word or style features in language, consider the following four dimensions of language based on the first letter of or last name. So read either if get disturbed based on your last name, think of a social science analysis or important legal or social dimension for example or judicial opinions for newspaper years for social media for political speeches, think of something that's interesting to you, perhaps related to your work or other things that can be either can be measured by capitalization punctuation would change depending on the use of software or we change depending on the use of simulator limbertizing and so we're actually at the the point of the break so is Let's think about this now and you can take a few minutes but also take a break and we're going to resume and a fifteen minutes a ten actor the hour and will give some examples of will have people to volunteer some examples for these categories so will see you in fifteen minutes stmepo degree to at it on the alright we're going to resume now so can I get a couple of designs from those where the last name we direct for something that can be measured with capitalization. Yes annual air last name starts with off someone who thought about capitalization being a relevant legal or social science point to deal animal which for you and do you want to start us So I'm asking for some volunteers for these four groups. So what is your last thing to start with? So great. So you are in the first category then did you think about this you come up to an idea for something social science on social media, law, political reached where capitalization would be this important dimension. Let's okay we havepouned the back here yes at and that's interesting right? So you can imagine you people who have the more years of schooling. maybe they have a longer's longer a more complex language it has or fewer periods but more cocombas for example interesting And the abundance of pertulatory events using doctors bacors be cause more along plots heat this shorter grade umation interesting right? So you can imagine like in liked transcription of patient's speeches. if there's more like dots that they're breaking up their language more because thneither'r taking more extra both is like that. That's interesting yes and at often the J owl are done although it are a simple passport and analysis version by the biological and that so just an etaspoto to the court other is so and what was the role of the porter semester in your approach analysis there can be built that would plate time analysing, forge in on and so porter sometimes I guess it depending on what you're in your context. the swimming was worth while because the world endings did not have as much information is needed. Is that fair to say He I see thank you any other exact yes constitution age weights that like the First Amendment ex ample or like you's constitution article there and do not be among number is for example woman numerals are often capitalized and so like luck it is going to be different than just looking for lower case and so interesting at at to can topecifoe report on a rise to iyaitionsa to get and of new truly vote changers sometimes right But I think that ye turkish and maybe Japanese like this too much to but through examples sometimes like a whole synthesis in one word right like you could have like addition or her like tin surface saves is one thing. So why you won you very careful what's standing in Turkish so I think that's a very good very good. totally once here and or son the second cussificationak shout interesting Totally right yet so like tweet the yealternating fur lower or K R I d' I did not put it here but I guess you could include a mode jesus punctuation as well about thinkffor a syntomate classieria all cap it's going to make a big difference. you have a well at of routine on tree was root seeing airport announced special rises of interesting totally right. So like all naninnity recognition for example. so if you want to who you let them second but like if you want to in English especially when things are capitalized they have some kind of special properties to get them of personal about. for example procedure for it is important another point there yet he and interesting right? So that's a interesting thought right? So you could imagine like there could be some systematic methodical way to see know for what words are, timing or limiting capital that's a good project idea potentially to right. Let's move on. Thanks everyone for those on the zoom if you want to add examples please type of Imachap. Also now we're going to move forward with recognizing the most basic unit of representation in a text is a token. So thats what were going to refer to as when we break documents down into pieces those pieces usually words are tokens and so you could imagine. One simple way to do this would be you represent documents as a list of characters or letters. The standard approach would be splitting into words so you go to like a split on the space and then the third would be immigrants or phrases so we capture local world ordered as it for example with direct becomes with sir So when this is like the kind of classic workhorse in help would be a backwards representations. Basically just a corpus is broken down to a vector and for each word in the vocabulary you also get a number corresponding to how many times that word showed up in a document. And so you know just to introduce some languages here which we'll use repeatedly. For a given word in a vocabulary we can say we can have the document count which is the number of documents that had that particular word tocentype appears in a in the curbs and then term count would be the total number of appearances in a corpus. In then we can define a term frequency within a document as the number of times a token type appears in a document divided by the documentligso there's the number of words in the document and so sometimes that we might be a bit imprecise, but going forward if you tried to use the word counts to refer to integer, the number of times it competent occurs and frequency will be the share are the count divided by the links. So as an application of this in the political realm Monro and all or this is bighton words paper which is linked on the bibliography they recognize congressional speeches and ten identify words that are distinctive of Republicans and democrats in the U's Congress. So to do that first they run a topic model late to reached directly allocation which we will talk about next week to identify speeches about abortion, reproductive rights. So that's what is. They use a top model for purposes sick so this is it. Think pretty useful in the congressional record because it's very diverse what they're talking about. Sometimes the speeches are just like procedural there is saying we should not on air, not on the and those are not going to be very interesting politically and in the paper that they provide a number of ways to identify language that's distinct five of the different parties. so you know the Democrats at the left wing Republicans are the right wing party. This is from around to thousand and five I think they hundred six Congress and so you can see that when you just looking at the difference in the proportions of how many words are used, the most democrat word is to do it's a stop word and the most republican word is the T. And so this shows that this very simple metric of just the difference between the frequencies is not. There's not to a very good job of extracting distinctive tycoons. The second approach that they do is compete the law of auzry. Go for the words. This you would think this would help a lot because as racioit adjusts for the proportions of other works that are used and within this actually or living works. even worse because the most democrat phrase is bankruptcy which has in snow which have nothing to do with abortion right. So this is not extracting an interesting dimension of partisanship. Then you can look at the paper for the statistical the mathematical specifics of this. but they provide this interesting multinational lesbian model underlying language. If they specify that model and estimate the associated parameters, you get a much more interesting ranking of the words. So for within the topic of reproductive rights, democrats talk about women's rights for example their decision up to the family whereas republicans are talking about abortion procedure and killing babies. So you can really see how this kind of like difference in the difference in the framing of this topic really comes out once you try to these other token drinking methods and then finally this one they had a regularisation parameters to really to shrink most of the word premieres to zero while still maintaining this distinctive language. Yet that's great. Yes so not now. and I think that's a very good question of this specific ranking of the terminology and their sporting well in this context, but it seems quite specific what they tried so we actually do not know. they might havetried a hundred other things. So in terms of exciting we do not want to draw from this we sold always apply this method in every context, but in in this course you you know when we're when we're doing the required readings, presentations who were running the response essays. That's exactly the type of question do you want to ask the morons and their co authors they wanted. They're also probably selectively presenting this right and so some other questions would be used so I didnt even include. that's a good one right. Will this work in other domains? We actually do not even know if it works on other topics, even within the congressional record. They only try their reproductive rights at the abortion topic and so get these nice ranking of the words that we kind of know intuitively that if we want to extract this of slain dimension of reproductive rights discourse that it worked. but if it was another topic where we do not know such about, how would we know to trust the results So so waiver this is kind of. It's lucky that they chose this topic because the results are indicative. So in another domain where we would not have a lot of previous knowledge about the topic in mind to work and we might not even barely equivalent in another question or comment about ewhyheard some words so out of that supporter batter. So when you do a porter simmer on words it will replace them with these kind of placeholder suffixes. So words that end in why given ending in eye. So this is because like baby or bad they both become paid and but that's a good question. Do I about the presentation of this right? Maybe they should have used a limitizer instead because it makes it were to and of have to read the paper to even know it. but that's a fact. So I think and that's kind of related to you known they were saying that the difference in the frequencies do not work. The mean if they had just dropped stock words beforehand this would have looked okay right because they just dropped software. Then you can say women's rights their the Democrats abortioned baby procedures and publican that would actually already looked pretty good. And so this is the type of kind. Like what else could they have tribe alternative methods to get it the same place? That's way we want to start speaking about the you mean on horizontal axis it says frequency of word within topic. So if I think it's just how its the log of the of the frequency of the words. So but I Do not think it played a big role in our analysis. I Think it was just a way to add more space on the graph. So let's not road these types of questions. Okay, well, you know, is this figure effective? What else could they have done? What information is being concealed or left out? That's what we want to ask for. All these applications appears okay. So building a vocabulary. As already mentioned, this is basically the goal of frustration or tocanization. We need to think about. Well, how big is the space where our documents are being represented And so no. there is number of ways to do this and depending on your context it might not be do not afford it. but there's just kind of simple historical that usually work pretty well. So like any word that shows up in less than ten documents, for example, you know it's not going to be very informative for any analysis that you undertake, so you should drop it from the clear. probably. You could also impose a more complex threshold. sigh you not needs to appear twice in over twenty documents. I Think there's kind of limited returns to this, but I Like the ten document minimum or take the top twenty thousand most frequent words is usually a decent approach. A more subtle way to rank words is in the frequency is called to term frequency inverse document frequency waiting. So what? this does is in addition to avoiding information about how often as a word or a phrase shows up, that's weighted by how often it shows up across document in the whole countries. And it's inverse document frequency waiting in the sense that words that show up in all documents should be downweighted because they're not very informative about topics. So if this will kind of depend on your task. but let's say you you're trying to classify documents on breeding topics or you're trying to learn topics accidently at a topic model. The words that show up in every single document like the or a those are going can be very informative and inverse Document frequency waiting is a way to address that. And there's some example text at the farm on this slide to get a littlebit of intuition, but basic by this will upright distinctive words that show up often, but not in all documents yet. I Think that they have like Enpsycli Burn for example you would put one plus they smooth. there's a good stomping farmer btyso this is like one example formula but as you said it would be undefinend for some works this way. but they they add might made partner in this danger of implementation. So this is the the psychic learned implementation that is talking about. So there's a few different options but I think in the default one it's one plus the log of the document count. To address this issue. The intevenominer. But in terms of life for most tasks especially for doing it classification task this is your friend of if your corpus will fit into memory run stiff vactorizer on it to transform the words the plain text to earnings. When do you're ready to go you have a data set already that can be used for anything and so in terms of the preprocessing options it will actually already remove accents. Things like this remove capitalization and drops up. Words is the question at so the the idea Let's say let's see how this a regression model and am packed at up. Let's say you this will make more sense when we're doing were like comparing documents so let's say it will do this next week. But you want you want to compute the distance between two documents and if you when you vectorize documents as frequencies over words you might want words that show up and all documents to cut less and so if if you use this transformation you're basically re scaling the dimensions of the base so that words that show up in all documents means there not very distinctive you downright that dimension and so he think it's an important question because the while this will matter all for concise coinkedocument distance for other downstream tasks that actually do not make a difference and lets but I want to come back to your question next week. Buyout On think of it as it's releasing the corpus so that releasing the corpus representation so that dimensions are words that show up and all documents count less in those distance calculations. So stiff Vactorizer the is waiting. Its its optional so you do not actually do not have to do it and there is options for something. The if waiting on things like this time I don.' There has not really been enough kind of systematic comparison of whether if waiting helps or not but this will like only be task specific. so there are this. representing documents is like counts over words or counts over in brands is really the standard thing to do. Counts or frequencies and you can pick up an infinite number of other things to do right. You could say well in actual I Want take the log of the frequencies or just an indicator like for whether the phrase appears or not. Quadratic paradise interactions these winners often not done because of any given like text classification or top of motor polemhis just adding dimensionality without helping and there are not any kind of rules of them for beyond getting the ignorance frequencies. what else? What other featureization we should travel. I think immigrants is the way to go. and so just to talk about immigrants a bit more, These refer to converting lists of words to lists of phrases. so you know for this is a sentence, you have a number of bargains like this is is a sentence and then atrtrigrams are the three word fishermen. That sentence it is a sentence and this is useful for oclassifiers and things because it captures the local word order and I mean you can. This is a sentence is not a very good example, but like we talked about, you know the word tax versus tax cut or tax credit. You can emit like these two word phrases really in of a lot of important information in legal and other documents. and so this is why you normally use infants. and there's this link on the cillabus is that this is that code may be a little bit out dated now, but I think most of it is probably still accurate with this Google developer's text classification guys. They tried perturbing some of these different choices or use of standard text classification tasks and when you have long documented or relatively few documents then stiff weighted bargains is the base lie. They recommit which you know because there are so many different things to choose. This is a very satisfying to hear. I Think there really simplifies things that if you and the even they even give you like this. these told that if rose the number of rose divided by the document length is less than fifteen hundred. used tide water bridges and so they tried like words for his characters immigrants versus sequence which we're going to talk about later just bygraundi usually enough and then if I did and so this simplifies our or task. How many managers do you need in your vocabulary if you imagine that there's a hundred thousand words in the English language, the set of possible toward phase hundred thousand times on hundred thousand times to and so that is not going to feed in memory prison. The same Google Developers guide they recommend picking were thousand bygrands based on their frequency which I think is actually a really decent rule of them And so in the physicist learned stiff facterizer you can just say I want bargains and I want twenty thousands of them and it will do it automatically based on frequencies. but even twenty thousand any thing is in marriage cases is more than you need I've just like tired this in different applied classification tastes and even like to thousand features is often enough in adding more as diminishing returns. But I think that and well talk about feature selection and a bit for the if you're doing classification old let's sixty thousand immigrants based the frequency and a du feature selection to select predictive features down to ten thousand. A totally alternative way to get at this issue of high dimensionality is bashing bacterizer at E all year. Good question, right? So that probably seems a bit inconsistent with what I said. is it can to go about of painting rat? So what I would recommend is in general, you should include them frequent including those frequent words and what it would normally do is any words that show up in more than like forty percent of documents those get dropped and then after that take the option. Thousand Bees formulate this. Another historic would be drop the two hundred most frequent words that is usually actually a r by document frequency. So that's like kind of a representative of these southward usual words that are not very important, the most corporate, drop the top two hundred in the ten thousand flat. So but I think that the idea is that are the most informative words are the most frequent order, not that are formative and all the very internet words are also not that informative. So in the middle and a year we also look out at the dipran Yes, so there will be very few of those. so to be a bit more precise I would say dropstopwards and words that show more than forty percent of documents then make bargains from the remaining vocabulary and include the top to on tousan on that probably has. Not only are when we get to partings ansyntact and sometimes is to date because the right I show you an example earlier. Beyond a reasonable doubt that's like a forward phase that contains a sword but you need it to get the meaning, let of that. So these are kind of rules at Dumb and and it's if important to think the ways where they find not work for some cases is her on what performance did when need breeds or so the best. The most systematic one that I know of is this one Here is at least for text classification right they found they did is own array of stated text classification datasets and for the biogramds work as well as strike so we'll see an example later. This gains crown should appear at two thousand and ten paper at least in terms of interoperability like kind of the quality of the features. The programs are pretty nice in terms of getting phrases that are saved like in important narrative or political information in them. The programs were nice, but I think that for you just the quality of classified bygrands are normausually as good or there's diminishing returns as you had not at ignoring pills but another way to kind of dodge this issue is the washing vacterizer because with a washing vacterizer you can add even an arbitrary length of immigrants and they all get just mapped to an arbitrary idea. So has anybody heard of this before today So you know washing is this you known way function technology thats often used in cart criptography Right where you say basically you can take the string output and output a lower dimensional string and there's no way to get back to the original string. But that function is deterministic. So basically you can imagine making string and mapping it to a number. That's what a hash function does. And so I can build a hash function that has ten thousand buckets in it and any string will be randomly matched to one of those ten thousand ideas. But once I have the washing function, it's deterministic. So once if I have a washing function, I can then vectorize the whole corpus to a vector of ten thousand items. But what those individual strings include could be anything. so it could take words, Bigrams, programs, and quadrans. All of them get mapped to into this ten thousand dimensional hash space. Kethis is the illustrating it here under traditional vocabulary construction and specify the number of everywhere. so bell cats and dogs. but with the washing trick, they just get a random number. but that number is always the same so it's comparable across documents because there's collisions. So just by chance, the same number at the end of the chasing function represents multiple words. so that's what this is trying to to represent here. This is called a collision in the sense that if you only have ten thousand items in your hash function at in the output space, but there's an infinite number of strings in the input space so You can imagine that if you have ten thousand items in your hash function, two words have basically one out of ten thousand chance of being up to the same number. and so every when you're going backwards, every number at the other hash output space represents a lot of different strings. right? And love. Tune in your purpose. And so that's why the fact that this would never happen in practice, right were two or superrare that two important words would be mapped to the same idea. The way that you address this is you can actually use two washing functions And so if you use to washing functions is in the rare case where two important features are confounded in in the hash output, it will only be for one of the washing functions. enough in output. Yeah, there's two reasons that's that's by the main one is that it works for any vocabulary. You do not have to build it a crucially so new documents, it will vactorize them correctly and the other is so you do not have to have the purpose a head of time the whole purpose. So that's Ikapty. Very useful and it helps with dimensionality. So like you, if you want to use triggers for example in your feature or quatrans in your feature space, there's meusilians in them, but that will not increase the size of your questioning. A passion bactrizer. It's also very fast, like the computationally so you guess might have heard of it. The text classifier that Facebook uses it's called Fast Texts the Washing Bacterizer I be the celebrity compound technological in the fast Text of Text classified algorithm. So basically what it does is it takes documents and represents them not as the list of word ideas, but a list of hatched in grant ideas. So you get a bunch of these hash numbers for every document and those can be very quickly computed and can input into a classified so ththere'there.'s few practical issues, but also complicate ya, the can is alright. like to about isinin outfits about grids, the w he ears there or do to say ye let me plazamia That's actually nice analogy, right? So basically even in the standard model language for the same in type you you will have multiple words on it anyway, right? And this is just adding to that. Making it slightly is making it more more crook, right? The court's representation: Multiple meanings or words are loaded on to the same string anyway. but this actually just takes that a little wooden worth and allows for even phrases to be loaded onto a specific string. Yes, so this fast text model is. were going to get into this like week as or so. but like the the Fast Text Word Embedding model or Document embedding model uses passion processing and so you can embed the washing ideas the same way that ipludin bad words. So I would say it. I Would say it's under explore though so there are a couple of papers doing it. but I think that this is a very interesting technology that has not been not that to be I think in principle yes but I mean if all you if all you had was the hash ideas it would be difficult to check this. but during the corrupt verification process you can keep track of swapping alongside do things in is and then one is. I think it's a little bit unpredictable but would happen but at a minimum it might add anyone to your embedding tribe because there is going to be some other words that are included that are mapped him to men and woman's the she site you like. it's interesting. I mean you could probably do some of this theoretically like the distribution of world frequencies like this on the world frietsay who get sagazip's distribution and I think in the fast text model they allow for like a million stews or something. so Whether using embeddings they're using this hash fertilizer as well as allowing to for a million different spots and so in that case like collisions are very rare but I'm sure you could work something and based on the distribution of frequencies it would right? So yes so so I mean especially why no if you're talking about the original fast text vactorizer on social media data so you know they are like hundreds and millions of documents and sminicalliio and even for strangers or quadrans. So the that's an interesting question. I Do not think that they get it that directly. Was there another? Yet to be sure that as the stand directors of Poet, traditional economic construction and artists basic met one producing yes, Andritan a permit evident yes Exactly as so. This is great. We're get in to that like week six about getting into to embeddings. But but the standard toconizer will give you and time with a list of words, a sequence of words or account over words like that. One of the atectet is on one hop in coding or for account in coding and that nose could be limited. so be the. But you can imagine that a document is just to list of adhesives as well. so it would be the same thing you could apply an embedding look up to like a list of words or a list of ash ideas. Okay, him going to speed up a bit and then we so we can finish the next week. So reaching vactorizer is one way to deal with hydimontionality, another is to filter the vocabulary based on your downstream task. So for example, if you just take all of the world frequencies and get the correlati with your outcome variable. so let's say' we're like in the morning at All the fighting were its paper we're trying tobasicaly that was kind of feature selection right? They were trying to check which features were correlated with the different political parties and so there is a number ways to do this. but Key Squared for example would be like a classic feature selection approach. Let me think about what I should cover now. So one issue which were going to come back to the lot in this course is relating statistics to meta data. So let's say in the fighting words paper A you the number of Republicans is increasing over time, but also the frequency of the word kill is increasing over time like over like a twenty year time periodand you would measure this correlation between Republic and a kill. But it's actually just a spurious correlation that if the word kill is increasing in both parties and the number of Republican is increasing. So this's like a kind of a classic correlation virus causation issue. You're actually not estimating a partisan dimension in text yours getting this time confounding trend and this is something that you nwits not recognized very much an nip, but when you're started applying to kind of social science issues, then this is going to be everywhere. And so rather than just estimate these rather as do feature selection based on the rock correlations, you might actually want to try to deconfound some of the frequency voters or at least the outcome vector. And for example, if you deny all the word frequencies by year and then estimate the relationship to democrats and republican, then you're getting just a within year variation and excluding it this time for point. And so this is something that will will continually come in back when we're trying to get up more casual or more informative representations and correlations. This is just like some side note is that you might have. you might want to get rid of both the time component of this correlation but also the geographical component. So link to Congressmen from taxes. for example, they might. you might have the word congresor from taxes might use the word to kill up often times, right? And so in order to remove both the year and the state variation, you can use a literary reason to do that and for why it would be the outcome and x would be the frequency of each word. You redress that against less categorical variables and take the residuals and then use those and your feature selection or machine learning to bask. So that's done. Found did feature selection. We're right at at four o'clock now. so I want to wrap up. And also for those of you who were doing a project I'd like you to stay for five more minutes to talk about it. We're going to write But now and I'm going to finish these slides at the beginning of of next time. Are there any questions or logistical or otherwise before we wrap up right? Thanks and will see you next week at the same time here in the same room. If you're interested in doing a project please stick around for a few minutes. Sigiisi A but cut in be on the tide process a holeayat at a of of more is is said to always if you have to wind up where the not object is too yes please he So if you have not sign for a peak yet and your he man to hear about it feel free to is to cart but you just go to hear about it. So if it's okay I'm just going to start talking and is freed to go yesterday. There is an additional to optional course credits for a course project. It can be done individually or in groups of up to force students and it's really just doing an application or a act based on what we're the continent of the course. so it's quite broad in terms of the topic. but usually its an app project on some social or political data or legal data but we're quite go and you can choose some thing and you're interested in. So just to give you some examples of the previous year's projects, so one of the top a better startup feature Startups: the Deep Judge team. Their project or their system started as the course project in this class and so they actually built this context sensitive legal search engine which's pretty amazing and they went on to get some by funding their headquarters at the I center. Now another group did some environmental regulation analytics and they want it in a swift grant. So just to show you how successful some of these projects have been, a number of them have been published and one on doing legal language modeling, another on inspiring legal language, another on medical documents summarization of medical documents, one before biting with a old student. Here he built a basically an automated question answering system for a coding class and then no Langu who's also a old student here publishes an attractive summarization system using reinforcement learning and so those last to those are individual projects. So even if you do a project by yourself, those have been in just being successful. There's a number other projects that that I think they have good chance of being published at some point either an top conference or a social science journal. We made a partisan treat generator when standard. Stump Bombelli who's another students here. She did an analysis of immigrant attitudes and historical newspapers. One group did and deep it like normal nets and instrumental liberals. The kind of caused a little of paper. One group did parties and question answering. it does not have to be text. One of the students did an audio an analysis which was fine. if we want to do audio or other images we're not in cover that in the course material. but you're welcome to do that for your projects and then some of them have just been kind of classification projects in terms of picking a topic, You're welcoming to pick one on your own and I can provide back if I think it's a good topic or not. for how to modify it, some of you are ready. Asked about this. we have a list of suggested topics like we have a short list of topics that I think are kind of interesting to do right now and then a longer list if you want to see that. just send opera, email wearing people, less economist everybody list of the same time. With some doubts it would also be good to think of which maybe one or two or three topics you're interested in because then I can provide some thoughts about advice about what's doable for course project One you take a look at these slides from the gethub, but once you formed a group, send us a list of their team members and it be useful. You do not have to be right a Cv for this or something, but would be useful. to know what courses you' we take in and so that we can set expectations and things and then some of the projects have either new or after we advise them but there's a number of other project advisors that will help out this. So for those if you need help picking a topic we can need to disc test. maybe it could be Onzumor in person and if you already know your topic we will also need so that I help you get start. So those are the points that I wanted to to bring up for Now you do need to say you do not need to pick a top topic for a month a few weeks from now so this is none of us urgent. but I'll send the list of project ideas A we go from there, are there any questions or concerns at the moment A right to keep the such with any questions and will see you next week. Teach changes are sejetd. |
| 2 | ASRnlp_law_lecture_week_3_part_1_v_2_c_transcription_3.txt | 07af2cf9-15a | ASR | I think we do reading as started. Thanks everybody for your patients with the hybrid learning experts. Just we face a bit of a tradeoff here. where in order for me to see the slides here we have to have those kind of zoom overlays and rather than me having to kind of a switchback, keep looking back and forth thing mass iksher lwilld but better so we'll try that this time. I Looked at the queue and a page this morning I Did not see any questions, were there any that have been added since then anybody wanted to to bring out loud off for those on the home I Can not see the chat so if you have questions type it in there and African can let me know. Do see any questions on there yet? Yes sure yeah each. so it did not see any there this time but yet we have Consimantos in the mood. swell do that going forward. So we ended a bit early. we did not finish the slides last time so I'm going to go through the rest of the week two slides today at the beginning today and then we'll go to the week three slides in a few minutes. We are in the middle of the discussion about immigrants which are perhaps so phrase representations of documents and why are used phrases is because in English as in many other languages the meaning of a phrase is more than some of its parts. A There's and honest clear examples of this are what you linguists would call collections such as kick the Bucket. So these are these sequences of words that you can not just say adding kick and then and bucket together here in any arrangum order is not the same as kick the buck at this particular sequence and this is what you would call on compositional non substitute So you can not just like swap and no synonym and maintain the meaning and non modifiable so you can not. These collections do not respond to the grammatical rules the same way as as normal language. So to get at capture collections in language a nice way to do that's a positive mutual information. So basically two words in a sequence A and be. If they are independent, then the probability they occur next to each other is just the the product of their probabilities that they happen on their own. So that's the denominator here. the probability that each word shows up in languages like me, the corpus frequency. If you take those two words independently, then this the numerator here. which is the probability they occur right right before and after each other. It will be the same as the denominator. And so if you compute this metric or point living mutual in motion between two words in this specific bathroom sequence, you can really rank the words in the vocabulary by their collocatiodness. They're how often they can relative to who occur a part. So I think thisactually makes for a nice feature selection approach. We're tough ways to reduce vocabulary. One way to do it would be on frequency for bankers and trainers. The pointlized mutual information is astonished to it because you capture these distinctive phrases. So back to to a social science application of of distinctive immigrants. So whether we are talking about feature selection approaches supervise feature selection is a way to identify words that are related to some metadata and in the morning at all Fighting words Paper We looked at words related to reproductive rights that were distinctive over republicans and democrats knocked and Chipiro to thousands and ten is a paper from media economics using that same type of machinery. but for a social science question of what drives media silent newspaper media and so what they do is they build this parallel corpus of next from you's daily newspapers as well as the text of political speeches from the U's congress and they have doing some pre processing So this is why the this after unharmed so you know more was looking at single words Ginscoach appear are looking at bargains and trademarks or two or three word phrases and they then use a related metric to mnmonro all do to identify polarizing phrases. So what they do is for each phrase you you have some frequency that a word shows up for democrats, a frequency use of the Republicans, and it a frequency for other phrases. and they compute this Kysquared metric for each word, which is the same as the key squared metric that you have Gipychit Learn for example for features, action. it's just the latest statistic for the equality between the political parties of using that case. Is it better? Much better. So this will be easier to read now, but still probably too small. Check out the slides if you can not read it from where you're sitting. But in terms of using these nip techniques for social science, something that readers will always ask for is some examples or some way to do some qualitative valuation of what the system is outfitting. And so here they show the lists of phrases, bargains and programs that are most associated with democrats here on the left and Republicans here on the right. And if if your European or your switch you I do not know that much about the you's politics so this might cope unomoush to you. But for those who are familiar with the Us' political setting these phrases are kind of intuitive and for four republicans for example used well site cell and around this time there of her thousands they were talking about ban and to his self research because it was time or in minority time cells that tourist reality reproductive rites and they're talking about kind of legal issues and government spending like this is in part of the Republicans being worried about about expenditures and Democrats on the other hand are more worried about the work war. So so this provides some kind of just qualitative evaluation that the distinctive phrases for democrats and Republicans they make sense this like way to add confidence or will trust them at bit more And so this is the main result from the paper. We can not see this, but this is just scattered plot of the newspaper slant so computed using this those distinctive phrases. So how often does a newspaper use republican democrat phrases and then the horizontal access is how many republicans there are in that in that congressional district. So you can see that this is consistent with the demand side media slant. So rather than you you get a republican, owner of a newspaper takes the newspaper republican. The consumers for a newspaper market are more likely to be republican and the newspapers respond to that by using more republican language vice versa At yes so this is it could be that the newspapers introduced Republican language and then over a time people became more republican for example Than there's the short answer to that is they looked at when the owner of a newspaper agent and it did not change the language. So that's that's somewhat imperfect response to that question because there's other research showing that when the newspapers do change their me as people do are persuaded by that allowing talk about an example of that mix to be in to actually know so one other just kind of like parenthetical feature to detail that I want to mention it is named in the recognition so these collections that is mentioned of low kind of high phrases with high prime phrases that tend to do locate you will get these nice phrases such as this order or meriquery which those words that that phrase is quite distinctive of those individual words. But if you want to do if you want to try to capture proper bounds or entities did a name indie recognized is the right way to do that is so space for example. it will have some examples in the notebooks where if you put a document into space it will identify different categories of entities such as people, organizations locations and so this could be useful in a range of things. So like for example if we look at news articles and you want to identify articles about a specific story then getting it overlapping entities would be the right way to do that. So you know another kind of shortcut to getting named inities would be it actually just like search for proper bonus which is a part of speech part of speech are kind of the basic functional categories for works in attendance. Solution is an object of verb is an action and adjectives and attribute and so on. And the the standard part of speech tag set in English has thirty six tag once even there and but otherwise kind of universal tag sets that work across languages as well, but kind of. These basic functional units in language are pretty universal across languages where a noun will refer to an object. So if you getting a topics in text and looking at bonus who make sense, adjectives correspond to attributes and so if you're concerned about the symptoms in the text to maybe you want to focus on adjectives and that's parts of speech. So okay, so one question you might have on to practically is why on what to do with out of vocal words So were not. We talk about the washing concrete last time. This is a method that we out. Vocabulary words are no problem, but if you want to limit your features that to ten those fifty thousand items, there's going to be a lot of words that that are not going to show up while you take it to new data. So what to do? It does. Think the standard is used to drop them. That's what psychiclearn Tfidvacterizer will do. By default, you could replace with a special unknown token. This is what is done in modern toconizers such as the housing face town years. You could replace them with their part of speech tag. That's kind of, I think a nice compromise or you could actually just have again auxiliary chasing vactorize for those out of vocabulary works. something else is thought. Bone was right in these slides. is that it you could replace with the hypernym from word not so like why if you see the word you no trout you could replace it with fish something like this. So another social science application for the specific case of parts of Arts Muse tags that is found is this Netsmer, Lamair and Person seems two dozen into paper windward sweat. They find that and the words that people use in loan applications are predictive of whether or not they'll they will pay back to loan and so maybe you guess withered this light already. But imagine that by the website exists where they have people like protein peer lending so you can actually lend to strangers on the internet if you want and one of them says I am a hard working person marry over twenty five years I have two wonderful boys please let me explain in why I need help? It would use the three thousand dollars alone to fix our roof. Thank you God bless out and is promised to pay you back borrow number two days. While the past year in our new place has been more than great, the roof is now leaking and I need to borrow the thousand dollars to cover cost of repair. I pay all bills eg. Carloan's cable utilities on time. So which of these borrowers is more likely to default Razor hadything its borrow number one of number two. it's about equal. may be a little more borroughor number one, but it's the case as borough number one actually is more like a default. And so in this paper, they have this really amazing data set of these types of snip bits and then date on where the people paid back the loan. And they find that loan requests written by Dealt and Bears might like to include words related to board's family, financial, general hardship, mentions of God I Think that's superinteresting the near future in pleading for help and also using verbs. In present, vendors feature tents so you can see that like they really are. These kind of tells in language for whether you pay back to eleven or no. And so this is one interesting craft that they show in the paper. Words like Reinvest Wedding student learn summer side. These are words that if you see this person is going to pay back to lean, they're pretty reliable. but then words that predict default actually he go. God and family is like a category if you see that in a lone application, you should stay away from this purists. which is maybe it makes sense in retrospect, but you can see like you know God bless, pay day loans, medical death. I need help. Let me explain the situation and these are pretty striking. And like this figure, they I think they used a topic model for this, but they had a way to classify words or documents into topics and these are the words that are associated with default but also typically related so they tend to show up in the same sentences. But they did that based some automated method so we do not have time to do this today. but we might come back if we have time for these four Social Science Applications papers that we talked about last week as practice. for this course and for the response essays. It's good to be able to look at these papers and think about them and ask these types of questions. So what is the research question? What data set is being used in? why? what is the paper trying to measure using the text data, and why? What's the purpose of it For the entering of research question? What no method is being used? Why was this method chosen? How was it validated? What else could they have done? And so you knowthese papers or from two thousand and nine, Two thousand and Ten inilp has changed a lot since then, right? and so often times the method they chose ten years ago it would be different now. Andasomething that that's what is that we want to think about it. time. Main results from a substantive social science standpoint: Why are they important You? The results seemed incomplete or non robust. So for you guys as mostly exams students on social science I would not expect you to know this just from run the abstract but when you're reading the paper like in the introduction they will normally have a paragraph explaining what are the other papers and so you do not need to be able to verify this objectively. but you need to be able to point in the paper where they say it and then what are the limitations in open questions? So again lie Now these one of these questions are and stars than others. but the in terms of what is the research question and what is determined in how they relate to each other. This and if I would be best to start practicing and thinking about so we talked about each of those questions a little bit in the race related research paper. the more fight and words paper what drives media silent then wind words slept so want to start practicing for the to say is tape I answered your question true love on your poll if it was saying a tenor's first montage she has on to train a time to to skin or if your zoom can you go you Asylum of right where go on the next slides. Now nobody knows how to get rid of this right? This everything only. So the week there lecture is about unsupervised learning and so this is really an example off what will be frequently doing in this course which is representation Learning Feminism reduction to extract relevant information from high dimensional data. Otherwise, in this text, next week based on the next two, We week four and week five are more on the supervise learning side where we for example have labels of documents and we want to teach a machine to reproduce those labels. If we do not have that then we we still want to extract relevant information and unsupervised learning is a way to do that. It's where an algorithm discovers themes and patterns in text and then we look at those topics or patients to interpret them so well we'll see is that you know they really is not like some kind of fine distinction here that is always clear cut. Often times will cause supervise learning to find the themes and patterns and unsupervised learning will often used to pursue a known goal or for making features for predictions and this is related to what we just saw at the end of the last weeks slides. which is these methods. If you were taking a regular plant course just learning the methods would be enough right? That would be tiny of it. But for us this is talent for law and social science and so we want to know what is the research question or the problem to be solved and then in service of that question, Why was this corpus or or data set chosen? How did they preprocess it? Do they produce enough descriptive statistics or visuals to really understand the data For unsupervised learning like a topic model, What are we trying to accomplish? What are we trying to measure by getting those topics and so other steps out of undertaking an undiscovered learning project known, pick a model based on your goal to probe sensitivity to hype parameters and try to provide indication that the model is delivering what we want and then number step forward. Empirical Analysis is doing a scope of science analysis. Usually this is testing some hypothesis like what drives media slant. For example or so the first set of methods that we look at is document distance. The classic way to do this is a set of algorithm is called text release algorithms like Smooth Watermen which really this is like Pleger is in detection software. So you know when you pick to documents and you find a sentence in one document is incidentally there and other documents. If the sentence is more than like five words there is something that's actually very unusual to five six words. Once you get to that length there's very few of them that are at tend to be shared across documents and so if you look for these shared sequences then you can really detect pretty well whether whether word text is being rescued or copy in tasted And so that's actually a shortcut that is much fat, so smooth water Inwhici looks for these shared sequences across documents That's computationally expensive. so for large corporate you can not live to it. And looking for shared hatched five grams like these quiet grams like so five ward sequences actually worked pretty well as a shortcut. If you have two documents that share a lot of five gardens, then they they're likely from the same source. So a more standard way of of doing unmet comparison, which is used for example in information information retrieval is basically just take factors of the word counts or the phrase counts in farm counts and then take the cozine similarity though those factors. And so that's what we did the last time is we take a document a list of words and convert it to a frequency distribution over words. And generally we will use the Tfida or inverse document frequent red word frequencies, and then you have each document as a non negative factor in an index space where index is the vocabulary size and so documents, Rays and similar documents have similar concerts. So you can imagine if you take a document to multiyou duplicate it so you have two documents that are like one after another, those rays are still pointing the same direction, right, but one is looted the other, but in order to make events of varying lengths more comparable. rather than use a Uclidian distance, we use this choice of the vector trees angels and so perfectly collinear documents have a concise similarity of one of documents are orthogonal means they have no words or phases in common, then they concise similarity of zero yes to and that depends so it will often provide better performance if the voters are normalized for example, so you just divide them by the norm poor even if you standardize them. let use as the standard secular in Psychic Learn which will center everything in the divide every dimension by the standard deviation. that will often improve performance and so if at those are the two kind of standard normalization approaches that you would use besides stuff and this is art of broader issue is that you often do not know which one to use when you're starting off and unless you have as if you have a metric for deciding which documents stimulate which documents you think should be together then you can evaluate it. But this is part of the broader issue that we'll talk about at the end. but often times for the document comparison document distance the method you use could to make a big difference ad so and if you do not have any document labels then you might have a problem and so in that case using the standard stupid of weighted in grams. the fact that so so many people use it is like a reason to start with it because people should have an intuitive understanding of it and it's become popular. And so the fact that we be Twenty Five indexing so like a lot of the standard indexes, comfort search and go in for databases queries use this way is a good reason for using it because at least just in practice people tend to agree with the results but will come back to that. So this is the formula for cuisine. Similarity is just the dot pro dust normalized by the norm of the victors. And you know in scholar you can get the concise similarity between all rows of a matrix or in just one line for example. So if you want to try to rank documents by their similarity, that's the way to do it. So if you have not done this before it might surprise you the first time that you run close and some party and you get Azilian metrics. So in times in this one so if you have a thousand simultaneously get a million scores back to be careful about that to if it will downright terms that appear in many documents. And so as it mentioned this stiff similarity is used in be the Twenty Five and elastic search. So that's like this standard robust matrix that's used across many domains. but this is a custodian one but there are other distance metrics that can be used. My sense is that there is not enough research kind of system out of comparing these, but this is what we have. So a research paper that we have linked on the syllabus that uses both be Twenty Five, the tariff similarity and the text Rescue Smooth Watermen. It is this burgus at a legislative influence detector. So what they do just to summarize ndthewill talk about it a bit. They take built texts legal texts across states and they compare both laws between states and across states as well as similarity to take to the model legislation provided by lobbying agencies. So here is an example they show in the paper of basically two bills between two states that are almost identical. So you would not be able to tell this but if the title of the bill is or of whosance sponsoring it or things like that. But it turns out that the same legal text is proposed in different states as it similar times. and so they use this text similarity matrix to look at the number of bills that are introduced from Black All which is this conservative lobbying group and they also look at the number of bills that are introduced by amid a liberal lobbying group. and I Think it's interesting that you would not be old to know this without this text similarity metric because they do not advertise it so likethey're kind of secretive about it. That's these bills that they propose on their own website. When a legislator is actually proposing them, they do not advertise that. and so it shows that they are trying to kind of secretly influence the legislative process. and this is a good example of a paper where it kind of makes this simple point of by taking a corpus of legislative documents and providing a measurement of similarity to these lobbying group documents. and you get this type of out of graphics and so just pink on your own. What is the research question here and why is it important and separately, what is the problem solved And I'll come back to that the second. What is being measured and how does the measurement help answer the research question. So in this class in the response essays in the exam, these are the types of questions that I want you guys to be able to answer and so noticed that it said what is the research question and what is their problem solved. Does anybody has not to guess why it put those separately, How would you distinguish those in this case or or generally setters. That's an interesting right. Let me see if let's say you could solve the research. Let's say you can answer the research question of nowhat is influencing legislatures. There could be a separate policy problem that just actually run this paper does not solve. Is that what you have had an ad in mind at Good At For Sure right? So actulyye to think it's even better. So it could be that this paper has a research question as they'd like to answer, but there are such statistical or computational other problems that have not been solved and so they actually could not answer the research question yet. And actually this paper is a good example of that. Where the problem is these lobbying organizations are secretive about further influence on legislatures and so this paper was to solve that problem of matching up the legislation with the model bills to measure. This shot is the problem that they try to solve, But the research question is what is influencing legislatures or what is the influence off of these model bills on legislative. This is important because this is not very democratic right? Like these are kind of the lobbing organizations and they are adding policies in this secretive manner. Other any questions and or reflections on that so ye comment or analyst at treatment in words and a person is brat more about brother as hasatist plan. Yes that's great right? So let's actually's another nice way to frame that distinction that there could be this technical problem right off measuring the influence of these model bills on legislation. but the research question is is this happening in the first place That's what's relevant for policy and for social science We there another point right? So he so in this case like the problem is like this technical question of how to to measure this influence. That's great. Okay so the next topic is do mention a reduction and this is something that elementary you as gin to prolyscene and other computer science classes. This picture is from the Aurelian Garden Book which I think that on the syllabus, the Aurelian Garden Book increase on machine Learning Physics learn is really the best way to read about the method side of what we're doing in this class. and the issue is that even if you have high dimensionality, data is not just randomly distributed across these all the dimensions in the data set and this latent dimension. So the Swiss role is an example of one is what you do call manifold and that manofhode can be learned. That's what dimension reduction techniques are designed to control this Swiss role for example or identify whatever other structure there is in other data sense. So what dimension reductions have we already tried? So let's think what we did in week one and week two in already today have we already done some dimension reduction to somebody? Want to think of a mission, an example or two? Ye totally right to the washing function. You go from having a vocabulary you know millions to acabular a tin thousand. Definitely dimension reduction. And actually even just the process of taking a plain text document and representing as a vector is dimension reduction. So the washing is an doing kind of two stages there out the totally year. Yet like that's compression right? So like unmoving software that's like basically taking out the extra noise data from the text even dictionary met this. Its is really extreme dimension reduction. like her you're saying if I want to say this is the bit that what I'm interested in. So what we've already done in the last two weeks and what we will do further down the road In this course most all of it can be understood as dimensioum reduction and the classic approach to dimensioum production that you would learn like to linear average class is a principal component analysis or a singular value deco position in and this is just as in the same way that we no and twenty five stiff Immigrants is like this kind of workforce for document similarity. Pace is its workforce for dimension reduction and the auralian Imaginary Book has a nice discussion of of this if it this new to you. But the most important practical piece is that you can just extract these these directions in the space, these dimensions that explain the most variants using just these three lines incycilar and you can imagine that this is a preventative algorithm that it captures the first dimenso that explains the most variant, takes that dimension out and you' get a second dimension, explain the rest the variants in the remaining data up until you have allowed the data explained and you can imagine. Rather than take the flood matrix of word frequencies in stiff matrix instead, you take the first ten principal components of that and then use that as your data representation. The problem with A is it? or I would say the advantage of Pace is that the distance metrics in the space are approximately preserved. So for many, perhaps most data sex, the tariff similarity, or the stiff distance between the full matrix of immigrants will be quite similar than let's say, the first hundred principal components of of that Natrids. So you could say like, rather than have an hundred thousand columns, you could have an hundred pounds and the distance between observations would be preserved, as that's very convenient if you're doing document distance. This type of dimensionof reduction could also be useful if you're doing supervise learning, so you could actually then use the principal components of the reduced matrix to our procedures, but this might not work well. This has not often done, But the other major problem with Pace is that the dimensions are not interoperable, so it's quite rare. If you lie, look at the first principal component or the tenth principle component. look at the words or phrases that are correlated with that. It usually do not be interpreted as a top. This is why it will. Beginning into other approaches to dimension reductions, there's topic models which do have this nice feature. It's something that is think's quite popular like in recommended systems but I have not easy to but none. Negative Matrix factorization in is it's a more interoperable. usually you get more interpretable outputs within me than in place so even try a that and Pace and Nmf it will dimension reduce your data to like a set of like tin components. Let's say where all of these are continuous and so rather than rather than have a dense one hundred thousand colunmatrix you'll have a dense hundred column trip. If you want your data to be sparse at the end and you want to just separate the data into groups then on the society that's called clustering is recommended for that. So may means clustering. Its an along with them that you take data that could be distributed in various ways with in a space and campaigns. Clustering will automatically find these cluster boundaries and assign your data into different locations based on the geometry. In terms of you knwthe terms of fixing the number of clusters this is a Hyperpremer You have to decide and if you end up using Cinemas clustering you should read this chapter in in the early Hungarian book because you can use something called the Satellite Score to find the optimal number of clusters. So in terms of other clustering algorithms to others that you him I think Kmeans is again we just like this standard approach that often works in different kinds of data but you could also try a medoid clustering which gets it medians rather than than mean asteroids do scan will work better if you have data and this so where you need to identify these continuous regions of destiny. Most of the time you do not know what your data looks like before this so you meet as why the sense came into clustering and to conclude this section on clustering just to give you an idea how this could be used illegal or such as science applications in Ganglemere and Ward law they apply a modified clustering so a clustering approach to different contract clauses and then you can imagine the clustersintroid as like a K of standard contract, the Boiler Plate contracts and then a contract that's kind of far away from this Detroit that's an outlier customised contract clause and so you can imagine that this is useful for a descriptive analysis. They can say you know what types of contracts have clauses that are far away from this indirect are more customised and they find it the larger deals so that the more valuable deals have more customised contracts which that make since deceptively probed and Phillips is a paper where they use in K filings so basically filing with the Securities and Exchange Commission and so this's like kind of finance that they apply a text clustering approach to the business description sections of those contracts and so this is basically a little texting bit describing what the business does and so then they can. if they apply taxes clustering to that, you get these kind of industries or these product groups and so and they use that to basically to analyze how companies are differentiating their products from each other and they do one antitrust analysis to that as well. Only so that's clustering. Were going to take a break now and will come back in eleven minutes at Three Fifteen to talk about top models. Ice are words heano on as a plan who they are inflicted for. Its up it right. We're going to resume with topic models. So as we mentioned a second ago, the one of the problems or the the disadvantages of our seeing pace for example as to mention and is that the resulting topics or factors or components are not interpretable and this is beyond the fact that if you have like an in grain representation of a document which is musiclikea vector or and hundred thousand numbers in it that's not going to be interpreted either and so ideally you'd want to look at us just on it like a single number. To summarize a document like this one is about law. This one is about policy Things like this and topic models are a way to do that. So the core methods and topic models they actually come from basin statistics Yes, Skleed is of pub his. it's I think the might have to go with that that stologay right on Toastis statue and on pins. So even though these topic models these methods were developed in computer science as way for information extraction, statistical representation, documents, marization and they ended up being quite popular and useful in social science. So basically starting off with documents you could then have to set of measurements that are on the topic, shares of different what do people talk about in toilet models will find a nice way to answer that question and this is. what's useful about them is that they're interoperable. So again, the work force relate to the depot model that is often used lady or lean dricklat allocation And this model basically assumes a structural model of language that each topic is a distribution of words the some set of topics out there in a corbus. Each document is a distribution of topics and so then you start off with a document being fifty topic one, fifty per topic, two and use sample words from those topics and that produces a document. And it turns out that mathematically geometrically is just another example of matrix factorization in the same way that pace and if we were factoring matricies and so what we assume is that there's a topics and that's something that you can choose when you're setting up your cot niwl. it's really the only thing you have to choose when you're setting up lady besides pre processing. So that's another reason that this is polar because is pretty simple to get a nice result. and like Pr, Nmf lady will take in your corpus which is represented by one bigmaker its x and factorize it down into two smaller motors is a document topic matrix and a topic term matrix and this is borrowed this scheme from some of branding's credit slides where basically all you a depict is capital kit here the number of topics and and then your label. read through your corpus and infer from the word counts in each document what are the or is the sound working now can you indicate in the chat? Thanks Got all of these details are as idea. you can actually just train kids withlke five lines using payments gensim and in the end you get this nice kind of statistical highlighted tool where you take in a document for example this newspaper article from Henoolix sites and it will identify the words in that document that correspond to these interpretable topics of these constellations of words that try to go occur and usually this works almost all the time. Where you get these topics like this one's like genes, genome sequencing, genetic that'sgens topic right and an organisms survived life that's like a biology topic and the computer numbers competition predictions that's like statistics topic, computer analysis and say they're starting off with try they behave that they spin to right well do this the last time. If it does not work now then I'll just others finish in the room without the zoom and then ultra record a separate version the zoom stream. So as I mentioned like you know so one you have lady trained you can then take out any document and get a a set of topic shares for that document. or you can just take the highest probability, highest probability topic and that is the time for the document and then once you have these paper proportions those then can be variables in a social science analysis. So an example that we linked in the sillabis is any calacatalinack's two thousand and sixteen paper. She ran a topic model on Japanese political speeches and identified topics that were related to national or federal or policy issues freeze topics that are related to local or which will talk about. So everybody on the zoom i'm I'm just going to it's going in and out so I'm just going to finish the lecture in person. In in the report are another version to add to for few guys later. so I realize that that's a bit inconvenient but we'll have a backup next time. |
| 3 | ASR-whisper-rpunctuated_Noam Chomsky, Fundam_1669853561_0_part1.txt | fed834b5-a04 | ASR_cleaned | Well, I'd like to, in these two talks, I'd like to talk about some foundational issues, in particular with the most important ones, I think, namely, what are the fundamental computational operations that enter into constructing syntactic objects and why these and not other ones. And It turns out there's quite a lot to say about that, since the last time I talked here, there were many problems, some solutions, I'll get to this in the course of the discussion as far as I can, but I think it would be useful to begin with a couple of comments on something more general, namely: what are we trying to achieve altogether in studying language, many different ways of looking at it. These Questions I think are in many ways more important than the particular technical results. They raise many questions about what is an authentic, genuine explanation, a genuine solution, and what is a maybe very valuable reorganization of data, posing of problems, often posing as solution, but not really achieving it. These Things are worth thinking through, I think. The Basic issues were formulated, I think, for the first time, quite perceptively, at the outset of the scientific revolution in the 17th century. Galileo and his contemporaries, who were raising all sorts of questions about received wisdom, turned their attention to language as well. And They expressed their awe and amazement at the miraculous fact that with a couple of dozen sounds, it was somehow possible to express an infinite number of thoughts and to find ways to convey to others who have no access to our minds everything that's going on in our minds. So In their own words, which I rather like, I'll quote, they were awed by the method by which we are able to express our thoughts, the marvelous invention by which using 25 or 30 sounds, we can create the infinite variety of expressions, which having nothing themselves in common with what is passing in our minds, nonetheless permit us to express all our secrets and allow us to understand what is not present to consciousness, in fact, everything we can conceive and the most diverse movements of our soul. Galileo himself regarded the alphabet as the most dependence of human inventions because it had these amazing properties and also because, as he put it, it allowed us to express all the wisdom of the ages and to it contained within it the answers to any questions that we might pose,. kind of like a universal Turing machine in our terms. The Port Royal Grammar on Logic, actually which I was just quoting a paraphrase of Galileo, had many insights into logic and linguistics,. it's in many ways the basis of modern logic. There was a rich tradition that developed exploring what was called rational and universal grammar,. rational because it was supposed to provide explanations, universal because it was concerned with what was taken to be common to the common human possession of language, was seeking explanations including descriptions of even the vernacular which was quite surprising at the time, innovative, but mainly explanations and universal Trying to find what's common to all languages. This Tradition went on for a couple of centuries, many contributions. The Last representative of it about a century ago was Otto Jesperson, as he put it, his concern was how the elements of language come into existence in the mind of a speaker on the basis of finite experience, yielding a notion of structure that is definite enough to guide him in framing sentences of his own, a crucially free expressions that are typically new to speaker and hearer. And also beyond that to find the great principles that underlie the grammars of all languages. I Think it's fair to, you have to interpret that tradition is metaphoric, often vague, but I Think it's fair to extricate from it. The recognition that language is the capacity for language, as well as individual languages are possessions of individual persons, they're part of a person, they're shared, was recognized throughout the species without significant variation, and recognized to be unique to humans in fundamental respects. That General Program falls within the Natural Sciences, within what these days is called the bilingualistic program. Of Course, it ran into many difficulties, conceptual difficulties, empirical difficulties,. the evidence was pretty thin and nobody really understood how to capture the notion,. Jespersson's notion of structure in the mind and what is that that enables us to develop, construct in our minds infinitely many expressions, and even to find a way to convey to others what's going on in our mind, that's called the Galilean Challenge, which is still extant. Well, All of this was swept aside in the 20th century by structuralist, behaviorist currents which very typically adopted a very different approach to language, taking the object of study not to be something internal to the person, but some outside thing. So Maybe a corpus, an infinite set of utterances, some other external formulation. And You see this very clearly if you simply look at the definitions of language that were given through the early 20th century by the leading figures,. So for example, a language is a kind of social contract, it's a collection of word images in the community of speakers. For Leonard Bloomfield, its language is the utterances that can be made in a particular speech community. For Harris, it's the distribution of morphemes in a set of sentences. For the philosophy of language, say, Van Quine or languages, as he put it, I'm quoting, a fabric of sentences associated with one another and with stimuli by the mechanism of conditioned response. Elsewhere, an infinite set of sentences. David Lewis, languages also took just languages, a language is some set of sentences which is infinite. Both Quine and Lewis crucially argued that it makes sense to talk about an infinite set of sentences, but not of a particular way of generating them, which is a very strange notion if you think about it because these are the leading logicians and philosophers. You Can't talk about an infinite set in any coherent fashion unless you have some characterization of what's in it and what's not in it. Otherwise, you're not saying anything. But The Behaviorist,. the pressure of behaviorist beliefs was so powerful that the idea that there could be a privileged way of generating that infinite set was, as Quine put it, folly, Lewis put it, something unintelligible. But Whatever any of these entities are, they're outside the individual. The Tradition was completely forgotten, people like Jesperson, the last representative, were literally unknown,. There's good review of this by a historian of linguistic. Julia Falk who runs through the way. Jesperson was disappeared in the first half of the 20th century, and the whole tradition way back also. In Fact. To this day, the even linguistics historical scholarship is pretty thin, it doesn't barely recognize any of the things I've mentioned. So Returning to the forgotten tradition, by the mid-20th century, there were clear ways of capturing the concept, the notion of structure in the mind, Jesperson's concept, touring other great mathematicians that established the tools for addressing the Galilean challenge, something you're all I'm sure familiar with. So Jesperson's notion of structure becomes what's now called the I language, the internal generative system, finite system that determines an infinite array of hierarchically structured expressions that express thoughts insofar as they can be expressed linguistically and it can be externalized in sensory motor systems, typically though we know not necessarily sound, we can call this the basic property of language. Well To meet the Galilean challenge, there are several tasks that have to be undertaken,. The main one, of course, is to try to determine the internal languages, the I languages of speakers of typologically varied languages, a huge task,. Then the question comes of how a speaker selects a particular expression from the internal I language, then how the expression once selected is externalized and the inverse how the externalization is internalized by the here, the last two tasks are both input output systems, we kind of grasp how to study those, and quite a lot has been learned about it over the years. The First of them, how the speaker selects a syntactic object out of the infinite array,. That's a total mystery,. there's nothing to say about it, that's true of voluntary behavior generally,. So actually here at MIT some of the two of the leading specialists on the neuroscience of voluntary action, Emilio Bizzi, Robert Adjemian About a year ago wrote a state of the art article in which they discussed how, what they know about voluntary motion, simple, not language, simple things like lifting your finger, you know, And they said well, They put it as they said, fancifully, that we're beginning to learn about the puppet and the strings, but we can't say anything at all about the puppeteer, so how you select what you're going to do remains the kind of question that you can't even pose intelligibly in the sciences at this stage here, as well. Well, the eye language, keeping to the tradition, is a property of the individual and also the species specific faculty of language, also an internal property, something which allows the eye language to be acquired and it has to meet a couple of empirical conditions, two conditions which are kind of conflicting, the conditions of learnability and the conditions of evolvability,. So whatever the faculty of language is, it's got to be rich enough so that possessing it, a child can acquire the eye language from the scattered and limited data available. And it is scattered and limited. and it has to achieve the internal system, which has all of these rich and complex consequences,. so it has to be that rich. But it also has to be simple enough so that it could have evolved. And now we can be a little more specific about that, because some of the conditions of evolution of language are coming to light and talk about it later. if there's time and the evolution has to meet those empirical conditions. Well, those are the conditions for a genuine explanation,. If some proposed descriptive device satisfies these conditions, then it's the basis for an explanation for addressing the Yellow Land Challenge as it was formulated and developed. In the tradition of rational and universal grammar, the general explanation is always at the level of UG, the theory of the faculty of language, and it has to offer some prospects of satisfying the conditions of learnability and evolvability,. that's a pretty austere requirement, very austere requirement, but it's the right requirement. Anything short of that is short of actually explaining things, it's maybe very valuable, maybe organizing problems in an interesting way and move on from there, but still falls short of general explanation. We Can now, I think, grasp somewhat more clearly what actually is a genuine explanation, something that was really not possible in earlier stages of linguistic inquiry, but again, any device that's introduced to account for something unless it can meet these joint, these dual conditions is short of explanation, maybe very valuable. So Many examples, take a concrete example to illustrate about something I'll come back to later if there's time,. an interesting paper by Djokovic, who you all know, on the coordinate structure and adjunct island constraints. What he points out, is that each of these constraints poses many problems, many mysteries,. but his paper is an effort to try to reduce the mysteries by reducing both constraints to the same constraint, using the device of neo-Davidsonian event semantics, which interprets a junction as a kind of coordination. So You can reduce both of the problems to the same problem of coordination, and then we still have the mysteries,. but now a simpler problem, one set of mysteries instead of two independent ones, tries to show that the problems then reduce this way. Well That's a step forward, it leaves the mysteries in a better position for productive inquiry, but it's not an explanation, he's quite clear about that. And I Think if you look over the field that virtually every achievement, everyone, is a partial step forward in this respect. There's very few exceptions, just barely coming to Light, which I think can count as genuine explanations,. They're important in themselves, and they're also kind of a sort of a guideline into how we should think about proceeding, and they may also tell us something about just how far it's possible to go. It's not so obvious, you can go much beyond what kinds of explanations that are now beginning to come to light,. I'll talk about that. Well, actually the earliest work in generative grammar, tried to meet even more austere conditions. It was heavily influenced by a work of people like Nelson Goodman and W.E.V. Quine, who were working on what they called constructive nominalism. No sets, very austere, just a mere illogical concept of a very limited kind. That was too austere, at least for the present, couldn't get very far that way,. there were several papers about it. So That was kind of dropped, at least for the present, maybe even come back to it someday, and the tension turned to something else, namely the vast range of empirical data from all kinds of languages that was beginning to appear as soon as the first efforts were made to write actual generative grammars. It Turned out that everything was puzzling and complex,. nothing was understood, it was just massive puzzles. Big Change From a few years earlier, during the period of structural linguistics, it was basically assumed that everything was known, everything was solved, we had the methods of analysis, you could formalize them, all that was needed was to just apply them to one or another language. That turned out to be radically false. Well, the first proposals, as you all know, were dual,. There were operations to deal with the problem of compositionality, very structured grammar, and totally different operations to deal with the phenomenon of dislocation, ubiquitous phenomenon, transformational grammar. Both Systems were far too complex to meet the long-term goals of genuine explanation, that was well understood. The General assumption at the time remaining for a long time, often broken up until today, is that the principles of compositionality are natural, you can expect those, something like very structured grammar, But the dislocation is a weird property that languages have, a kind of imperfection that we have to somehow, languages for some reason have this,. formal languages would never be constructed with that property. And That is still a widely held view, I Think it's exactly the opposite of the truth, the opposite I Think turns out to be true, that more recent work suggests that dislocation is kind of the null hypothesis, it's what's expected on the simplest grounds, and it's the most primitive of operations, I'll come back to that. But Let me just take a brief look at the steps that were taken to reach what I think is this conclusion. Well In the 60s, phrase structure grammars were basically eliminated. A Phrase structure grammar is far too rich to be contemplated as relevant to describing languages, so there's nothing in the theory of phrase structure grammar that prevents you, say, from having a rule, you know, VP arrow NCP, let's say, fine phrase structure rule, doesn't make any sense. It was just assumed, you just can't do that sort of thing. But The right theory has to rule that out as unacceptable. And That step was taken by the late 60s, basically led to X Bar Theory. X Bar theory had interesting consequences, which weren't really fully appreciated at the time, they're obvious in retrospect., For One thing, X Bar theory, notice, has no linear order. So Japanese and English, roughly mirror images, have about the same X Bar theory, linear orders on the side somewhere. That was a step towards something which I think is now much clearer, namely that the surface order of expressions is not strictly speaking part of language. It's something else. We'll come back to that. But If you just look at X Bar Theory, it's already a step in that direction. Another Thing about X Bar theory is it forces a theory of parameters. So Japanese and English, say, differ, and they're going to differ in some choice that is not determined by X Bar theory. So Some, the speaker and the hearer, who's using a linear system of externalization, you don't have to use that. But If you are using it, you're going to have to make a choice as to the order in which you're going to externalize the internal system. So X Bar theory itself, first, is a step towards separating a linear order and other surface organization from what we might think of as core I language, the I language that's dealing with the Galilean challenge, constructing the set of linguistically articulated thoughts, putting externalization in some medium to the side. And I Think that picture is becoming clearer. We'll come back to that. Well, There are also, along with the clear progress of X bar Theory, there were very serious problems which weren't recognized at the time. The Main problem is it excludes the possibility of exocentric constructions. Everything has to be endocentric in X bar theory. And That's just false. There are exocentric constructions all over the place, simple things like subject predicate, or for that matter, every case of dislocation, without exception. All of these give you exocentric constructions. There's no way to describe them in X bar theory. Now In order to describe them, many artifices were developed. So For example, if you have a subject predicate construction, maybe it was called a TP or an IP or something like that, or a VP. But That's just stipulation. You could just as well call it an NP. And This runs all the way through the descriptive apparatus. So There was a serious problem not really recognized until a couple of years ago. My Own feeling is it's pretty much overcome by labeling theory, which tells you in a principled way in terms of minimal search, a simple computational principle, when movement, internal merge, may take place, when it must take place, when it need not take place. There are many interesting results and plenty of interesting problems about this, a lot of very intriguing material, most of which I presume you're familiar with. Well, by the moving up to the 1990s, it did seem to a number of us that it's enough had been learned, so it might be possible for the first time to confront the problem of genuine explanation. That's what's called the minimalist program. Pursuing That program. If you want the, if you want a genuine explanation, you want to start with computational operations which meet the conditions of learnability and evolvability. Well, the easiest way to meet the condition of learnability is to say that learnability is zero. It's just innate, nothing to say about it. And The easiest way to meet the condition of evolvability would be to say, let's find a computational principle that had to evolve. There was no way for it not to have evolved. Well, if you look at those two conditions, they're satisfied by the most elementary computational operation, what's been called merge in recent years, which incidentally has many problems that I'll come back to. But Basically just the operation of a binary set formation. It has to be there because the basic property exists. And That means at least, at the very least, the simplest operation must exist, maybe more complex ones, but at least the simplest one. So We know that it has to exist, had to evolve, so it meets the condition of evolvability. That leaves the question of just how it happened and what the neurological implication is. But Whatever the answers to those, this is an operation that had to evolve. And Having evolved, it's innate, so it meets the condition of learnability. So If you can reduce something to that, you do have a genuine explanation. That's as far as it's possible to go. If It doesn't,. if you can't go that far, it's a description. It's not a genuine explanation. Again, this is a pretty austere requirement, but I Think it's the one we ought to have in mind when we're thinking about the goals of our efforts in inquiring into language. Well, So let's, I won't give the details because I think you're familiar with them, but the simplest computational operation, then merge binary set formation, meeting the no tampering condition, least possible computation. You Don't modify the elements, don't add any more structure. Interesting Things to say about this, which I'll come back. There is a good deal of current literature which tries to show that you can reach this operation in steps. That's incoherent. You can't have partial binary set formation. You can't reach it in steps. You Either have it or you don't have it. There's nothing simpler. Again, lots of literature about this, but it's just beside the point. There's actually a recent, interesting recent paper by Rene Heubrichs analyzing some of the recent proposals and showing why they don't make any sense. But If you think about it, they can't make sense. The Simplest case of merge is going to have at least, maybe at most we would like to show, but at least two cases. One of them, external merge when you're taking separate things and forming the set. One Internal merge when you're taking one thing and something inside it, forming the set of those. Those are at least the two simplest possibilities. Notice There are only one operation. There's no two operations, just one operation with two cases. Much Confusion about this in the literature, but that should be obvious if you think it through. Well, notice that this whole program is a program. It's not a theory. The Program is to see how far can we go if we take the simplest possible operation and try to give genuine explanations in terms of it. Maybe That's impossible. Maybe You have to find more complex operations. But In that case, it's going to be necessary to demonstrate how they can be acquired, how they can be learned, and how they could have evolved. And That's not so trivial. You Can't just say, well, natural selection does anything I like. That's not an explanation. You Have to give a real explanation. Very difficult in biology. In The biological literature, it's pointed out that it's a fiendishly difficult standard phrase to give an account of the evolution of almost any trait, even the simplest ones, like having blue eyes, for example., And It's not the kind of thing you can hand wave about. So Either you can try to meet that condition or recognize that you don't have genuine explanations. Well, there have been, I think, substantial achievements in the last recent years in trying to gain general, genuine explanations. They do have problems. I Want to return to the problems later, but I'll put them on the shelf for a moment. The One achievement, which is not trivial, is to unify the two traditional kinds of operations, compositionality and dislocation. They are unified once you keep to the simplest computational operation. So Far from being an imperfection, as was always assumed by me in particular, it would take a stipulation to bar dislocation. If You have no stipulations at all,. you get dislocation. Furthermore, As I mentioned before, that's arguably the simplest case of merge. Actually, you can't have only one and not the other, because once you have merge, you have both. But If you're looking for one that's more primitive, it's probably internal merge. The Reasons for that are quite straightforward. External Merge requires enormous search. To Put two things together that are separate,. First of all, you have to search the entire lexicon. Then You have to search everything that's already been constructed and maybe is sitting there somewhere waiting to be merged. With Internal merge, you have almost no search at all. So One reason for regarding internal merge dislocation is more primitive. It requires a tiny fraction of the search. But There's a good deal more than that. There's some interesting suggestions in the literature. They're not definitive, but they're suggestive. So One was some work that was done by Marv Minsky a couple decades ago. He and one of his students just explored what would happen if he took the simplest touring machines, smallest number of states, smallest number of symbols, and just let them run free and see what happens. What Turned out was kind of interesting. Most of them crashed, either got into infinite loops or just stopped. But The ones that didn't crash, all of them gave the successor function. Now, what's the successor function? Well, One thing the successor function is, is internal merge. So If you take merge and you have a one member lexicon, just run three, you get the successor function. Minsky's argument at the time was that probably evolution,. in the course of evolution, nature found the simplest thing. That's what you'd expect. So It found the successor function. And That happens to be internal merge, not external merge. If You look at other organisms, a way down to the level of insects,. they have, they count. So An ant, say, can count the number of steps it's taken. It's got a counter, maybe a set of counters inside. And If you look at just the mathematics of successive counters, they kind of tend towards the successor function. It doesn't take a big step to move them up to the successor function. So From various points of view, it seems plausible to think that of the core operations, the most primitive one is actually dislocation, contrary to what was always thought. And As you get richer constructions, you have external merge and it gives you richer kinds of languages. We Plainly have it in natural language. It's not just internal merge. An Interesting question is why. It probably has to do with argument structure, which is uniquely related to external merge. We'll come back to that. Well, What's with the unification of internal and external merge, compositionality and dislocation,? what was suggested by X-bar theory, as I mentioned before, becomes much more clear and explicit. So It seems that the generation of the CI interface, sometimes called LF, what gets thematically interpreted, linguistically articulated thoughts, that's, we can call, core I language. And That just keeps the structure. No Linear order, no other kinds of arrangements. So Why is there linear order in spoken language? Incidentally, not strictly in sign language. So in sign language, which we know to be essentially equivalent to spoken language, there's different dimensionality. So You can use visual space. You can use simultaneous operations, the facial gestures and motions. So It's not strictly linear. It makes use of the contingencies allowed by the space that's of externalization. But Speech happens to be linear. You have to string words one after another. So If you pick that particular modality of externalization, yes, you're going to have linear order. But Does linear order have anything to do with language? Well, you know, depends what you think you want to call language. But What it really has to do with is an amalgam of two totally different independent systems. One of them, internal language. The Other, a particularly sensorimotor system, which has absolutely nothing to do with language. The Sensorimotor systems were around the hundreds of thousands, maybe millions of years before language ever appeared. They Don't seem to have been affected by language. At Most, there's very minor suggestions about slight adaptations that might have taken place for, say, changes of the alveolar ridge and clique languages. Very Small things. But Basically, the sensorimotor systems seem independent of language. But If you do externalize the internal system through this filter, you're going to get linear order. But Strictly speaking, that's a property of an amalgam of two independent systems. And In fact, that's true of externalization altogether. And Notice that externalization opposes a hard problem. You have two completely independent systems. They have nothing to do with one another. You have to match them somehow. You can expect that process to be pretty complex and also to be variable. You can do it in many different ways. Also, to be easily mutable, can change from one generation to another under slight effects. Putting All these expectations together, what is a natural expectation? And I Think it increasingly is coming to be imaginable, Maybe true, is that the variety and complexity and mutability of language is basically a property of externalization, not a property of language itself. And It could turn out to be true. It's a goal at the moment that the core I language is really unique, may not vary from language to language. Actually, that much is pretty much tacitly assumed in essentially all the work on formal semantics and pragmatics. It's not assumed to be parameterized from one language to another, or to be learned somehow. It's just there, which means if we ever understand it properly, it will be reducible to elementary computations, which just don't vary. That's the way the internal system works. That should be the goal of inquiry in those directions. I should say, just as a terminological point, what's called formal semantics is actually a form of syntax. It's symbolic manipulation. Technically, something becomes semantics when you relate it to the external world. And That's a tricky business. Even Things like, say, event calculus,. if you think about it, events are really mental constructions. You can't find them in the outside world. You construct them there. And The task of relating what's internal to the external world, dealing with questions of reference, is no trivial matter. A Lot to say about this, but I'll put it aside. But It seems to me we can see a goal for all of this work to try to reduce it to computational operations that do meet the conditions of genuine explanation. Again, a very austere criterion, but I think one that's worth keeping in mind. Well, these are all possibilities that I think are looking increasingly plausibly and the field may go in that direction. It'd be very striking discovery if it really does. Well, let's go on with genuine explanations. One of them is dislocation, putting it together with compositionality. And Notice that that includes automatically the basis for what's called reconstruction. You Keep to the no tampering condition. You Automatically get what's called the copy theory of movement. That's the basis for the complex properties of reconstruction. There's a lot to look into, but that's essentially the basis for it. You Don't need rules of reconstruction. They're just there. That's automatic. Well, of genuine explanations, the most interesting case, I think, is the old principle of structured dependence. This was discovered back in the 1950s. This is a really strange principle of language, which had never been noticed, namely that the rules and operations of language, the ones that yield interpretation of sentences, don't pay any attention to linear order. They Just deal with structures, which is extremely puzzling when you think about it because linear order is what you hear. It's 100% of what you hear. You Never hear structure. Furthermore, at least superficially, it seems that computations on linear order are simpler than computations on structure. From Another point of view, that turns out to be false, but at least superficially that looks right. So What it seems, and what always seemed extremely puzzling, is that the syntactic rules and the rules that yield semantic interpretations don't pay any attention to 100% of what you hear and to the simplest operations, which is a pretty puzzling fact. We Now have a simple explanation for it. It follows from the simplest computational operation. If The entire internal language is based on the computation of the simplest merge operation in its simplest form,. Then you automatically get structure dependence for operations of movement of construal of interpretation of everything else. I Won't run through examples. I Assume you're familiar with them, but that just seems to be a fact about all constructions and all languages. That, if it's correct, is a genuine explanation of a fundamental property of language, maybe the deepest property of language, that the core language just doesn't care about order and arrangement. It only cares about structure. And A child learning language just ignores everything they hear. By Now, there's interesting independent evidence supporting this conclusion. So For studies of language acquisition, which have proceeded in very sophisticated ways by now, have now gotten down to the point where 30-month-old infants have been shown already to observe the principle of structure dependence. That's almost no data, remember, and it's a very abstract principle. There's other work, earlier work by Steve Crane, Nakamura, who's got a lot of evidence. The Three-year-olds have mastered it. Recent Studies have it down to 30 months. If We have better studies, which, as they keep improving, it'll probably be earlier. What that means is you're just born with it. So It meets the condition of learnability, namely zero, and it has the condition of evolvability. You Have to have this particular operation at least, maybe more, but at least this one, because you do have the basic principle. Well, there's also, as many of you know, neuro-linguistic evidence. The studies of, inspired by Andrea Moro of a group in Milan, Musso and others, have shown, many of you know this, that if you present subjects with invented systems of two types, one which correspond to the rules of an actual language that the subjects don't know, the other, which uses things like linear order, you get different kinds of brain activity. In The case of, say, having a negation be the third word in the sentence, a very trivial operation, you get diffuse brain activity. If You follow what look like more complex rules of actual languages, you get activity in the expected language-specific areas, the brain, Broca's area, and so on. That's been, by now, replicated many times. It looks like a pretty solid result. There's also psycholinguistic evidence of other kinds. The Moro-Musso Experiments were actually suggested by work of Neil Smith and E. Anthony Timpley on a subject who they've been working with for many years, a young man they call Chris, who has extremely limited cognitive capacities, almost none, but tremendous linguistic capacities. He picks up languages like Ken Hale, like a sponge, in other words, just picks them up immediately. Neil Smith Tried these same experiments before the neuro-linguistic ones were done. He Just tried it with Chris and turned out when he gave Chris a nonsense language modeled on an actual language, he learned it easily, like every other language. When They gave him the very simple language, things like negation being the third word, he couldn't handle it at all. It was just a puzzle. He Can't deal with puzzles. That's what inspired the neuro-linguistic studies. I Think those are the most interesting discoveries so far in the brain sciences related to language. It's a direction in which other experimental work had gone. Looking Back at this, it seems to be one of these very rare cases where you have converging evidence from every direction, leading to the same conclusion that poor I language just is independent of linear order and other arrangements. You Have linguistic evidence, psycholinguistic evidence, neuro-linguistic evidence, evolutionary considerations, anything you can think about. Now There's a very curious fact. There's a huge literature in computational cognitive science trying to show that somehow this principle can be learned, which is a very weird fact if you look at it. It's like trying to find a complicated way to disprove the null hypothesis. Things Like that just don't happen in the sciences. I Mean, here you have the absolute optimal explanation and a huge literature trying to show, look, there's a very complicated way in which maybe we can reach the same conclusion. It's an enterprise that's kind of senseless at the base of it. Of Course, when you look at the actual cases, it never works. It's not going to work. If It did work, it would be meaningless because it's always asking the wrong question. I Mean, suppose you could show that by a detailed statistical analysis with recurrent neural networks and so on of many layers of, say, the Wall Street Journal, you could find evidence that a child might have used at 30 months old to discover that you have structure dependence. You're not going to find that, of course, even though there's literature claiming it. But If you did find it, it would be completely meaningless. Of Course, the only question is, why is this the case? Why Is it that in every language and every construction, this is the way it works? If You could find a way of showing, well, here's how it might work in this language, tells you nothing. It's answering the wrong question. And Furthermore, as I say, it's trying to find a complicated way to disprove the null hypothesis. The Whole Enterprise is completely senseless. It's actually probably the major effort in computational Cognitive science to try to find a basis for some linguistic principle, huge literature on it, new papers still coming out. A Very strange thing, papers trying to show that, as they put it often, you can get structure dependence without what's sometimes called an inductive bias for structure dependence. But There's no inductive bias. It's just the null hypothesis. Make No assumptions. This is what you get. There's no bias. It's just given. So I Think an interesting question about, many interesting questions about how linguistics is done. But One of them is why things like this go on. I Think it's worth thinking about. Well, there are other successes, but what I'd like to do is turn to problems. There are a lot of problems about merge, and there are some paths to solution. So One problem is what I already mentioned, exocetric constructions. So It takes a NPVP. Let's assume, since Dominique is here, let's in his honor, assume the predicate internal subject hypothesis. So you put together a subject and an NP and a VP. The NP's are often called DP's. I'll come back to that. I Think it's probably a mistake. Let's just call them noun phrases for the moment. You have a noun phrase and a verb phrase. You Put them together. That gives you the basic Theta structure. Well, the noun phrase and the verb phrase have to be independently constructed, which means you have to have some kind of workspace, something that Jonathan pointed out years ago. You Have to have some kind of workspace in which you're constructing these separate things. And If you think it through, the workspace can proliferate, not indefinitely, but can get larger, where you're just doing parallel things and putting them together. So It means that the operation merge really ought to be revised to become an operation on workspaces, not on two elements, X and Y. It's an operation which changes a workspace to another workspace. And Then the question comes, how it does it. Well, I should say I'm very pleased to be back at a nice low-tech institution like MIT with blackboards and no PowerPoint and no projections and any of that stuff, which they have in Arizona. So What we want is some kind of operation that says it's called a capital merge. We'll look at its properties, which takes two things, call them P and Q. The Guys were going to merge on a workspace and turns it into some other workspace. So What's the other workspace? Well, it's going to include the set PQ, the two guys were putting together. In Fact, let me use a different notation for reasons all mentioned. A Workspace is a set, but we want to distinguish it from the syntactic objects, which are sets. So A workspace doesn't merge with something. So Just for convenience, just use a different notation. So The new thing will include the set PQ and a lot of other junk. And The next question is what's the other junk in the workspace? That Turns out to be not a trivial question. A Lot turns on what the answer is. So Let's take the simplest case. The Entire workspace consists of two elements, a column A and B. That's the workspace. And Suppose we decide to merge them. So We get the new workspace, which includes the set AB. And Does it include anything else? So For example, does it include A and B? Well, If we think about the way recursion generally works, it should include A and B. So If you're doing, say, proof theory, you're generating a proof. You Construct a line from axioms and former things. And You can go back to that line, if you like. You Can always go back to anything you've produced already for the next step. But There's a good reason to believe that for organisms, and particularly humans, it doesn't work that way. And You can see that if you think what would happen if you did allow this to happen, suppose you allow this, then you could go on to construct some much bigger object here, including AB as a term. But It could be of arbitrary complexity, any kind of complexity you like. And Then you could take A and merge it with it and get xA. But Then A would be up here. And A would be down there. And There are two copies. And They would violate every imaginable constraint on movement. So If you allow this, you're going to get total chaos. Every constraint on dislocation will be violated. No matter how radical you make the violation. Well, that tells us something. It tells us something kind of surprising, and I Think significant, that the kind of recursion that takes place in human language, and probably organic systems generally, cuts back the number, the set of items accessible to computation as narrowly as possible. Let's give it a name and call it resource restriction, RR for simplicity. It Looks as though a very general, this is merely the first example. If You think it through, it works for millions of things. The Same model of refutation eliminates a whole set of possible extensions of merge that have been proposed over the years. I'll come back to examples. But You notice what the problem is. The Problem is that if you allow the normal kind of recursion, no constraints, no limits, then you're going to find that by legitimate means, you can get illegitimate objects. Now That has to be barred. You can generate all kind of divin expressions. That's not a problem. But You don't want to have legitimate means for generating things that violate every possible condition, descriptive condition. Anything Like that is wrong. Well, in this case, and it turns out in a great many cases, you can bar this outcome simply by limiting the resources that are available. Now, what are the resources? The Resources are elements that are accessible to the operations. So The real condition says limit accessibility. Keep Accessibility As small as you can. We Already have examples like that that we're familiar with. One of them is the phrase impenetrability condition. If You think about what that condition says, basically it says when you're generating something, you get to a certain unit, a phase, talk about what it is. Anything Inside the phase is no longer going to be accessible to operations. That reduces the amount of computational search that's required, but it's a way of limiting accessibility. It says those things down there aren't accessible anymore. Another example, and this may be the only other example, is minimal search. This is what's often called a third factor property. Third factor, for those of you who are not familiar, comes from the just simple description of the elements that enter into computation, into learning. So What enters into acquiring a system is three things, external data, internal structure, and basically laws of nature, which are independent of the system in question. So If you're studying growth of arms, let's say, humans grow arms, not wings, partly because of nutrition to the embryo, and partly, in fact, largely because of internal structure, just genetic determination, and extensively, simply because of the way physical laws operate. There's only certain ways that organisms can develop, other ways just aren't possible. You Put these together, you account for any kind of growth and development. The Same is true of language. There's external data, whatever it is, it's going to determine whether you end up with the Golag or English. Internal structure, which at least includes merge, more, no doubt, but at least that. And In fact, anything that can be explained in terms of that, it does yield a genuine explanation. And Then laws of nature. What are the laws of nature? Well, language is a computational system, kind of unusual fact. That's rarer and organic nature, maybe unique even, aside from counters. But Anyway, that's what language is. So Among the laws of nature that you would expect would be things like elimination of computational complexity, doing things as simply as possible. Several Reasons for that. One of them actually goes back to Galileo again. One of Galileo's precepts was that nature is simple and it's the task of the scientist to prove it, whether it's falling objects or flight of birds or growth of flowers or whatever. That's a kind of a prescriptive hypothesis. You Can't prove it, but it's been extraordinarily successful. In Fact, the whole success of the sciences in the last 500 years is based on that. And That's a good enough reason to assume that it works for us, too. So Reasonable to accept that. There's a general point, which just has to do with the nature of explanation. It's just a fact about explanation that the simpler the assumptions, the deeper the explanation. That's just kind of logic. So There's a lot of, I should say in the case of language, there's another reason to believe it which is unique to language. And it has to do with the conditions on evolution of language. We Don't know,. very little is known about evolution altogether. As I Said, to try to really account for the development of any particular trait is very hard, even in simple cases. In Sort of the evolutionary psychology literature, everything looks easy, happened by natural selection. Why Not? But When you really try to explain something, it turns out to be hard. In The case of cognitive development, it's uniquely hard because you have no fossil records, no tape recordings of people doing whatever they were doing 100,000 years ago. Furthermore, when you deal with language in particular, it's super hard. With Other organic systems, say vision, you have comparative evidence. You Can study cats and monkeys which have essentially the same visual system. And With them, we rightly or wrongly allow ourselves to do invasive experiments. So You can stick a neuron into one cell in the striate cortex and see what's happening and so on. And You learn a lot from that. That's how we know about human vision. But In language, you can't do it because there's no other analogous system. It's a unique system, nothing analogous in the organic world, so there's nothing to test. So It's kind of uniquely hard. And Nevertheless, there's some evidence. The Evidence at Bob Berwick and I have a book reviewing it. By Now there's better evidence than what we had in the book. There's by now genomic evidence that Homo Sapiens began to separate roughly 200,000 years ago. That's when you get the separation of the Sun people in Africa from the rest. Interestingly They have unique forms of externalization. These turn out to be essentially all and only the languages that have complex click systems. There are what apparently look like a few suggestions, exceptions, but they seem to be borrowings or something accidental. There's a kind of very interesting paper by Rini Huyper on this recently. So One thing we know pretty convincingly is that roughly around 200,000 years ago, humans began to separate. They shared the faculty of language at the time. So There's no known difference between the faculty of language of the Sun people and everybody else. Nobody knows any differences, group differences in language capacity. There happens to be a different form of externalization, which suggests, and in fact Rini goes into this in detail in his article, that this particular forms of externalization develop later. As A matter of logic, the internal system had to be there before you can externalize it. That's not debatable, but he suggests there's a gap when the system was there roughly 200,000 years ago and began to be externalized in somewhat different ways later on. When Did Homo Sapiens appear? Well Here, we have reasonably good fossil record which shows that anatomically modern humans emerge appear roughly at that time, maybe 250,000 years ago, which is essentially nothing in evolutionary time. So It looks as though the language emerged pretty much along with Homo Sapiens, with the faculty of language intact. Another Kind of evidence comes from the archaeological record, which gives a lot of information about rich symbolic activity. It Turns out that almost the entire rich symbolic activity that anybody's dug up so far is after the appearance of Homo Sapiens. Well rich symbolic activity has been naturally taken to be an indication of the existence of language. Also more complex social structures, you know, burial practices, all sorts of stuff. So Putting it all together, it looks plausible that language emerged suddenly in evolutionary time. Along With Homo Sapiens, some whatever change gave rise to Homo Sapiens seems to have brought language along with it, and it apparently hasn't changed since. That's independent reasons to believe that whatever is in there is probably very simple, along with the Galilean precept and the general principle. If you want explanation, you want simplicity. So It makes good sense from many points of view to assume that the relevant laws of nature here are avoiding computational complexity, computational efficiency. That's what It means to call it a third factor property. Well One particular case of computational simplicity. Hi, how are you doing? What Are you doing here? You Graduated years ago. Go Back to Tufts. I'm wondering where you put among the three factors: the fact that the computation has to run on neurons. That's third fact. First Of all, it's not necessarily the case. That's a myth, you know. There's a myth that neural nets are what do the computation, but there's pretty good evidence that that's not true. The Neural nets just don't have the computational principle. I'm talking about real neurons. Real Neurons. I'm talking about real neurons. Real neurons may not be the elements that enter into computation. There's by now reasonably strong evidence against that. Randy Galastol's book with William King is a very good case where he gives strong evidence, Randy, that if you look at neural nets, you simply cannot find the basic elements of essentially Turing machines. You can't find the core kind of computational element that yields computational activity. It's just not there in neural nets. So What he's arguing is that people who've been looking for neural net accounts of computation are like the traditional blind guy who's looking under the lamppost for his lost keys because even though he lost them across the street because that's where the light is. So Yes, we know something about neural nets, but happens to what we're looking for somewhere else. There's a lot more evidence. The Speed and scale of computation is way beyond what neural nets are capable of. And by now there's including. Randy's particular proposal is that the computation is actually down at the molecular level, having to do with RNA and so on. There are other proposals by not inconsiderable people like Roger Penrose For example, that the computation is being done by structures that are internal to neurons which have vastly greater computational capacity. There's chemical processes that go on in the brain that aren't captured by neural nets. It's been known as far back as Helmholtz. That speed of transition of neurons is just kind of too slow to be doing very much. So We're going to have to look elsewhere to find the implementation of computational systems. There's something there, and it's going to be a third factor property, something about the brain, clearly. We Talk about this in our book. So Yes, there's surely we're going to try to relate whatever is going on ultimately down to the level of cells. That's science, try to reduce everything. So That's all third factor if you like. But Just talking about neural nets is kind of like talking about natural selection. I Thought you brought up neurons. Something In the brain, yeah. Surely Something in the brain is responsible for this. not the foot, let's say. You Can amputate your leg, you still have language, you have to take your head, you don't. So Yeah, we agree. there's something going on in there. It's a tricky question. It's a very hard question, even for simple traits, not just language. Very Simple traits. As I Said, you look in Technical Studies of Evolution, the phrase that's often used is fiendishly difficult to find the evolutionary basis for even the simplest traits. So To think we're going to suddenly find them for language is a little misleading. There are some interesting ideas. So Angela Ferrichi's book that came out recently, MIT Press book on the state-of-the-art and neural linguistics, gives interesting suggestions about what might be involved in the probably small change in brain rewiring that led to the appearance of merge or something like it. You Know, closing a certain circuit in the dorsal and ventral connections. It's an interesting proposal, but certainly not going to be trivial. You Know, it's a hard problem. Yes, but yes, that would be third factor. So Now where was I? He's always been very disruptive ever since he was a student. Well, maybe I'll just kind of end here and go on next time. But One principle that we're going to expect for many reasons is computational efficiency. Minimal search is the strongest form of computational efficiency. Search As little as possible. And There is a case of restricting accessibility that we're familiar with, which reduces to minimal search. That's the case of successive cyclic movement. So Suppose you've taken a Wh phrase and you've moved it up to here and then you've moved it up to here and you keep going. And Suppose both of these, and neither of these, let's say, is blocked by the phase impenetrability condition. Suppose That only blocks things here. Well, the next thing that's raised is this one, not that one. We Just take that for granted. Nobody talks about it. But If you ask why, it's again a minimal search question. Whatever's selecting the operation, argue about what it is, that goes back to that mystery I mentioned is going to take this guy because that's the one that's going to find by minimal search. So We have at least two cases already that we're familiar with of limiting accessibility. One PIC, which is pretty broad, the other minimal search, which we've just taken for granted. And Maybe that exhausts it. But Now I think there's a broader general principle which says just restrict resources. That will have a lot of effects. I'll come back to more examples of that next time. There's a temptation at this point to relate, restrict resources to something that Ray and I were just talking about. The Fact that the brain is just slow, doesn't work fast, works quite slowly. And There's many domains in which that is the case. So In many ways, the most striking one is vision. If You look at the sensory motor system, the visual system, The cells of the retina are actually responsive to single photons of light. They're maximally, they give you a maximal amount of information. The Brain doesn't want all that information. It's just the way overloaded if it ever got that kind of information inside. So Whatever the visual system is doing is, the first step it's doing is throwing out almost all the information that's coming from the retina. And Apparently every sensory motor system, every sensory system, is like that. The First thing it does is throw away just about everything and try to get down to something limited enough so this slow brain up here can deal with it somehow. That looks very much like a general property of which Resource limitation of the kind That says, don't do ordinary recursion, but restrict the resources of which this is some special case. All seems to converge kind of plausibly. We're very familiar with this in the study of language acquisition. So As you all know, an infant acquiring a phonetic system is basically throwing away information. It's throwing away tons of information in the first months of life, and maybe in about nine months or a year, saying these are the only things I'm going to pay attention to. The Same thing goes on through the rest of language acquisition. If You look at something like Charles Yang's general approach to language acquisition where you're just shifting, you start,. the child starts with all possible grammars by languages. And Then it changes the probability distribution of them as data comes along, reducing the probability of things that don't have evidence for them so that they become fundamentally invisible. It's also a matter of throwing away lots of information and converging on just a little bit. The Development of the brain is constantly losing neurons because you don't want all this junk around. You Want just what you need. And The resource limitation fits pretty naturally into that system. I Think I'll stop here and try to come back to more detail examples next time. Unless Somebody else wants to disrupt. |
| 4 | ASR-whisper-rpunctuated_Noam Chomsky, Fundam_1669853631_0_part2.txt | aa279e3b-2d1 | ASR_cleaned | Well, last time I Talked about a number of things and got up to the point of beginning to discuss the problems that exist with the concept merge that was developed back in the 90s and has been used in many ways since. There's a kind of a simplest version of merge, which was the original intention, which just had the two special cases, external and internal merge. As I mentioned last time, the more primitive of the two is actually internal merge. But Because of the fact that language has exocentric constructions that can't suffice, I mentioned some of the things that you can explain on the basis of merge, and also wanted to make the point that a genuine explanation in linguistics will, if we're viewing the study of language as part of the study of nature, basically the bio-linguistics program, which I think has roots back to the 17th century, as I mentioned last time,. although you have to skip the structuralist, haveless period, if we're engaged in that enterprise, then a genuine explanation will always have to meet these austere conditions of learnability and evolvability, which are very hard to meet anywhere in biology and in particular here. There's some reason to think that they might be, conditions might be attainable here because of the specific conditions of human evolution, which I mentioned briefly last time. If That picture's correct,. there's some antecedent reason to believe that there might be success in the enterprise, which is rare in the biological sciences. Well, the concept merge does happen to meet those conditions. It meets the condition of learnability because there's nothing to learn. It meets the condition of evolvability because since, in fact, the basic problem, the basic principle does exist, there had to be something to evolve the computational procedures that yielded it. And It would obviously be at least the simplest one. So We can be secure with explanations based on the concept merge. But Anything else is problematic. It's a very austere condition, but it's one that really has to be met. Well, I then started in on, talked about some of the examples where you can get an explanation, some interesting cases. But There are problems. The Problems are that the concept was very loosely defined and all sorts of other applications, the implementations have been given, which kind of more or less fall within the original loose definitions, but I think are probably illegitimate. I'll talk about that today. And I Think If we think through the matter carefully, we end up with just allowing what was originally intended and none of the extensions for good reasons, which leaves us with many problems, some of which have, I think, potential solutions, others look quite mysterious. Well, I mentioned last time, something which is a kind of a paradigm for many cases, the simplest case when you have only two elements. First Of all, since we do have exocentric constructions, it's going to be necessary for the operation merge to actually operate on a workspace, not on elements, because you're always changing the workspace every time you apply merge. So We have some sort of a definition that we call a capital merge, which says take two things that you want to merge and a workspace that exists and form a new workspace, which will include at least this, and then other things you don't want. And I Suggested a notation, which I sometimes will forget. Sam Fixed it last time. I will use square brackets for the workspace and curly brackets for the syntactic objects. There is a crucial difference between them. The Workspace is a set, but it's not an accessible object for operations. So We'll just distinguish them by that notation. And The notation actually means something. So For example, if suppose the workspace consists of just x, we want to distinguish that from X. So The singleton set is different from its member, because the workspace is not a syntactic object, and x is. On The other hand, for the syntactic objects, we want the opposite convention, namely that the singleton set is the individual element. There are good empirical reasons for this, which go back to Phrase Structure grammar. So In Phrase Structure grammar, you just, by convention, didn't allow rules like, say, Np arrow Np or E arrow V. That's assumed not to be a reasonable rule. That's essentially saying that a singleton set is identical with its member. Now, this was fudged often in the use of phrase Structure grammar. So There were allowed rules were allowed like this. When You move from phrase structure grammar, which is totally unacceptable for language for myriad reasons, as was recognized since the 50s,. When you move from that to x bar theory, then this becomes meaningless because there's no Vp. if it's only V. Actually, despite the fact that it's meaningless, it is used. So For example, if you try to implement Richie Kane's LCA, try to get it to work, you're forced to have rules like this, which is a serious problem in the LCA system, I think. You Have to argue that if you have a verb object structure, the object, even if it's a pronoun, still is complex. Otherwise, you don't get the right ordering. But That's technically illegitimate in an x bar theoretic structure. And Here, the analog of that illegitimacy is this convention. So We want to accept this convention, which has a number of consequences for syntactic objects and this convention for workspaces, which are different kinds of things, though they're all sets. Well, In the simplest case, we just have a workspace consisting of these two guys. And We merge them. Merge AB gives us a workspace which contains the set AB, which we've merged. And The question is, what else? Now, if it was normal recursion, like say, proof theoretic recursion, you would have here A and B. But You can't have it for language, for reasons which I mentioned last time. And This is a kind of a paradigm that applies to a great many cases of the extensions of merge. The Reason it doesn't work is that this can be built up to an object of arbitrary complexity. This, since it's accessible, can then merge to it. That gives you a relation between the thing that's merged up here and the thing down here, which violates every imaginable condition of movement. So That's illegitimate. And We do not want to have legitimate operations which yield illegitimate conclusions. That's elementary. So Therefore, we conclude that, surprisingly, if these things aren't here for language, recursion for language is different from general recursion. Namely, it has the property that I hold last time restricting resources. And What computation for language, and presumably for organisms generally, is doing, is trying to keep the resources as limited as possible. You Have to get something new, or you don't generate anything. But You want to generate as few things as possible. This really turns into a number of subcases. One Subcase is limiting accessibility. So Accessibility means something's accessible if the merge operation can see it and do something to it. We Want to limit accessibility. If We allowed general recursion, we'd have too much accessibility here. So We want to limit it to the minimal amount. It's tempting, as I mentioned last time, to try to relate this to a more general property of brain computation. Namely, the brain is pretty dumb and slow. So What it does is throw out tons of data that are coming in. In Fact, that's its main activity, is to get rid of lots of stuff that's coming in. So In the visual system, the sensory part of the visual system is essentially perfect. So You get a cell responding to a photon of light. Can't do better than that. But That's pouring into the brain tons of information that are going to totally overwhelm a computation. So The sensory system throws out almost everything, get down to just the limited part. Same True in language acquisition. The Phonetic system is throwing out just about all the noise that comes in, picking only very limited kinds of phonetic properties. And Even those are being thrown out very quickly in early language acquisition. That's the main part of language acquisition. Charles Yang's model for general language acquisition kind of exploits this generally. Notice, incidentally speaking, of Yang that if you have this property, you infer determinacy. It Turns out that if you think it through, when you limit the resources available, you're also going to force determinacy, meaning the operation will be uniquely determined by what it's looking at. That's not a trivial property. It wasn't true, for example, of standard versions of phrase structure grammar. So In a phrase structure grammar, if you reached a point where you had something of the form NP, the NP, and you have a rule expanding NP, which one you apply it to is indeterminate. It's a normal part of phrase structure grammar. But If you have in this much narrower system, resource restriction, you get determinacy. And That's kind of important, because Charles's work on the price of productivity are very important work, I think. That work depends, personally, on the assumption that the operations are determinate. So If you want to get those rich consequences, we want that property. It's one of the few examples I know of of the work in computational, statistical or computational linguistics that has real consequences, very rich consequences, very important work. So Resource restriction makes sense, as a property has a lot of interesting consequences. Lots follow from it. Now, resource restriction is going to have two components, sort of. One of them is restrict computation. The other is restrict resources. Restrict computation means limit yourself to the minimal kinds of computation that are possible. Well, merge is one case. It's the least possible computational operation. But It also wants to operate in the most limited way. So One of the consequences would be keep to minimal search. Don't use deep search if minimal search already works. Many Empirical consequences to that. Notice That limit accessibility already includes things like PIC, the phase impenetrability condition is one general condition that limits accessibility. Minimal search is another. It means, for example, in successive cyclic movement, when you move to the next stage, you don't look down. You only find the first thing that you find. That raises many interesting questions about possible ambiguity. Are There ambiguous cases? I'll come back to that. quite interesting question later on. But Those are the things that are in the background. Now, if you look at the original definition of merge back in 1995, it actually had this property inadvertently. It wasn't noticed, particularly. But The operation of merge, as it was defined, was defined basically as replace, which says, don't keep what you already had, but get rid of it. That was not particularly noticed. But It's a property of the original definition. And There are good reasons for it. We Can now see good reasons for it. If You don't accept it,. you do get legitimate operations, which yield illegitimate conclusions, the clearest sign that something's radically wrong. This is a paradigm case, but it extends to many others. Go into that. Well, So for example, take something that I presume no one's ever proposed. I'll draw trees. But Let me make, I should make it clear that trees are very misleading notations. One should be aware of them. For One thing, a tree notation suggests that this exists, that the root exists, but why does the root exist if you just have merge operations? In Fact, what the root is understood to be is something that comes from some other source, namely projectability. But That should be a completely separate property. Projection seems to have nothing to do with the compositional operations. The Tree notation kind of sticks them all together and misses many questions about what projectability is,. a particularly interesting case in exocentric constructions. So What's projected? Well, that has interesting consequences. Whole theory of labeling deals with that. I Assume you're familiar with it, or I'll put it aside. But We don't want that. The Other thing that tree notations allow you to do is draw lines in all kind of complicated ways. Draw A line from here to here, and that seems to mean something in a tree, but it means absolutely nothing in a merge-based system. And You really have to be very careful about that. So I'll use trees for exposition, but with a condition that you don't take them seriously. Well, let's take something that I presume nobody's ever suggested. Suppose That you have a structure like this, and you decide to merge these two. The Original definition doesn't say you can't. So That would mean you're forming the set xp, yp. But It has nothing to do with this original set, just something added on. And Notice that it has exactly the properties that are barred here. It adds accessibility. It's adding new accessible items, which will then be subject to exactly the problem I already mentioned, namely: this object can be made as complex as you like. You could then merge one of these guys to it, and you violate all conditions. So We're not allowed to do this. As Far as I know, nobody ever proposed this, but there's a good reason why you can't do it. However, There are things that people have proposed, and I think they're all ruled out for the same reason. I'll kind of leave it to you as an exercise to work out why, but if you think about things like parallel merge, which is usually written in trees like this. Again, the case where the tree notation seems to be saying something, but it's not because there's no way to construct this. But If you think about what parallel merge is doing, it's increasing accessibility runs into this very same problem. Parallel Merge is the basis for many, a lot of work in the literature, which yields multidimensionality. The Idea of multidimensionality goes back to the 70s, work by Jim McCauley and others. But If you try to reconstruct it in a merge-based system, you can draw the trees with funny lines like this. And The way of constructing it is through parallel merge, which has this lethal property that it yields illegitimate consequences from legitimate operations. So Everything That, if you look up the handbooks of contemporary syntax, there's a chapter on multidimensionality, which has many interesting consequences about ATB, across the board movement, parasitic gaps, and so on. But None of them work, because they're all based on an illegitimate operation which has this efficiency. Same True of sidewards merge as the same problem. The Same is true in spades this time of late merge, which is widely used. I've used it a number of times. Many others have. Late Merge, first of all, has this problem. It's creating a new, when you draw a tree, it looks as if you can do it. You Just add a line to the tree, and you've got late merge. But If you try to spell it out in terms of the merge operation, you're first creating a new object, which is bad enough, because that's already illegitimate. But Then you're adding a new operation, a substitution operation, which inserts what you've just created at just the right point inside the tree. That's a pretty tricky operation. Try to formulate it. It's a new, complex operation. So Late merge has a double problem. One, the problem of not restricting accessibility. Second, the problem of invoking a new operation, which is really unformulable. It's quite a complex operation, if you think about it. And It's way out of the framework of anything we're talking about. There's many very interesting results that follow from late Merge, very much like multidimensionality. But I Think the way to look at those results is as problems, problems that have been constructed in an interesting way. So We have organized data instead of chaotic data, which is a step forward. But It's only a step towards eventual explanation in terms of something that meets the austere conditions that we're interested in. And that sets interesting empirical problems to address. I Think there are some answers in some cases, which I'll. OK. OK. I've been to that question a little. Somebody, yeah, that's right. Figure out what's going on. Let There be light. OK. Is that the second day of creation, I think? Thanks. You're divine. OK. So Where are we? Sorry. We Need a different God. OK. OK. Are You guys doing that? OK. Well, There's a lot that follows from all of this. I'm kind of leaving it as an exercise to think it through. But If you think it through, what you'll find is that all of these applications, extensions of merge, including the kind that nobody's ever thought of, including others that have been used widely, all have the same problem. They all have exactly the problem that you see with the simplest case. And The problem, again, crucially is that they are constructing what are alleged to be legitimate operations. But When you apply them, you get illegitimate conclusions. And That is the sign that there's something seriously wrong. Obviously, that can't be. So All of those, the entire literature, big literature, that yields very interesting array of results is not legitimate explanation. It's a proposal of problems. It's posing problems that are interesting, important, big step forward. It's useful to have organization of data instead of just chaotic data. But That's not explanation. That's not the goal of linguistics, at least as a science. Well, All of this suggests a kind of a research program. First, determine which subclass of the loosely characterized operations of merge, which subclass of them are, in fact, legitimate. That's a research problem. If You run through it, I Think you think through it, case by case. Again, I Believe this is an exercise. I Think what you end up finding is the originally intended ones not defined properly, but the originally intended ones are probably the only ones. The Rest of the extensions are not legitimate. Interesting Consequences, but not legitimate. The Next problem is to formulate merge. So It gives you just the right ones and then try to explain why that definition of merge is the kind that should be reached on general considerations, general conditions that any linguistic operation ought to meet. And This should be deducible from them, along with third-factor properties, minimal computation, minimal resources. Well, I Won't run through the cases, but we get something that looks like this. And We have to put various conditions on X1 to Xn. So What are the conditions? Well, I Won't bother spelling it out formally. I'll just say it intuitively. Don't lose anything in WS. Spell it out. It means if Y is in WS and Y is distinct from P and Q, it's got to be in the X's. So You don't lose anything. We Don't want something to just disappear in the course of the operation. Second Condition, limit accessibility. In Fact, limit it to one. It has to be at least one because you're constructing a new object. Otherwise, you're not doing anything. But Don't do anything beyond that. So No new accessibility should be permitted by merge. And A third condition is x1, Xn should be minimal. In Other words, don't throw in some new junk that has nothing to do with the operation. Well, what we want to do, of course, is get rid of these. We Want a definition of merge, which has no conditions. But Notice we basically already have that the first one that follows from the no tampering condition. The No tampering condition has to be revised now so that it doesn't apply to syntactic objects, but to the workspace because all the operations are on the workspace. The No tampering condition says if something's in the workspace, don't change it. Well, the most extreme form of changing is to delete it. So You can't delete it. So For many reasonable interpretation of the general condition, NTC, which is part of the strong minimalist thesis, it follows. You're not going to lose anything. This One, we've gotten rid of. It follows from resource restriction, which is a special crucial property of organic computation, it seems, at least for language, but probably quite generally. Probably Related to the general brain activity of massively reducing the data available for computation. So We've gotten rid of this one. But This one implies this one. If You're going to add any more junk, it'll increase accessibility automatically. So Therefore, we don't need that condition. So Therefore, we can get rid of all of these. They All follow from plausible, in fact, necessary conditions on general computational procedures for an organic object. So That gives us the best possible definition of merge. And If you think it through, on principle grounds, and if you think it through, it restricts it just to the original intention, which was never captured by the actual formulations, just kind of in mind. Turns Out that what was in mind was actually created correct, and all of the extensions have to be barred. Now, I should say one word about one of the general conditions, kind of a meta-condition, descriptive adequacy. I've just been assuming that we want them. We Want the operations, of course, to be descriptively adequate. But That's not an innocent notion. You Don't know from data whether they're the right data. Descriptive Adequacy, and this is not just linguistics, all through rational inquiry, all through science, is a theory-determined notion. It's not innocent. You Get a lot of data in, say, chemistry. You Don't know. is this real data or not. There are two kinds of problems the data could have. One, it might involve too many variables, lots of other factors that you're not interested in. In Fact, if you just look at the phenomenon of the world, they're just worthless. Too Many things are going on. So You don't develop physics on the basis of just observation of the phenomenon of the world. If You're in. Silicon Valley, that's the way you do it. But I'm talking about science now, not Silicon Valley. So What you do is try to get rid of the data that doesn't really matter. It doesn't have to do with what you're interested in. But that's a theory internal notion. The Other problem is that you look at the phenomena that are around you. They Usually don't include the relevant data. They don't include the critical experiments, the kind that matter. These are all problems that were faced in the early days of the scientific revolution. And They were sort of settled for the sciences. But Linguistics and the soft sciences haven't really internalized it. So If you go back to, say, the 17th century, the Galilean effort to try to rebuild science on firm grounds, throwing out the neo-scholastic occult properties and so on,. what was important. And He had a hard time convincing the funders, the aristocrats, not the National Science Foundation, convincing the funders that there was some point in this. So It was very hard for the aristocrats to see why. You should care about what happens when a ball rolls down a frictionless plane, which can't happen. Why Should you care about that and not leaves blowing around in the wind, which you see all the time? That's a big move, actually. And If you think about the problem, it's not trivial. So Why is the rate of fall independent of mass? I Mean, if Galileo had done experiments, they wouldn't have worked. Too Many other things would have happened. So What he did was just a thought experiment,. neat thought experiment. Suppose You have two masses which are identical, two objects which are absolutely identical. And Suppose they fall. Well, obviously, they'll fall at the same rate. Suppose You bring them a little bit closer together. That's not going to make any difference. They'll still fall at the same rate. Suppose You bring them so close together that they actually touch in a point. Well, that can't change anything. But Now it has double the mass. So We've proved the theorem without an experiment. Most of Galileo's experiments, if you run through the dialogue and so on, are really like this. So For example, another problem that was bothersome is what happens if you have a sailboat sailing through the ocean. And You drop something from here. Where Is it going to fall? Is It going to fall here? Or Is it going to fall here? Aristelian Physics says it'll fall here. Sailboats Moving forward. So Of course, it'll fall behind where you dropped it from. Galileo Wanted to argue that it's going to fall here because the mass is getting accelerated with the sailboat. Suppose He had done experiments. I Leave it to your imagination to see what you would find about where the thing is falling. You get junk. So You don't do experiments. You Just do critical experiments, often just thought experiments. And For linguistics, that happens all the time. So You read the literature these days, linguistic papers, cognitive science papers, stuff coming out of Silicon Valley, Google. One Of the great achievements heralded is to be able to parse 95% of the sentences that you find in the Wall Street Journal. Suppose You could parse 100% of the sentences and get the right result with training. It would mean absolutely nothing. A sentence in the Wall Street Journal is just an experiment. Is This an acceptable sentence or not? You Don't care if you can match 100% of random experiments. That's of no interest. First of all, a lot of the experiments have the wrong data, too many variables. The Other thing is, they don't include the critical experiments, the kind you're interested in. Can You get parasitic gaps, for example?. Can You get garden path sentences? Well, It turns out when you look at the critical experiments, they fail almost totally. They get maybe 95% of the data, but that's a result of absolutely no interest. And A lot of the field is sort of going off in that direction. Even In the linguistic literature, you find that anyhow, without going on, this concept is not a trivial concept. It's not an innocent concept. A Lot follows from trying to understand what descriptive adequacy means as a theory internal notion. Anyway, we certainly want to be able to achieve the level of understanding that was reached in the 17th century in the sciences. I Think that's a fair goal to try to achieve, understand that there's something significant and serious and theory internal about what we call descriptive adequacy. There are other conditions that have to be satisfied, one of them. Most of them we just take for granted, but one of them call it stability. By That, I mean in the course of a derivation, a syntactic object can't change its interpretation. So For example, if you topicalize, say, Mary's book, I want to read, in the internal system, the non-externalized system, it's going to be Mary's book. But These two objects have to have the same interpretation, like this one. You Can't be saying, I want to read the book that Mary owns, but I'm talking about the book that she bought, let's say. That's sort of taken for granted. Same for ellipsis. If You say, I read Mary's book, and so did Bill, What Bill read is the same Mary, and if she owned it, it's ownership in both cases. So You have to have a general principle that's telling you that anywhere through a derivation, you can't change the interpretation of the expression. Doesn't matter for right now how you express this fact, but it's got to be somewhere in the deep inside the theory. And That has a lot of consequences. We'll come back to that. At This point,. notice at this point, we're getting into a very interesting area of the area where we have to identify what are called copies and repetitions. So Here, the two cases of Mary are copies of each other. Copies are symmetrical. The term is a little misleading, but you have to recognize it's symmetrical. So These are basically the same entity. They have to have a precisely the same interpretation for ellipsis for any operation. They could be copies. Like If I say John saw John, then these two are repetitions. If You look at the generation of the sentence, you had the same formal object, but they were generated independently, and they have nothing to do with each other. This one might as well have been Bill, let's say. That Distinction between copy and repetition is a tricky one. There's an interesting paper by Chris Collins and Eric Groat, which goes through, it's in Ling Buzz, I think. I Don't think they published it, which goes through a lot of problems in trying to distinguish copies and repetitions. I Think we can cut through all those problems in a non-trivial way. And Again, I'm going to leave it as an exercise for you to work it out. But If we appeal to a general, a very general principle, which seems overwhelmingly true, looks false in some cases. But In that kind of situation, it's reasonable to assume that if we understood enough, it would always be true, the principle that's sometimes called duality of semantics. What It comes down to saying is that argument structure, Theta theory in particular, is determined by pure compositionality. So In fact, the strongest conceptual reason, I think, for Dominique's predicate internal subject hypothesis, is that the subject gets a Theta rule. Dominique and Hilda have a lot of other arguments for it. But The basic conceptual argument, I think, gets a Theta rule. So Therefore, it ought to be determined in the general VP system. And In fact, anything that gets a Theta rule ought to be in this system. What About things that are out? And That includes functional categories. They have an argument structure, but they're going to be determined by just essentially by external. What About internal merge? Well, that's always yielding things which have no independent interpretation. So Internal merge is sort of determining aspects of semantics, which have to do with discourse with information structure and so on, not argument structure. But That looks like a very sharp distinction. And If we accept it, then it follows that when at the phase level, when interpretation is trying to determine what's a copy and what's a repetition, it can stand on a pretty high ladder. If Something's in a non-theta position, it's a copy. If It's in a theta position, it's a repetition. That cuts very sharply. Now, It does leave possible cases of ambiguity. If You think through the possible cases, there are some that seem not to be determined by this. But Here, I'll just give you a thesis and ask you to prove it in your spare time. It Turns out, I Think, that there is a kind of a conspiracy of other principles that solves the ambiguities. The Things that seem to matter are Verneu's abstract case theory, which makes distinctions that you don't see sometimes in the morphology. And That turns out to be quite important. Another is Ricci's left periphery theory, which assumes that there are actual positions, like topic, focus, and so on, that a raised element moves to. And Third, connected with this labeling theory, which tells you when those movements are legitimate, when they give you a real interpretation, when you have to move on, when you don't have to move on. I Think If you put all these together, it probably solves the ambiguity problems. But I'll leave that as another exercise. Very Interesting question. You Might think through. One of the tricky cases, which is not easy to deal with, is small clauses. So You might want to think about that. So It has interesting consequences when you try to think how that would work for this. But Again, I'll just leave that as a problem to solve. But If we can solve it this way, then we can solve the copy repetition problem simply by looking at internal merge and external merge. We'll say that every merge operation yields a copy. Nothing else yields a copy. In The case of external merge, the copies that are yielded disappear under the replace interpretation of merge. So You don't happen to see them. With Internal merge, they remain. And Then this phase level algorithm, based on duality of semantics, along with the conspiracy that language is kind enough to preside us with, should resolve the ambiguities of interpretation. That's the general picture. You Can think about it and try to fill in the details. Will? No. I Still don't understand this basic AB tree and the BC tree. The Two trees you put down right at the start, AB and BC. But Parallel merge. Yeah. OK, So let's. I don't understand why that's. You Seem to say that that was ruled out by something. Yeah. So Suppose you have AB and C. And. Now you copy B to C. The Way people draw it is like this. But We're not allowed to draw trees. What you're actually doing is forming a new object, BC. So Now you have two objects, AB and BC. Now You take this one. You Make it arbitrarily complex. Anything You want, you're allowed to merge this. This is accessible, remember. It's a copy of that. So You can keep merging it again. You Merge it to this. You Now have a copy relation here. But It violates every condition. It's back to the initial case. All The cases follow from the simple case. So Parallel merge is out. What's the part of the definition of merge that forbids the original? It blocks this. Not Adding accessible things. If You do this, you're adding this thing, which is accessible. And You're also adding this, which is accessible. And In fact, this, which is accessible. So You're adding three accessible things to the workspace. Actually, this one. These are all new, remember. They're not the ones we started with. There are copies of it. They're new objects. And They're now accessible. This One, personally, because this is the one you've moved. And Now you have two copies of it, both of them accessible. Plus the pair, which is accessible. So You're violating resource restriction, which is the crucial condition. Most Everything is following from resource restriction, which I think is probably a deep property of organic computation. Same is true of side-bridge merge. Collapses for the same reason. Notice In duality of semantics, that has consequences that may be objectionable. So For example, it seems to rule out Norbert Hornstine's theory of control. Norbert's interesting theory of control relating control to raising raises a controlled element to a new theta position. So that's giving internal merge to a theta position, violating duality of semantics. So Here we have a problem. Either Norbert's theory is wrong, or duality of semantics is improperly formulated. OK? Is This the same reason why side-bridge movement is ruled out? Side-bridge movement, same reason. Same Reason, OK. Even More reasons, because at least this, it does form a new object with more accessibility. But There's also the question about how you connect these two separate things, which is another problem. But At least it has the problem of more accessibility. That One runs across all of the extensions of merge. Look Through them, all of them have this property. So They all basically reduce to this very simple operation that I had down here somewhere. In the simplest case., That serves as a paradigm for just about everything. So We're now down to, with various problems hanging around on the side, like what about Hornstein's theory of control? We seem to be converging on exactly what we want, some the simplest possible operation, conceptually justified, which gives us exactly the cases that don't yield illegitimate operations and a limit to the cases that do. Well, what about all the problems that are left over? So It takes a across-the-board movement. The Parallel merge and multidimensionality gives you interesting ways of describing ATB. Notice describing, because none of them are legitimate. But At least you can kind of draw graphs that kind of look as if they're doing something, even though they're not. So What do we want to? A Clarification question. Yeah. Is The point,: was it so doing parallel merge or whatever similar side we're going to merge, you want that to be bad due to the constraints imposed by whatever gives us a restriction of resources? There's not enough stuff accessible to do that. The Way that we restrict resources is just via phases, I guess, if I understand things right. Does That mean that there's a phase-y explanation for the line? You can't have any operate. As Soon as you look at the operation, if it adds accessibility, it's out. Because There's a meta-condition on linguistic operations, probably on general organic computation, but at least on linguistic operations, which says you can't add resources. When You have an operation, it can create one new accessible object. The Object that you've constructed can't construct any others. Is Agree Also sensitive to this constraint? Agree is an operation, is he also sensitive to it? Agree has the same property. So It's also a big position. I'm going to generalize it. It holds also for Agree and labeling and other things. But The interesting cases merge. But None of the operations should add new accessible items. This should be a general property. You have the two AB and BC. Why are the BCs the same? Parallel merge case, yeah. So You end up with two element sets, which you see, AB and BC. Why Do you call the elements of the BC? Why Do you call those accessible? Accessible, There's no reason why they're not accessible. Unless You stipulate somehow that they're not. Why Are they accessible? If This one was able to move from here, then it's able to move from here. I Mean, it's the same structure. Unless You impose some other rule that says I'm not allowed to move. But There's no reason for that. Remember, this is a copy in IM. Copies are always allowed to move. So Unless you simply stipulate, unless you simply stipulate, I'm not allowed to move, which gives the game away, then you're not allowed to have the operation. Because There's no basis for that stipulation. It was originally, notice that it was originally allowed to move. That's the whole point. So Therefore, it should still be allowed to move. It's the same structure. OK? Both Move now. There are two of them. Yeah, either one or the other. So I Don't think there's any way to solve it. It's just out, like all of them, and late merging for further reasons. Well, let's go on and take a look at the things that the empirical cases that were described by these illegitimate operations, like, say, ATB. So What did John buy and read? Well, how do you generate this, first of all?, Well, it would start with, I'll forget the reduced order, and all that stuff, simplify it. What John Bought, What,? let's remember we're talking about the internal core language, not the external language, and what. John Read, what. That's the thing that we generate internally. Well, why doesn't that give the interpretation, standard answer? Because The thing that he read might be different than the thing that he bought. So It's not giving it the right interpretation. However, think it through for a second. This is a copy. This is a copy. This, of course, is a copy. This is a copy. What is the property of copies? Property of copies Is they all delete, except for the top one. That's a general principle of computation. So Like in successive cyclic movement, you delete all the copies. There are languages where you lead some residue somewhere, but we'll forget about that. Basically, you delete all the copies. That's general computation. So That means that in the externalization that we delete this, of course, we delete these. Well, notice that that gives you the right form. It gives you the form. what did John buy and read. Does It give you the right interpretation? Well, In fact, it does, because of the principle of stability. Remember This principle. It says in ellipsis or topicalization or anything, in the CI system, you can only delete if you have absolute identity. Otherwise, you just can't delete. General Property of interpretation at CI. Deletion requires perfect identity. But Notice that that gives you ATB automatically. Nothing else to say. If You think about parasitic gaps, essentially, it works the same way. There's actually an interesting paper by Rene Huyberg. It's not in print, unfortunately. But It's circulating around the internet, which goes through the details of it. But In fact, the basic idea is pretty straightforward. So ATB and parasitic gaps get straight. Notice Also, in the case of parasitic gaps, it follows that A movement won't yield parasitic gaps, because you don't have the initial WH phrase, which will allow the second one to be a copy. So Parasitic gap with A movement would be something like John made a sandwich, which book did he file, or something. Professor? I Don't understand. I Understood you to be saying that in the case of what did John buy and read, it's the same thing that he bought and read. Has to be. Otherwise, you can't delete it at the CI level. Right. But We can say things like, what did John buy and married? What did he read? Oh, that's fine. There's nothing wrong with this interpretation. It's just not ATB. So How do we derive, what did John read and marry each? You Just decide not to call this a copy. Call this something. Call this a repetition. You can always call something a repetition. The book was John saw John repetition. Generate them separately. If You don't say anything, one of the options is ATB. Another option is the other interpretation, which is all you want. And Notice that if you use the ATB interpretation, then you're deleting. So If you're giving the ATB interpretation, you're reading it as ATB with a deletion, because the deletion is dependent on the stability. That's quite understood. I Didn't understand a sentence like, what did John buy? And what did he read? No. What did John buy and marry self? That's fine. Well, it can be different things. It's just totally different sentences. What did John buy and Tom went to the store? You can generate that. But There's no question in the second part. But It just has the interpretation it has. But Isn't that what you just said? Doesn't that force the meaning that there is a thing that John bought and married eight? If You have two WHs, which are copies by definition, because they're in IM non-theta positions. So Therefore, there is the option for them to be interpreted as copies. And if they're identical to delete, which gives you the correlation between the ATB semantic interpretation and the externalization with deletion. You don't have to delete it. You could be saying, what did John buy and what did Tom eat? That's a fine sentence. Just Nothing happened. I have a follow-up question. So I think if one person asks, what did John buy and Mary read, it's possible to answer that question. John bought a book, and Mary read a journal, something like that, no? Not if it's ATB. It has to be the same thing. That's the point about ATB. If You delete, if you delete, you can't get the different interpretation. It just feels like a normal dialogue. And Then it would mean that the first was not an ATB interpretation. What did John buy and Mary read a book? No, no, no, no. What did John buy and Mary read? It's the same thing. Is It possible to answer if John bought an apple and Mary read a book? That's a fine sentence. But There's no doubt. And you have an answer to the ATB question. No, if either it's ATB or it's two independent constructions, that's what ATB is. Think about it. Otherwise, there wouldn't be any ATB phenomenon. No, I Think that the facts are, at least I Think that for many people, what did John drink and Mary read is fine as a question. Well, for those people, they don't have ATB. We're talking about ATB, the phenomenon, right? If Somebody says my language, sorry. But We're talking about people who have ATB, which has this interesting property that you delete the things and interpret them identically. That's the whole point of ATB. So. We're not going into the question of, does somebody have some different language? We're talking about the languages which have a problem to solve, namely ATB. If. You don't have the problem to solve, you don't do multi-dimensionality either. Are There speakers who have both ATB and the meaning ATB meaning and the one which you're saying is the non-ATB meaning? So What did Mary drink and John buy? For Me, that's ambiguous. It could mean that, obviously, for some bizarre reason, Mary drank and John bought the same thing. Or It could be John bought a Cadillac and Mary drank a Pepsi. I Mean, if that's your interpretation, then you don't accept the multi-dimensionality analysis either. Oh, no, I don't. OK, fine. That's fine. I Don't care. I Mean, if somebody has a different system, we can try to describe that one. But I'm talking about ATB. OK, if you don't have it, fine. No, you don't. You have something different. You have something that is different from what is described in the ATB literature, OK? So Then we can talk about that, yeah. But I'm trying to say that what is captured by parallel merge in the multi-dimensionality interpretation, in fact, is yielded automatically with nothing, OK? Same with parasitic gaps. If You look at parasitic gaps, there's a million different problems about them. And This doesn't address those problems. It's just saying the basis, the very basis for parasitic gaps, we already have without anything, including the fact that they are conditional on WHO movement in the first sentence, not A movement. All of that follows right away. Then You get into the morass of problems about different kinds of parasitic gaps, OK? A different kind of question. Yeah. I'm sorry, one more question. I Believe in ATB. It's not the same question. So Do we end up with a system where, if you're in a Theta position, if you have two things in theta positions, they are either copies or repetitions, and you get to choose which one? No, see, that gets back to this conspiracy that I just talked about. If It's John likes John, then they're repetitions. What I'm saying, what I'm giving you as an exercise is to show the following.: And I Think it works. That If you accept duality of semantics, the general principle, and you look at these properties of language, I Think you end up resolving the ambiguities, determining what are copies and what are repetitions. That's a claim, OK? It's up to you to falsify. So A movement, no parasitic gaps, but A movement is claimed to show across the board movement. So This is just dealing with ATB with WH movement, OK? There's other questions about A movement. Yeah, but I'm not talking about this. The Main ones are the ones in the multi-dimension element. The line of reasoning that rejects parasitic gaps with A movement. That explains why you don't get it with A movement. You Don't get it with A movement because the element in the gap, the operator in the gap, has nothing to be a copy of. So These two things are just totally independent of one another. It's like John ate a sandwich before reading. This doesn't mean anything. But It can do it in the case of across the Board. Like, would that also lead us to expect that there is no across the board A movement by the reasoning provider? Across The board A movement is a different problem. It's not dealt with in the multi-dimensionality literature of the kind that I'm discussing. I think it works, but it's basically the same. That's a fair question. We should look at that. Well, this question of the. I have a question over there. Sorry. I have a qualification question about parasitic gaps. So in a sentence, what you don't read after buying? What Did John read after buying? What did John read after buying? The gap in buying, what is the higher copy of it that you would like? Well, that would be something like, what did John read before what John buying? And The two what's are copies. So You get the same phenomenon. If You want the details, look at Rini's paper. But That's basically the structure. And It gives you the core properties of parasitic gaps. It leaves lots of questions open about different kinds of them. Well, let me. Any more? Yeah. The Question might have been had in mind was to do a comment on a sentence like, what two graves were one friend buried in on Monday and the other friend buried in on Tuesday? This doesn't go into that. And There are many other questions about interpret. Anybody Who's interested in the kinds of questions that Barry's raising should? look at the Book, the longest book in the literature with the shortest title. It's called, it's called, And, and the author is sitting over there. It gives the hundreds of pages of very interesting examples of ways of interpreting conjunction and complex structures based on a kind of a neo-Davidsonian event calculus. And There are tons of problems there that are really interesting to solve. But What we're interested here in is asking, what is the basis in the structures for yielding those interpretations? Barry actually doesn't go into that question in the book. I Think you tell me if I'm falsifying it, that Barry describes it fundamentally as a kind of conjunction reduction. But If you think about that formally, it can't mean, I Think, this is a hope, a friendly amendment to the book. It Doesn't mean that you first generate all these huge conjunction things, infinitely many of them for a short sentence, and then get the syntactic structure. It Must mean, that's the thing I'm going to get to next, that you have a syntactic structure. And There are some kinds of interpretive rules that we have to figure out that give you this mass of amazing stuff that you find in there. That's the challenge to face. And If the event calculus approach is the correct one, that will yield the event calculus interpretations of the structures that you generate. Unfortunately, Barry tells me he's now working on another book called, But I Hate to Think About That. OK. Question Here, how do we handle copies, not copies, resumption under this? How Do we handle resumption? How Do we handle resumption, resumptive pronouns? We don't in this system. That's something else. A Lot of things I'm not talking about. OK, yeah, it's a fair question, but not talking about. Not That there's any other way of talking about it, right? It's just that if it exists, we're going to want to have an explanation of it in terms of operations, which meet the austere conditions of explanation. That's the general point. Whatever Problem You're working on, whatever it may be, phonology, semantics, syntax, morphology,. If you want an actual explanation within the context of a program that regards language as part of the natural world, if that's your framework, you're going to have to have explanations into terms that meet this highly austere condition of learnability and evolvability. And About the only thing we know that meets those conditions is merge. So If we can account for things in those terms, like, say, ATB and parasitic gaps, we're in business. Otherwise, we have problems. OK, let me turn to another one, which is problematic and is related to what Barry was just raising to get to it. Well, let me just make one more comment about this. I Won't bother spelling it out. You'll notice that this is a sketch. I Haven't really formalized it, but you can figure out how to formalize it. But When you're left with this definition of merge, the simplest one, and the one that I think is principled, then you can reformulate it in the usual style of transitive closure, friggin'' ancestors. Like, If you're characterizing the set of integers, the standard way of doing it is to say the set of integers includes, say, one, and it's the least set containing one and the successor of any member of the set. That's the standard transitive closure ancestral definition. The Analog here would be the set of workspaces for a given language is the set, not the least set. The Set We leave out least, which includes the lexicon and merge of any triple. That's the set of workspaces. We Don't have to say least, because that's already incorporated in resource restriction. Otherwise, you get the standard recursive definition of the set of workspaces. So It sort of fits the norm. You Can work out the details pretty straightforward. Instead Of that, let's go to something else. There's pretty good reason to think that in addition to merge, which maybe we've now got in the optimal form, there's probably another operation, at least one other operation. That is asymmetrical merge. There are strictly asymmetrical structures, like, say, Young Man. Young Man, the structure is a noun phrase. The Two elements in Young Man are inaccessible. You Can't extract this and leave this the other way around. So We have an asymmetric structure where young is attached to man, and the whole result is still basically man,. adjuncts, essentially. All Adjunct structures, I Think, require pair merge, which is the next operation to look at. Now, There's a very interesting property of pair merge, which has been a thorn in the side of all generative systems since the 1950s, namely unbounded, unstructured coordination. So Things like young, happy, eager, go to Harvard, you can have an unbounded, unstructured coordination. This is a real problem. You Can't generate it by phrase structured grammar. Even Unrestricted rewriting systems, which are universal, in the standard interpretation, don't give you the structures for unbounded, unstructured coordination. Now, notice that since they're universal, you can code it, but that's not interesting. They are universal Turing capabilities. You can find a coding for it, but that's not of interest. If You look at the generation by Phrase structured Grammar, you need an infinite set of rules. It was thought for a while by George Miller and the papers back in the 50s that you could get around this with generalized transformations. But Howard Lasnick had a paper showing that the same problem arises. You'd have to have infinitely many of them. So You can't do it by phrase structured grammar. You can't do it by transformations. There's no way of generating it. It's been a big problem all along. But Notice there is a way of dealing with it in terms of paramerge, namely super-multimensionality. So You have, say, man, and you link to it with any number of adjuncts. They're all on different dimensions. But There's no limit to the number of dimensions. You can paramerge to the element. There's no reason to believe that, just because blackboards are two-dimensional, so is the mind. It does whatever it does. So It could have any number of possible dimensions attaching to a particular point. So For simply, that would give you something like unbounded, unstructured coordination. This can incidentally become extremely complex. Here We get into Barry's type of questions. So For example, one of the conjuncts could be a disjunct, could be John is young, angry, either going to Harvard or to go to MIT, so on. So You can have unbound, and the disjunct could also be unbounded. So You can have unbounded, unstructured disjunction inside of unbounded, structured, unstructured coordination. And This can yield incredibly complex structures. I Leave it to you to give the semantic interpretation of it. But In the nature of the system, you can see that this is possible. Well, instead of trying to, actually, the formalism is not very difficult, so I won't go into it. But Just take the simplest case and take a look at that unstructured. If We can deal with unstructured, unbounded coordination, Then the simplest cases of adjuncts are just automatic. They're the case where there's only one element instead of an unbounded sequence of elements. So We will get simple adjuncts if we can handle the unstructured case. So Let's take a look at that. That's the essential case. Notice A few properties of it. For One thing, it matters what the order is. So If I say, if the order of the adjuncts is young, angry, that's different from angry, young. The Reason is because of something that Jim McCauley noticed back in the 70s, the word, respectively. So If you think of structures with, respectively, then the young, angry man ate the turkey sandwich, and the young, angry men ate the turkey sandwiches and the chicken sandwiches, respectively. The Order of the adjuncts determines the nature of the interpretation. So Somehow, the object that we have in unbounded coordination is actually a sequence. Furthermore, it can have iterations. Like You can say, John is young, angry, young, tired, and so on. You can iterate them. So Basically, the problem is we have unbounded coordination or disjunction either, which has a sequential structure with possible repetitions. And That sequence is interpreted both at the CI level and the externalization level. Now, this does not tell us that linear order enters into syntactic operations. It just tells us there's some object being constructed, which is going to be interpreted in terms of its order and spelled out that way. So We're not crossing the barrier into believing that externalization feeds CI. That's important, even though there's order involved. So What we have to have is something that works sort of like this. If You just think of the general properties, you have to be able to pick out a set of things that are going to be adjuncts. And You have to form from that set a sequence where the elements of the sequence are drawn from the set, but in any possible way. That requires an operator. It's actually an operator that's familiar in logic. It's Hilbert's epsilon operator. In Hilbert's formalization of metamathematics, the core operator that he bases it, is this epsilon operator, which says that out of a set, you can pick an element. Basically, an. It's like indefinite articles. So We need an operator like this, which tells us that given a set in the generation of an expression, given a set, you pick out a sequence. And Then somehow, the elements of this sequence link to something. Each of them is going to link to it independently. So If I say young, angry man, the man is both young and angry. So Independently, they're going to link to something. So What we're getting out of this is a set which, first of all, it'll have to be identified as either coordination, or conjunction, or disjunction. So We have an element here, call it k, which will be plus or minus. Conjunction will be one or the other. And Remember, they can be interspersed, but that's just more formalism. And This sequence will include the pair-merged elements, y1 and some link that it's linking to, all the way up to Yn, and a link that it's linking to. And These links have to be identical all the way through. Like If one of them is a Wh phrase, they all have to be Wh phrases. If You think about unbounded coordination, you can't stick a question in one of the positions. And Of course, you can't have different links. So We have an object that looks like that, and that has to be merged into the general expression. That's the formal problem of dealing with a junction. I Won't bother spelling it out. It raises interesting questions. So For example, one question is, what do you actually link to? So Suppose you have a coordinated noun phrases. John, Bill, Tom, Mary, the guy I met yesterday, et cetera, et cetera. Each of those things is going to link to something. What Is it going to link to? Well, the natural interpretation would be that the individual items here, Y, K, L, if it's a noun phrase, they should all link to whatever is common to noun phrases. Some Thing, call it N. Notice Here that I'm not using the DP analysis, which I think is a mistake. I won't go into it here. But It seems to me that nominal phrases should be regarded as nouns, not as determiners. Determiners are probably something that hang off the outside. And Definiteness is probably a feature of the whole noun phrase. So I'm assuming that the Semitic is the universal language, that the determiner is just a feature on the noun phrase, which distributes somehow differently in different languages, depending on externalization. So You have a feature of the noun phrase, specific, nonspecific. You have a structure which is basically N, with determiners hanging on somewhere, adjuncts that you don't care about. What Is this N? Well, Here we get back to something that Hageet suggested years ago, that the basic structure of language is, again, kind of like proto-Semitic. You have roots which are unspecified as the category. And Then you have categorizers that determine what they are. So For example, N and the root probably paramerged, probably in the lexicon. That's probably the first operation, going back to the paper of yours. The First operation is probably a lexical operation. There are many operations inside the lexicon that involve merge-type operations. One of them is probably categorization. But Notice that this N that's determining that this root is a noun can't be the same as this one. Same with the verb. The V That's determining that something's a verb has to be distinct from what we usually call V or V star up at the phase level. Those are just different elements. They shouldn't be confused, I Don't think. In Fact, these, the ones at the phase level, I think should simply be regarded as phase markers, independent of category. Categories Decided down below, probably in the lexicon. At The phase level, you have something saying Amaphase, a phase marker. And Notice that if you take a look at noun phrases and verb phrases, they have some interesting similarities, some differences, but some similarities. One Similarity is that both noun phrases and verb phrases can be either call it strong or weak with regard to extraction. So The complex noun phrase constraint, it's well known, is strong for specific noun phrases. So It's basically inoperative for non-specific noun phrases. That's the same, sounds like the same distinction as between strong phases and weak phases. So transitive verbal phrases are strong with regard to extraction. You Have to move to the edge. Weak Ones, you don't have to move to the edge. It looks like the same property as weak noun phrases. So Possibly, what we have is something like this. Going Back to classical Greek grammar, philosophy, linguistics, didn't distinguish, we have the notions substantive and predicate, which gives us a four-way classification of things that are substantive, non-predicate. Those are the nominal phrases. Substantive, predicative, adjectival phrases, non-substantive, less predicate, verbal phrases, and then non-either, which is all the junk, prepositional phrases, and so on, some structure like that. And The phase, the crucial phase operations seem to be restricted to those that are the perfect elements, pure substantive, pure verbal, with either the strong or weak property. Now, one of the curious distinctions between noun phrases and verb phrases, which has kept, prevented the thinking of noun phrases as phases, is you don't get the normal escape hatch. But I Discovered a couple of days ago, thanks to somebody, that Uli has an escape hatch for noun phrases. So that fits in the gap that we were worrying about. So Let's say Uli and proto-Semitic are the four languages, and that would have put noun phrases and verb phrases together, ending the idea of using the same notion for the categorizer and the phase marker. They're probably different notions. Adam, yeah. I Would like to go back to the definition of merge, as you mentioned. The Original definition. They're the first. No, the one you put on the board, the first one is not. Per Merge or merge. And you said PQ And the work space. And This is the stuff I Saw. The Original one. Yeah. If You run the clock backwards to the first merge, the stuff you talked about. And Q then becomes the work space, right? Because In that definition, Q equals the work space. There's nothing. The Work space is empty until you put in there. If You start with just two things. No, you start with one thing, right? Before You start merging, there is a work space which is empty. Well, there's no work space until you put some. The Work space is a set. OK, so the work space is the set. PQ in this definition. Because Why do you have three elements? This is the confusing part. The Work space can't have PQ unless they get into the work space. You Merge them. So How do you get them into the work space? OK, So in this recursive way, Q is the prior work space, right? So At some point, you finish merge, and you have a work space. Before You start another set of merge, right? I Don't follow. Before You do any merge, you have nothing. You Just have a lexicon. But In order to put P, let's say, and Q even, into the work space, the work space has to be prior of P and Q being there. It just means there is the option of creating a certain set, which you can put things in if you want. OK, So the set, the work space is essentially something that comes out of PQ being put together. Yes? The Work space has nothing in it unless you put something in it. OK, And then the question is, so the work space, at some point, there's Q, though, if you have another. Is Q equals Q, right? No, only if you've put Q into it. Yes, So the next step, let's say. And You have another merger, which is, let's say, P1, Q1. Then Q1 was the previous work space. I Don't follow. Let's imagine that we haven't. We're just beginning the computation. We Take P and stick it into the work space. Now The work space is the singleton set, including P. Now We put Q in the work space. Now It has two elements. Then We decide to merge them. We Get a new element, set PQ. But not P and Q. There is an interesting empirical question here. How Do you start? And There are various options, and they have lots of consequences. One Possibility is that the only thing that goes into the work space at the beginning is things you've already merged in the lexicon. So If the first, remember, inside the lexicon, there are constructional operations going on, like the words in the lexicon already have structure. Part of the structure, if Hageet is correct, and I'm assuming she is, is taking a category like N, V, maybe broken down into substitution and predicate, and categorizing a root as one thing or another, an operation inside the lexicon, which is giving you a pair, like the pair V hit, let's say, the verb hit. Then That thing can be put in the work space. It already has two elements paired. Then You could put something else in the work space, begin to merge them, put more things in, build up the work space, and so on. Is That the account for the increasing accessibility constraint? No, it doesn't. Because You're basically pre-packaging things to make them accessible for merge. Yes, but you're going to have to. Every Operation that you carry out is going to create something. That's what an operation is, create something. Now, the resource restriction says, don't create too much. Create As little as you can, at least the thing that you're forming, but nothing else. OK? On The accessible? Yes, because I Noticed that when you put in the pair root verb, neither is accessible. Because An adjunct structure, That's what I said before, if it takes a young man, you can't extract young, you can't extract man. Notice, incidentally, that this approach, I mentioned earlier, less time, that there's a way. I've mentioned a paper of Jelko Boskovich, pointing out a way of putting the adjunct island and coordinate island problems in the same package, making them essentially the same problem based, again, on the idea of event calculus, which is a somatic event calculus, which treats a junction like coordination. So It kind of unifies the problems. And That happens automatically here. The Pair merge structure gives you both the islands of conjunction and the islands of adjunction. Now, notice that it leaves the mysteries, just like Jelko's paper does. So If you look at, say, the adjunct island effect, which Jim talked about years ago, it has interesting properties. There are some languages where you can't extract the adjunct. There are other languages where you can't extract from the adjunct and other distinctions of that sort. Those are interesting problems that remain. Furthermore, if you look at adjuncts, they're not uniform. There are some kinds of adjuncts which you can extract from. There are other kinds which you can't extract from. So The notion adjunct is too diffuse. We Have to sharpen it further to find different kinds of probably different kinds of pair merge. Those Problems all are sitting out there, more problems to solve. But You begin to get a unification of the problems. The Adjunct Island, conjunct island effects do reduce the same structure, this sequence that you pick out by the epsilon operator. Now, As far as the point that you're making, as I understand it, it does raise, It leaves open some questions about how you get things from the lexicon into the workspace. There are a couple of ways of thinking about this. And They have different consequences. One Approach is just to take something in the lexicon, insert it in the workspace, and then go on from there. Another Approach is to take something from the lexicon, and to merge it with something that's already in the workspace that's formally slightly different. It has different consequences when you spell it out. You may need both. But Those are questions that you want to resolve, certainly. But I Think it's plausible to believe that the whole system of operations begins by just forming categorized roots inside the lexicon, then building up from there. I Don't see a problem with that. I Don't see the problem you're raising. These categorized roots are completely invisible to syntax. Yes, because the categorized root itself is, but not the part. You can't just raise the root and leave the categorizer. Or conversely, because they're parametric. And That's essentially the idea of development. They don't have to be just moved. They can be targeted by a Greek, for instance, or other things. Targeted by a Greek, for instance, or other things. So They're also invisible to any operation, not just moved. There's other sort of syntactic operations, which will look at the root and move. That raises other questions. Here, we're talking about accessibility to merge. There are questions you can raise about whether you can have agreement into an adjunct, let's say. That's a different question. Here, we're talking about accessibility to merge. Lots of other questions. Let Me: just get in kind of late. So I Mentioned a couple other things that you might deal with in terms of paramerge. There's lots of interesting questions hanging around that have a potential, I think, paramerge analysis. So Let's take one that's been a crazy problem for a long time. There's a strange restriction on extraction from causative-type verbs and perception verbs. So If you look at structures like John Sawville walking down the street, you can passivize this. You can say, Bill was seen walking down the street. On The other hand, if this was a bear verb, walk, then you can't do it. You can't say, Bill was seen, walk down the street. This also holds for the kind of causative-type verbs, verbs like let and make. They don't have the full paradigm, but they have part of it. I saw John, I let John walk down the street. But You can't say, John was let walk down the street. Now, there's a long problem in the literature about how to deal with this. The Only partial solution I've seen is a paper by Norvin. I don't know if it's in print, even, left. The Paper by Norvin in terms of contiguity analysis, which gives a description of how you could lock back. OK, I mean, come back in. I'm insulting you. So The only paper I know that says anything about it is Norvin's paper, which this is the blocking of pacifization out of perception, verbs and causative verbs in terms of contiguity theory, which is an interesting description. But It doesn't cover the whole set of data, because the fact of the matter is, you get the same property without extraction. So For example, in English, these structures are a little bit odd. In Other languages, they're normal. But Things like, there were seen last night, three men walking down the street. And You can't say, three men were seen. You Can't say, they were seen last night. three men walk down the street. So Even without extraction, you get the same property. So It can't be based on extraction. It's got to be based on blocking pacifization. Now, if you think about it, with the let-make type verbs, you can think of those as being basically causatives. They have essentially a causative structure. And In fact, the verb cause itself is kind of resistant to pacifization. So John was caused to leave and that sort of thing. So Suppose we think of the let and make as being essentially causative affixes, the kind that show up in many languages. That would mean that they're pair-merged with C, probably in the lexicon. Well, that gives a unit that's invisible to the operation of pacifization. It's a pair-merged element, which is resistant to whatever we think pacifization is, maybe eliminating the case structure. That would block both the in situ cases and the raising cases. In Fact, it's the only way I know of dealing with that. Now, it's natural for let and make, because they are kind of quasi-causative. But There's a very interesting question that goes way back as to why perception verbs should act the way the causative type verbs do. Actually, Jim Hagenbotham has an interesting paper on this in the 60s, 70s, 80s, I guess, in which he tries to argue that the complement of the perception type verbs is basically some kind of nominal expression with a there verb. Maybe That's an avenue to explain it. But At least using the device of pair-merge, you have an opening to try to account for this strange phenomenon. I Don't see any other way of dealing it. Another Kind of case that's quite interesting is head movement. Head Movement has always been a terrible problem. It doesn't have any of the right properties. It doesn't fit anywhere in the movement system for all sorts of reasons. There is an approach, an interesting approach, by Pisa Kitahara, as the paper is on this, in terms of pair-merge. I'll just give the simplest case. Take T to C movement. So You have a structure of C, T, V, whatever. And At some point, this moves here. How Does that work? It's one of the cases of head movement. Notice That the thing that's moving is really not T. It's V. This is an error of the traditional head movement analyses. But The thing here is usually described as a T with a V adjoined to it by an adjunction operation. But It's actually the other way around. It's a V with a T adjoined to it. One Of the reasons the traditional adjunction operations just don't give you the right result. There are many reasons. What Pisa suggests is that when you get to this point, you've created this object. You have a C. And Then the next operation is to form C, T. Notice That the elements of C, T are not accessible because that's a pair-merged adjunction structure. So You've only added, you've actually enlarged the workspace. But You've only added one. Actually, you haven't even because you've taken C and added T to it. You've kept the workspace the same size. But The only accessible thing you've added is this. So That's permitted by resource restriction. Then The next operation is, you've got this thing. And This thing is just to merge them. When You merge these two things, you get what you wanted. The Structure, C, T, with T down here, V. So That gives you a possible way of looking at head movement. Notice It has a problem. It has the same problem as all of the examples. back to the original one of resource restriction. What happens if you then make some new thing here, x? You start building it up. It ends up being of the form T, V. And Then you decide to merge this to this one. That's crazy. Doesn't make any sense. What blocks that? That's the same paradigm we always have. Now, this gets kind of complicated. But I Think there's a way out of this problem. And I'll leave it to you with something to think about. There's a way out of it by sharpening the notion of restricting computation so that it tells you at each point to add as little as possible to the workspace to still continue. If You think about that, it gives an interesting direction into perhaps blocking this option. It amounts to a condition that will say, you're going to have to merge this one before you create something new. Now, you can't make that too strong, or you won't be able to build up exocentric constructions. You have to put conditions on it that allow just the right ones to block the wrong ones. I'll leave that as another exercise to the reader. HISA has a paper which doesn't go into this. It just gives the proposal. Well, there's a lot more that could be said. I Think I'll stop at this point. These are the kinds of problems that arise when you try to give a principled approach to the nature of explanation. You get some interesting results, get a hoard of problems. The Problems may be presented in an organized form, which is helpful, but we want to go on to try to find real explanations for them. Sometimes You can, as in the case of unifying compositionality and movement or structure dependence. or the basis for reconstruction, things like ATB and parasitic gaps, maybe some of these things. But There's a mass of problems out there to try to deal with in a principled fashion. So That's why it's an interesting field. CHEERS AND APPLAUSE. |
| 5 | Emie_dissertation_cleansed.txt | 7a72cd85-984 | academic_paper | Ms Emilie Szemraj Pembroke College 10 June 2021 Dissertation MPhil in Film and Screen Studies Supervised by Dr\nLisa Mullen\n\nUrban Space, Materiality, and Movement in Post-War American and British Film Noir: Act of Violence and The Man\nBetween\n\nReferencing Style: Chicago, Notes and Bibliography Word Count: 14,952 Words\n\nThis dissertation is submitted for the degree of Master of Philosophy. This dissertation is the result of my\nown work and includes nothing which is the outcome of work done in collaboration except where specifically\nindicated in the text. I confirm that the dissertation does not exceed 15,000 words.\n\nEmilie Szemraj 10 June 2021\n\nAcknowledgments\n\nI would like to thank John David Rhodes, Laura McMahon, and especially Lisa Mullen for their\ncontributions to my development as a scholar and writer during this incredible year of study. It has\nbeen an absolute privilege to work with each of you and join the MPhil in Film and Screen Studies\nProgram this year.\n\nThank you to all my course mates; for your witty debates, aesthetic inspirations, and building of\ncommunity during an incredibly difficult and divided year marked by social distancing and lockdowns. We\nhave mastered the joy of watching film alone, together.\n\nMy heartfelt gratitude to the individuals of Botolph Lane, as well as those at Pembroke College and\nPembroke College Boat Club who I have had the privilege to call friends this year. Special thanks to\nSebastian Benzecry and Sarah Sharp - my time at Cambridge would not be the same without both of you.\n\nLastly, I am beyond thankful for the love and encouragement of my family. Mom and Dad, thank you for\nyour endless and unconditional support. Peter, Kiki, and Lillie, thank you for attuning your\nscience-orientated minds to my love for film and always being the first to listen.\n\nTable of Contents\n\nIntroduction 5\n\nMateriality 8\n\nMovement 12\n\nChapter One: Act of Violence 18\n\nEnley's Flight 22\n\nRegistering Traces 28 The Camera's Reconciliation 33 Chapter Two: The Man Between 41\n\nFragments and Ruins 43\n\nTrajectories 46\n\nRenegotiating Identity 56\n\nConclusion 60\n\nEnd Matter 65\n\nAppendix to Act of Violence 65\n\nAppendix to The Man Between 72\n\nWorks Cited 80\n\nWorks Referenced 81\n\nFilmography 82\n\nIntroduction\n\nModernity is marked by those who move through it. Whether through the figure of the promeneur, walker,\nor fl 1/2neur, those who have considered the individual and his relationship to the city of modernity have\nfound his movement revelatory in its ability to create and investigate historicity while the individual\ncomfortably loses himself in the crowd. Skipping beyond Baudelaire, De Certeau, and Benjamin's\nconceptions of the walker or fl 1/2neur to what Marc Aug 1/2 terms super-modernity, the individual moving\nthrough the city becomes anonymous to the point of similitude, and alienation becomes normality. Yet in\nbetween modernity and super-modernity, the moving individual navigates a city that has literally or\nemotionally undergone cataclysmic catastrophe. He navigates the new spatialities created by the ruins of\nmodernism, and in doing so plants the seeds for the spatiality of the super-modern. What happens to the\nindividual moving through urban space after the Second World War? To investigate this question, I turn\nto an American and a British film noir: Alfred Zinnemann's Act of Violence (1948) and Carol Reed's The\nMan Between (1953). These films respectively form the beginning or end of each director's trilogy of\npost war film noirs. Odd Man Out (1947), The Third Man (1949), and The Man Between (1953) constitute\nReed's set, while Zinnemann's trilogy includes Act of Violence (1948), The Men (1950), and Teresa\n(1951). In considering these two films noir of different national cinemas, I suggest their\nrepresentations of material reality aid in aesthetically defining the murky postwar transition between\nmodernity and super-modernity. In Fred Zinnemann's Act of Violence (1948), American WWII veteran Frank\nEnley sets out on a panicked flight across Los Angeles after the appearance of his fellow veteran, Joe\nParkson. When the two men were trapped in a Nazi POW camp together, Enley informed a German officer of\nParkson's plan to escape via a tunnel dug with the other men from their platoon. Now bent on vengeance,\nParkson tracks Enley throughout Los Angeles. Enley drunkenly allows for a hit to be placed on Joe in\norder to save his own life, but ultimately he saves Parkson and kills both himself and the hitman in\ncrashing the hitman's automobile. In Carol Reed's The Man Between (1953), Susanne Mallinson's tourism of\na divided Cold-War Berlin transforms into her navigation of East Berlin with and romantic interest in\nGerman Ivo Kern after her kidnapping by Eastern Sector gangsters. In returning Susanne to the safety of\nWest Berlin, Ivo Kern attempts to rid himself of the stain of his past war crimes as a Nazi supporter\nand simultaneously gain the good opinion of the Western authorities so that he may relocate to the West.\nWhen he successfully sends Susanne across the border checkpoints, Ivo is shot, and he dies in the no\nman's land in between borders. Both Act of Violence and The Man Between follow their protagonists'\nmovements through urban space. In illustrating how Reed and Zinnemann's cameras engage with the material\nworld in following these movements, I argue that the protagonists of both films move through their\ncities with the aim of distancing themselves from materiality to escape the lingering trauma of the\nSecond World War. Yet, stricken by identities troubled by the materially registered trauma in the very\nsame urban spaces, these protagonists cannot find reconciliation except in facing the consequences of\nmaterial reality. My analysis engages with and departs from Edward Dimendberg's idea of parallel\nspatialities across cinematic traditions. In investigating the possibility of a line of inheritance from\nWeimar cinema to American noir, he concludes the two cinemas evoke each other while retaining their own\nsocial and historical referents. Yet, their parallel does not "imply a common origin or goal."1\nDimendberg uses Siegfried Kracauer's earlier writing on materiality, especially his 1946 essay\n"Hollywood's Terror Films: Do They Reflect an American State of Mind?" to compare and contrast the\nphysical reality in "shots of street life" between these two cinemas. Kracauer hypothesized that both\ncinematic traditions emphasize "mental disintegration," and Dimendberg concludes that it is "the\nrepresentation of threatening urban spaces during cultural crises accompanying postwar demobilisation\nthat unites these two cinemas."2 In the following analysis, I disagree with Dimendberg in that I do not\nconsider the urban spaces in Act of Violence or The Man Between particularly threatening, and Kracauer\nin that these American and British post-war films noir do not emphasize mental disintegration in their\nspatialities. Instead, I apply Kracauer's argument in Theory of Film of cinema's recording the physical\nworld in order to illustrate how the camera uniquely registers urban spaces that hold and emanate the\ntrauma of history in films of two different national cinematic traditions. In following how the\nprotagonists of Act of Violence and The Man Between move through such spaces, I reach a different\nconclusion than Dimendberg. In these films, the material world precipitates disengagement and yet\nnecessities a renegotiation of identity that these characters ultimately seek. Character disengagement\nmay appear similar to "mental disintegration" - as when the materiality of the tunnel in Act of Violence\nresonates with such painful history it drives Enley to the verge of suicide - but unlike a film noir\nlike Scarlet Street (1945), the protagonists of Act of Violence snap back from their disengagement to\nviolently confront the material world. It is space in these films that is fragmented, and Zinnemann and\nReed's cameras offer a reparative materiality by which the spectator can understand its disintegration.\n\nMateriality In the following chapters, my analysis will illustrate that both Act of Violence and The Man\nBetween demonstrate what Siegfried Kracauer calls an awareness of the tensions their materiality produces on\nthe content of their narrative. This "awareness" manifests in how their protagonists navigate space in their\nefforts to confirm, establish, or transform their identities. Ultimately, both Act of Violence and The Man\nBetween suggest that the camera undertakes the agency of noticing and revealing the physical reality of urban\nplaces, though they accomplish this by different means. Kracauer's Theory of Film offers a rich contradiction\nwith which to view the representations of physical reality in Act of Violence and The Man Between. Like Walter\nBenjamin's "text-as-city"3 of the Arcades Project, Kracauer's Theory of Film is perhaps best viewed as a\ncollection of material analyses rather than a unified theoretical work. Noel Carroll suggests that Kracauer\noften relies "on question-begging assumptions that should be the results of his arguments rather than their\npremises,"4 and he especially critically engages with the medium specificity of Kracauer's argument - that\nphotography is film's basic element and hence film should emphasize its photographic nature. Carroll points\nout that Kracauer fails to consider several kinds of non-photographic film that would undermine his argument\nthat photography uniquely creates a film's content (contradictory examples include animation and flicker films\nof alternating leaders).5 I agree with Carroll that Kracauer's work is idiosyncratic in nature and his\nontological claim can be problematic. However, Kracauer's firm directives on how cinematic films reconcile\nrealist and formative tendencies provide a useful framework for investigating how films noir represent\nmaterial reality. Kracauer's contention that "detective films and films of intrigue traffic in narratively\nclosed universes"6 and are thus uncinematic serves as an inciting backdrop to my investigation into how Act of\nViolence and The Man Between reconcile realist and formative tendencies. Working from Kracauer's definition of\nthese terms, in this dissertation realist cinema denotes film's ability to capture the material world as it\nexists without distortion, while formative cinema is a distortion of the natural world as the filmmaker seeks\nto express ideas other than what is inherent in the physical matter before the camera. Kracauer asserts that\n"it is the nature of photography to record and to reveal physical reality,"7 and therefore the formative,\ncreative work of the artist and director should be secondary. This creative work "is endorsed so long as it is\ndedicated to revealing physical reality, rather than, say, to concocting reflexive abstractions or imaginary\nworlds." Kracauer considers many genres of films too formative to accomplish this task, including the\ndetective film, the "theatrical film," and the films of German expressionism, whose use of stages and canted\nangles "neglect the external world." Even as art films, he thinks expressionist films "frequently ignore\nphysical reality or exploit it for purposes alien to photographic veracity."8 In considering two stylized and\noften theatrical films noir that seem to contradict Kracauer's thesis, I suggest these films demonstrate a\ntransnational occurrence of material realism. I argue that both Act of Violence and The Man Between engage\nwith their physical world, and secondly that the material recording (or realism) of the camera proves just as,\nif not more crucial to illustrating their protagonists' struggles with identity than the revealing (or\nformative) work of the director, screenwriter, or studio. Though British and American film noir operated in\ncultures with different registers of the trauma of the Second World War, these films record the material\nreality of their cities to reveal how their characters both struggle to renegotiate their identities. I\nsuggest such a shared use of material reality expands the aesthetic definition of post-war films noir. In his\nchapter, "British Noir," amongst the ambiguity of defining noir as a genre, Jim Leach offers two defining\nfeatures: "1) a corrupt and threatening urban setting in which crime is endemic, and 2) a visual style\nemphasising low-key lighting, deep shadows and unusual camera angles."9 I would add that the camera's\nengagement with physical reality is a further defining feature of the visual style of post-war films noir. In\ndiscussing the problems with overarching definitions of noir as a genre (the scope of which is outside the\npurpose of this dissertation), Andrew Spicer also directs that "some commentators have sought to unify film\nnoir through its prevailing mood or tone, one that can be characterised as cynical, pessimistic, paranoid,\nmorally ambivalent, with a strong sense of alienation and that existence is meaninglessness and absurd."10 The\ncamera's recording of the material world to express the protagonists' changing identities and subsequent sense\nof loss actively contributes to the alienated or morally ambivalent tone of Act of Violence and The Man\nBetween. This contribution suggests that materialist aesthetics also contribute to forming the distinct tone\nthat defines film noir cycles.\n\nMovement The transnational materiality of these films noir is produced by the individual moving through space.\nAs Graeme Gilloch outlines, the crystallization of the fl 1/2neur by Edgar Allan Poe in "The Man of the Crowd,"\nand its recognition by Charles Baudelaire and Walter Benjamin suggest that the detective story and film noir\nare actually born "in the interplay of American and French writers." From Paris, "noir is then re-exported\nback to the United States where it is adopted as the definitive signature of a particular literary and\ncinematic subject matter, style and sensibility."11 The transnational context of Act of Violence and The Man\nBetween as part of the film noir genre especially resonates in each director's nationality as compared to the\nsetting of the film noir, and, in Zinnemann's case, the national film industry in which the film was produced.\nA British film director, Reed shot The Man Between on location in Berlin, as well as at Shepperton Studios\noutside of London. Fred Zinnemann was born and raised in Austria, began working in film as a cameraman in\nBerlin, and then emigrated to New York before moving to Hollywood. Walter Benjamin's fragmented writings on\nmodernity, drawn together in The Arcades Project, provide another set of contradictions for a consideration of\nthe individual moving through space. An application of Benjamin's writings on the fl 1/2neur to sequences in Act\nof Violence and The Man Between will make clear that dualism and bifurcation are central to the individual who\nmoves through space. Benjamin posits the fl 1/2neur as a hunter who can also be hunted; Act of Violence\ndemonstrates such a reversal when Enley switches from fleeing Parkson to locating him to save his life. In his\nbold claim that, "we know that, in the course of fl 1/2nerie, far-off times and places interpenetrate the\nlandscape and the present moment,"12 Benjamin suggests that the fl 1/2neur experiences a doubling of time through\nplace.13 Faced by the abrupt return of his war-time past, Enley faces a doubling of identity that he attempts\nto flee through movement across the city. Despite his efforts, the events of the war surface as he walks\nthrough Bunker Hill, a surfacing precipitated by the richness of the neighborhood's material reality and\ntraces recorded on film. Bifurcation occurs in The Man Between through the literal divide of Berlin between\nEast and West and Ivo's movement between urban place as he attempts to renegotiate his relationship with the\nEast and West Sectors. The crisis and trauma of the Second World War becomes explicit as it is materially\nregistered in the ruins of Berlin. The final dualism these films noir share with Benjamin's ideas is in the\nindividual's movement in the post-war moment between anthropological place and non-place, as defined by Marc\nAug 1/2 in Non-places: Introduction to an Anthropology of Supermodernity. Anthropological places include domestic\nspaces such as a family home and cultural spaces such as music halls and cafes, while non-places include train\nstations, shopping malls, and airports. Aug 1/2 defines anthropological place as "relational, historical, and\nconcerned with identity," while non-places cannot be defined in these terms.14 Benjamin considered the tension\nin modernity between exterior and interior places as manifesting in how the fl 1/2neur does or does not find\nhimself at home in such places. Similarly, Aug 1/2 considers the relationship between an individual's identity\nand space and how these are historically located. He suggests, "identity and relations lie at the heart of all\nthe spatial arrangements classically studied by anthropology. So does history. For all relations that are\ninscribed in space are also inscribed in time."15 After the Second World War, the film noir protagonist\nrenegotiates his identity through movements between anthropological and non-place, rather than simply interior\nand exterior spaces. The need for increased movement results from the rise of centrifugal space from the\nAmerican perspective and the division of space from the British perspective. The protagonists of Act of\nViolence and The Man Between must adapt their movements to stabilize their identities in these altered\nspatialities. In doing so, they step into the space left behind by the fl 1/2neur who has vanished with the\nfragmentation of Europe's cities into ruin, and they navigate between the places of modernity and those of\ndeveloping super-modernity to wrestle with and reconcile identities traumatized by the events of the Second\nWorld War. The nature of Los Angeles and Berlin's post-war spaces both necessitate that their protagonists\nmove at greater speed and that they disengage from material reality in order to ease the post-war trauma held\nin its material traces. This phenomenon, illustrated in the analysis of the following chapters, raises\npoignant questions. If the protagonists' movement creates their disengagement and the camera instead registers\nmaterial reality, who is the descendent of Poe's detective, the hunter of traces? The protagonist, or the\ncamera? In investigating this question, I will consider four kinds of movement that the protagonists of these\nfilms either undertake or are subjected to. These are walking, running, the automobile, and the train. They\nform a progression in their speed and distance from the material world and are differentiated from each other\nby the agency they offer the moving individual. The first two, walking and running, are bodily movements of\nwhich the individual is the agent. Though the classic fl 1/2neur walks, Benjamin suggests "an intoxication comes\nover the man who walks long and aimlessly through the streets. With each step, the walk takes on greater\nmomentum; ever weaker grow the temptations of shops, of bistros, of smiling women, ever more irresistible the\nmagnetism of the next street corner, of a distant mass of foliage, of a street name."16 Walking transforms to\nrunning given enough momentum and speed. Speed has a distancing effect on the individual which may cause\ndisorientation and lack of attention or focus. Walking and running share this distancing quality with the\nautomobile and the train. The automobile and train contrast with walking and running in that they reduce the\nmoving individual's agency and distance him further from the material world by placing him in interior space.\nThis contrast is significant in both films, for when the protagonists become overwhelmed by the trauma\nregistered in the material world that they encounter when walking or running, they seek the increased speed\nand distance of the automobile and train. Conversely, while the automobile and train evidently serve to\nconnect the individual between the distant spatialities of the centrifugal or divided city, they inversely\ndistance him from the material reality of urban place in doing so. It is important to note that neither in Act\nof Violence nor in The Man Between does a protagonist enter the interior space of the train. Rather, the train\npresents the opportunity for movement or the threat of movement's deadly possibilities. Finally, each of these\nfour methods of movement has its own association with different kinds of city place. The walking or running\nindividual is most at home in the urban street, while the automobile is associated with the suburban. The\ntrain provides the most publicly accessible form of movement, but it also dehumanizes space as it slices\nthrough it on established, linear rail-road tracks. As the protagonists of Act of Violence and The Man Between\nuse these movements to form trajectories between place, the camera's observation of physical reality as it\naccompanies the moving individual aids in rendering frequented place into space. "It is the pedestrians who\ntransform a street into a space," Aug 1/2 points out, drawing from De Certeau. In the dialectical relationship\nbetween individual and the place of the street, what does the camera's recording of the protagonist moving\nthrough the material world illuminate about his troubled identity?\n\nChapter One: Act of Violence\n\nIn Fred Zinnemann's 1948 film noir Act of Violence, Parkson's return triggers Enley's struggle to\nreconcile two identities: his role as patriotic veteran and war hero in the eyes of his wife and\ncommunity, and the return of his war-time identity as a traitor and informer. Like Enley's flight from\nParkson, this chapter's consideration of movement begins and ends with the automobile. When Enley first\nlearns from his wife, Edith, that Parkson has visited their home, his first instinct is to hide in\nplace. He latches the door, draws the curtains, and turns off the lights in the front of the house. When\nParkson inevitably comes to their home that evening, Edith and Enley are trapped inside their domestic\nspace until Parkson leaves. Racked with anxiety, Enley departs in his automobile in the middle of the\nnight. Rather than show Enley exit, Zinnemann uses a close-up of Edith sleeping as the sound of an\nengine rumbles. The shot tracks her waking and running to the window, at which point the film cuts to an\noverhead long shot of the automobile reversing from the garage [Fig. 1]. On the spatial level, this shot\nshows Enley drawing away from the domestic space, while on a thematic level he is reversing from his\nidentity as a war hero and upstanding citizen. In the next shot, the automobile continues to reverse,\nbut the film cuts back to Edith at the moment that the automobile pauses to change gear before it will\nmove forward, towards the spatial possibilities the street offers. As we learn from Enley's note to\nEdith, the car heads towards Los Angeles and the Builders and Contractors Convention. In "The Divided\nSelf and the Dark City," R. Barton Palmer identifies Enley's struggle with a double identity to "three\nforms of backward movement." These are movements in the sense of narrative action "interrupted by the\nfilling in of some bypassed gap" (whether through flashback or narration), characterization where a\nnarrator relates the past, or "the return of characters who were thought to belong to the past and who,\nit seemed, had been bypassed as the protagonists embarked on a fresh start."17 All three of these\nnarrative strategies are found in Act of Violence: the first and second in Enley narrating to his wife\nhis betrayal of his fellow soldiers in the POW camp, and the third in Parkson's arrival. These narrative\nmoments lead to the splitting or doubling of Enley's identity between his wartime self and present\nsuburban identity. In fleeing from Parkson and Edith, "thereby surrendering what anchors his self to the\npast and the present as well," Enley moves "toward a reclamation of his true self, as he pays the price\nfor his betrayal and simultaneously saves Parkson from having to commit murder."18 Enley first\nconsciously disengages from his present and past identity, and then in the conclusion of the film\nconfronts both to renegotiate a new identity. This chapter suggests that these past and present\nidentities are literally evoked in the film's material world and that Enley's engagement with the\nmaterial world through physical movement illustrates his renegotiation of identity. Secondly, in\nregistering Enley's engagement with material reality, Zinnemann's camera acts as observer in Enley's\nstead and marks the transition between the kinds of urban place through which he moves. In considering\nhow the material reality of Enley's movement either turns him regressively towards his past war-time\nidentity, progressively towards his identity as a hero, or towards a new identity altogether, I\nunderscore that Enley and Parkson have lost the kind of observational supremacy the fl 1/2neur or\ncity-walker of modernity once held. This loss results from their rushing through the city, disconnecting\nfrom space and leaving an observational void which the camera fills. In his travel between suburb and\ncity, anthropological place and non-place, domestic space and urban space, Enley's movement is\nrepresentative of a decentered Los Angeles that evokes the spatial parallel Edward Dimendberg finds\nbetween Weimar cinema and American film noir. Dimendberg notes that the lingering violence and threat of\nwar present themselves in "a social space permeated by anxiety, a reality in which things have fallen\napart and the center no longer holds."19 This decentering is integral to Dimendberg's constitution of\ncentrifugal films noir in Film Noir and the Spaces of Modernity. Whereas the American film noir cycle of\nthe 1940s captured anxiety in the interconnected spaces of the urban city, he writes that American film\nnoir of the 1950s demonstrates centrifugal tendencies. In these films, characters travel between\nincreasingly distanced locations, while the city's newer suburban spaces manifest the anxiety, fear, and\nviolence previously found in the urban metropolis. In Act of Violence, the automobile is the primary\nagent of movement across these distances and through centrifugal space. Yet Enley most vibrantly\nencounters the material world when he moves by foot, or when he moves in proximity to the violent threat\nof trains, which would render him still (and dead) through nascent motion. In this manner, movement is\nthe primary means through which Act of Violence registers the "the continuum of physical existence."20\nThe automobile especially presents a double opportunity for a director to reveal material reality;\nKracauer categorizes "mad automobiles" as among the inanimate objects that when recorded correctly,\ncapture the spectator's attention as much as the figure of the protagonist. Secondly, movement dominates\namong the aesthetic routes available to the director to capture the material continuum of physical\nreality. Kracauer splits these into the categories of recording and revealing functions. For the\nrecording functions, these are the category of movement, including the chase, dancing, and nascent\nmotion, and the category of inanimate objects. For revealing functions, they are "things normally\nunseen," including the small and the big, the transient, "blind spots of the mind," "phenomena\noverwhelming consciousness," and "special modes of reality." Kracauer points out five affinities of\nfilm, four of which - the unstaged, the fortuitous, endlessness, and the indeterminate - film shares\nwith photography. The fifth affinity, unique to film, is the "flow of life," which Kracauer relates to\nthe cinema's fascination with the street and the representation of the stage within film itself. This\nchapter primarily considers the use of movement in combination with the flow of life in the street,\nwhile the next chapter will discuss objects and the stage.\n\nEnley's Flight\n\nEnley departs the anthropological place of the home for what appears to be the anthropological, crowded\nhuman place of the Blake Hotel. The hotel lobby provides context for how Kracauer considers the fl 1/2neur\nfigure of modernity moving through urban space. In "The Hotel Kracauer," Marc Katz considers Kracauer as\na "hotel fl 1/2neur" because Kracauer viewed the hotel's interior topography as a reflection of the modern\ncity.21 During the Weimar period, the hotel came to be seen as a "distinctly cosmopolitan self-image," a\n"micro-city," in that it housed "exchanges of all sorts - information, money, services, goods." Katz\nargues that these exchanges also include identity, since the hotel is a site of displacement.22 This\nrecalls de Certeau's argument that through walking, the fl 1/2neur practices spatial appropriation through\ndisplacement. Thus "merely crossing the threshold of the hotel carries with it, intentionally or not,\nways of resisting the hotel's manufactured sense regime."23 Drawing on Kracauer's essay, "The Palace\nHotel," Katz also highlights Kracauer's analogy between the fl 1/2neur in the hotel lobby and the\n"moviegoer," as one who can observe "the comings and goings of an international clientele in a lobby\nwhich, [Kracauer] says, has the auratic 'distance-effect' of a film set."24 Sitting in the hotel lobby,\nhowever, Enley fails to observe. Instead, he sits at the bar, drinking with colleagues and answers a\ntelephone call from Edith, who warns him that Parkson is headed to the hotel [Fig. 2]. When Enley raises\nhis eyes to at last observe the room, Parkson is already there. Enley pushes through the crowd and\ndeparts the hotel, setting off on his flight through the urban streets of Los Angeles and ultimately to\nthe derelict neighborhood of Bunker Hill. Zinnemann commences this onset of motion using sequence and\ntracking shots, both principles of editing which Kracauer suggests are particularly disposed to\ncapturing the continuum of physical reality. While Enley speaks with his wife, the camera tracks forward\nto show Parkson, standing elevated above the crowd. The next shot records the moment Enley turns and\nsees Parkson in the same frame [Fig. 3]. Having revealed Parkson first to the camera and then to Enley\nusing the tracking shot, Zinnemann continues to build the sequence using the same tracking motion. In\nsuch sequencing, these close-up and medium shots both reveal spatial and narrative trajectories of cause\nand event and establish the indeterminate expanse of physical reality. Similarly, in its ability to show\nthe "ensemble and diverse elements" of a crowd,25 the traveling shot allows for a more graceful and\nsophisticated power of observation. Kracauer directs that "faced with the task of capturing the\nsubstance of a large-scale landscape, film ought to proceed like a tourist who, in strolling through\nthat landscape, lets his eyes wander about so that his ultimate image of it will be composed of sundry\ndetails and vistas."26 This description again recalls Benjamin's figure of the fl 1/2neur, who observes as\nhe strolls through the city, and Kracauer's own hotel fl 1/2neur, whose eyes wander and note the\nindividuals passing through the hotel lobby. In the Blake Hotel sequence, the camera takes on the\nobservational powers the urban fl 1/2neur once held. The sequence in the lobby continues with backwards\nmovement and then a vacuum of motion shown through sequence shots. Parkson descends the stairs and moves\nthrough the crowd [Fig. 4] until Enley jumps out from behind a pillar and punches him. When Parkson\nfalls, it is as if his nascent motion has stopped the camera's movement as well. While all the\nobservational powers of the fl 1/2neur in the hotel space lie with the camera, Enley retains the agency of\nspatial displacement. His movement draws the camera's attention just as his force in stopping Parkson\nsimilarly jars its motion from fluid travel in tracking shots to sequential movement. While the film\ncaptures Enley's first instance of flight in his departure from the domestic space of the home in the\nautomobile, in this second instance of movement, Enley flees the capitalist space of the lobby on foot.\nIf Enley's backward movement from the home demonstrates a reversal or drawing away from his identity as\na heroic war veteran and model husband, his flight from the hotel and the Builders and Contractors\nconvention within precipitates the reversal from his identity as a developer of suburban homes. Enley's\nmovement away from suburbia engages with the idea of the "single family idyll" characteristic of films\nnoir concerned with centrifugal space. Dimendberg identifies the transition from centrifugal to\ncentripetal space as marked by "novel perceptual and behavioral practices - new experiences of time,\nspeed, and distance - no less than new features in the everyday landscape."27 The automobile proves\nessential to this shift to centrifugal space as it provides the speed by which the individual\nrenegotiates his relationship to such space. Where Dimendberg concludes that "the film noir cycle can be\ncomprehended as a means of generating spatial knowledge, a cultural strategy to bridge the gap between\neveryday life and institutions,"28 movement and speed in Act of Violence also bridge the domestic place\nof the suburbs with the non-places of the Blake Hotel's Lobby, the streets of Los Angeles and Bunker\nHill, and the train tracks near which Parkson and Enley have their final confrontation. Enley's movement\naway from suburban space also offers a criticism of the suburban domesticity that is integral to the\nfilm's centrifugal spatial relationships. In Film Noir and the Cinema of Paranoia, Wheeler Dixon\naddresses how American post-war noirs explore the light fa 1/2ade of American suburbia to reveal a dark,\nthreatening world underneath. Such films noir present "the inherent corruption and complacency of\npost-war life, when forced consensus and idealized conformity were prized above all other\nconsiderations."29 In her text, In Lonely Places: Film Noir beyond the City, Sarah Imogen Smith\ncontextualizes this idea in regard to Act of Violence's suburban space. Because the film centers on "a\nsurvivor plagued by memories he can't share, a victorious hero who wonders what he fought for, the\nreturning veteran personifies the noir view that post-war prosperity rests on unsteady or corrupt\nfoundations."30 The film puts the suburban promise in crisis, revealing that "Enley's new life is as\nflimsy as the bungalows he builds."31 Act of Violence thus presents a criticism of and backward movement\nfrom the "wholesome, sanitized domesticity" of postwar America.32 Indeed, Enley is just as quick to\nabandon his own suburban bungalow as he is his colleagues at the Builder's convention. If Enley flees\nboth his past and present identity, what identity does he have as he runs towards Bunker Hill? Smith\nnotes the converse of the American ideal of mobility: "identities can't shift as fast as bodies can\nmove, and travel can leave people psychologically stranded halfway, neither here nor there . . . the\nloss of ties can result in a sense of being a stranger everywhere - in Dimendberg's words, being "in\nexile at home.'"33 Enley's leaving behind of his present and past identities through movement suggests\nhe moves through the urban streets of Los Angeles alienated and without ties, an idea echoed by Aug 1/2's\ndiscussion of place in the journey narrative. Building from De Certeau's idea that space narratives\n"'traverse' and 'organize' places," Aug 1/2 argues that the plurality of places in the journey narrative\ncan cause "a break or discontinuity between the spectator-traveler and the space of the landscape he is\ncontemplating or rushing through. This prevents him from perceiving it as a place, from being fully\npresent in it."34 Zinnemann's presentation of Enley running through Bunker Hill presents exactly this\ndiscontinuity and subsequent alienation. As Enley submits to the disorientation of movement, in the\nabsence of his attention the camera becomes the traveler attuned to the material world. Traveling on his\nown two feet and relinquished from the speed of the automobile, Enley should be fully present to the\nmaterial world as he moves through the streets of Los Angeles. However, gripped by the terror of\nParkson's pursuit, he only appreciates space for its vacuum and the possibilities of escape it presents.\nHe departs the Blake Hotel in a mad dash and pauses at a corner to consider what options are available\nto him. After he picks a direction he runs, glancing backwards only to ensure that Parkson is not\npursuing him [Fig. 5]. Here Zinnemann does not offer a sequence shot but keeps the camera distanced,\nswitching its position in long shots to show Enley running towards or away from the camera. A singular\npanning shot breaks up the camera's still position as it tracks Enley crossing a hill with downtown Los\nAngeles in the background. In this sequence, the film's more hesitant distance from Enley allows him to\nestablish the material continuum of urban space, rather than demonstrating more agency in establishing\nsuch a continuum itself through the incorporation of close-up shots. However, the camera in this\nsequence accomplishes something Enley cannot. Where Enley fails to perceive his surroundings beyond\ntheir possibilities for escape, the camera records the full material details of the urban streets. It's\nattuned presence to the material world reveals the traces of history in its material surface and the\nboundaries between kinds of urban place.\n\nRegistering Traces\n\nThis attunement begins when Enley turns the first corner, and the camera records the ruffling of the wind\nin the leaves of the trees. The capturing of such detail demonstrates film's affinity for what Kracauer\ncalls the endlessness of the material world. Film finds rhythm in the indefinite; it "may represent an\nindefinite number of material phenomena - e.g. waves, machine parts, trees, and what not - in such a way\nthat their forms, movements, and light values jell into comprehensible and rhythmical patterns."35 In this\nsequence, the rhythm of the material world works in tandem with the cadence of the music. As Enley crosses\nthe hill, the music begins to crescendo, and the camera records an even more rapid fluctuation of the\nleaves as they are stirred by a seemingly growing wind. When Enley descends a side alley staircase, the\nwind has grown forceful enough to wave entire branches beside him [Fig. 6]. He leaves behind these foliage\nlined paths for a Bunker Hill street devoid of plant life and runs under the Angels Flight Funicular while\nthe jazz piano refrain crescendos [Fig. 7]. Zinnemann's camera then begins to record the material\ntextures of this new environment through the litter that spins around Enley. The litter first takes flight\nas Enley passes under Angels Flight Funicular, literally crossing a threshold into this different urban\nspace. This reminder of the more rapid speed of transportation Enley has left behind returns in a\nfollowing shot where Enley checks over his shoulder to see an automobile turn around the corner, as if\nmight be following him [Fig. 8]. Litter continues to scrape along the ground and softly tumble through the\nair in every shot until Enley comes to a stop against the exterior wall of the bar, where he will meet the\nprostitute Pat. Whereas Enley's panicked rushing renders him blind to its details, the camera captures the\ncadence of the material world. It sustains movement while also marking the transition from the foliaged\nlined streets to the desolate stretches of Bunker Hill, a place markedly more decrepit as the rhythm of\nits materiality is demonstrated through litter instead of leaves. In a later sequence, the materiality of\nBunker Hill revives the wartime past Enley attempts to flee. Dimendberg considers Zinnemann's portrayal of\nBunker Hill as suggestive of assuming "the role of the repressed historical unconscious of Los Angeles in\njuxtaposition to its recently constructed suburban present."36 Playing off the typical depiction of Bunker\nHill in film noir "as a haven for alcoholics and criminals," Enley retreats to this space to find a haven\nfrom Parkson's pursuit. Yet instead, the historical traces still held in Bunker Hill resurface the trauma\nof the past Enley attempts to escape. When Enley converses with a gangster who he might hire to kill\nParkson, he is clearly disturbed by the gangster telling him, "you're the same man you were in Germany,"\nas a reason for Enley to overcome his moral qualms in arranging for Parkson's murder. Distraught, Enley\ncowers against the wall before exiting the restaurant into a side alley. Here, the gangster's words echo\nin his head. Enley moves from the alley into a larger tunnel whose space offers a material resonance with\nthe tunnel that his platoon attempted to escape through. As Enley registers this space through walking,\nthe trauma of his betrayal leaks forth, and he begins to hear voices from this fateful event at the POW\ncamp. He grows so distraught at these that he cries out, "don't do it, Joe!" as if he could still stop\nParkson from attempting the escape from the camp [Fig. 9]. Dimendberg suggests that Joe's outburst in this\nmoment results from the temporal and spatial slippage created by Bunker Hill's materially visible history.\nThe space of Los Angeles renders trauma physically, in that "Enley's scream alludes to multiple spatial\ntraumas of modernity, the 'un-homely' condition of the postwar metropolis in which space acquires a\ncomplex layering of temporalities and resonances of the past erupt with explosive force."37 I suggest\nEnley's past erupts here through the combination of Enley's movement and how Bunker Hill makes the past\nvisceral through its material traces. Looking backwards to Walter Benjamin's concept of the trace as well\nas the fl 1/2neur's "descent" into history as he moves through the city aids in elucidating this duality.\nBenjamin defines that "to dwell means to leave traces." Such traces include the imprint that an individual\nleaves on a plush cushion or other soft furnishings in his or her apartment. Benjamin suggests the\nbourgeois individual fills his apartment with such soft furnishings and objects because he no longer can\nleave individual traces in the city, so he must preserve them in the interior.38 Graeme Gilloch proposes\nin Myth and Metropolis that the trace establishes the dialectical relationship of the individual in the\ncity. He explains that "for Benjamin, 'the original social content of the detective story was the\nobliteration of the individual's traces in the big-city crowd,' and that "in the detective story the\nindividual is both enthroned as hero (the detective, the hunter) and denied (he or she vanishes without\ntrace, becomes merely one of the crowd)."39 Both the individual's searching for these traces and his\nability to move within a crowd without leaving traces are characteristics of the fl 1/2neur. Thus,\nDimendberg's notion of eruption suggests Enley's sensitivity to the material traces of Bunker Hill,\nsimilarly to how the fl 1/2neur's movement through the city is a means of descent into the past. The Arcades\nProject includes an excerpt in which Benjamin writes: "The street conducts the fl 1/2neur into a vanished\ntime. For him, every street is precipitous. It leads downward - if not to the mythical Mothers, then into\na past that can be all the more spellbinding because it is not his own, not private . . . In the asphalt\nover which he passes, his steps awaken a surprising resonance. The gaslight that streams down on the\npaving stones throws an equivocal light on this double ground."40\n\nThe fl 1/2neur's excursion into the past is the process of his reading the traces of the city rather than his\nown. While the contrast of Bunker Hill's traces compared to the newness of the suburban should manifest in a\n"resonance" and activate Enley's memories, as he runs from the Blake Hotel he is too preoccupied with forward\nmovement. Rather, Bunker Hill at last becomes a "double ground" when Enley slows down and walks through a\nliteral tunnel. Without speed and its disorientation, Enley becomes mentally rather than physically\ndisorientated. At the speed of the fl 1/2neur, he registers the historical resonance with that of the tunnel from\nwhich Parkson and his platoon tried to escape. The Bunker Hill sequence demonstrates Enley's ability to\nregister material traces, but this occurrence crucially depends on Enley slowing the momentum of his flight\nfrom Parkson. As a singular occurrence amongst Enley's movements through Los Angeles, the sequence emphasizes\nthat the fl 1/2neur figure and his usual powers of observation have vanished along with his cityscape. Movement\nand momentum are normality; strolling is the exception.\n\nThe Camera's Reconciliation After being immersed in the materially produced trauma of the tunnel, Enley looks\nto motion to undo the distancing of identity it has already created. After he yells at Joe to stop, Enley runs\nout of the tunnel to the converging train tracks. He collapses against a chain link fence, where he listens to\nthe diegetic sound of the train. He walks towards the tracks and pauses before deciding to step in the train's\npath [Fig. 10]. A close-up reveals him stunned by the onslaught of the train and its light bearing down upon\nhim. At the last moment, he jumps aside. The train passes between him and the camera, and then we see Enley\nstanding limply before collapsing to the ground. He cries in desperation that he does not even have the\nstrength to kill himself. In the train's potential to end life and identity, its threat of motionlessness\n"produces a shock effect." It demonstrates just how broken Enley has become now that he has left behind both\nof his prior identities. Having fled from them, he languishes, unsure if he is the same man he was in Germany\nor someone else. This scene epitomizes what Kracauer suggests is the chief consequence of nascent motion,\n"that we acutely realize the significance of movement as an integral element of the external world as well as\nfilm."41 The train as an agent of movement is a threatening converse to motion that parallels the dark\nundercurrent of suburbia identified by Dixon and Smith. In the object of the train, Zinnemann suggests that\nthe vehicles or technology that offer the post-war individual speed or connection - the automobile, the\nfunicular, the train, and the telephone - can be deadly. In a sinister relationship, their distancing of the\nprotagonist from anthropological place can drive him to seek a terminal respite by the same speed and\nconnection that first offered him freedom. This nearly terminal crisis of motion runs its course the following\nday. Enley wakes up on Pat's couch and learns that she and the gangster have set up a hit on Parkson for nine\no'clock that evening. Enley returns home to Edith, leaving the cabbie with instructions to come back to pick\nhim up that evening. He then blatantly lies to Edith that Joe is leaving town and "it's all over." The couple\nsit awkwardly in their living room, and Edith suggests Enley might be lying to her. Enley responds, "Lie to\nyou? No." A few minutes later, he slips out of the house and into the cab; the automobile once again carries\nhim from away from his home. This time, he moves toward reconciling identity rather than abandoning it. In the\nclimactic scene of the film, Enley redeems the lies he has said in his identity as a suburban husband as well\nas the betrayal of his wartime identity. He has the cabbie drop him off at the train station, ironically\npassing by Joe's fianc 1/2, Ann, as she enters the station. Enley ventures behind the main train and locates\nParkson at the other end of the pavement bordering the rails. Zinnemann's framing of Enley as he walks along\nthe tracks parallels the earlier sequence when Enley emerges from the tunnel before attempting to commit\nsuicide. This time, Enley's expression is the calmest it has been since the film's beginning [Fig. 11]. Enley\nand Parkson approach each other across the tarmac, in what Wheeler Dixon identifies as "in the tradition of\nthe western genre 'stand-off,' with strategically intercut shots of increasing tightness as the two men\napproach each other."42 When Parkson raises his gun, the sound of the train drowns out Enley's warning of the\nwaiting hitman [Fig. 12]. Without hesitation, Enley steps between the hitman's gun and Parkson and is shot.\nWhere earlier Enley could not submit his fate to the motion of the train, in this sequence the train forces\nhis movement. In saving Parkson, Enley resolves the doubling of identity identified by Palmer. Yet Enley does\nnot die when he is hit. He rises and latches on to the hitman's automobile as it moves away, and only then\ndoes Parkson run after Enley and the automobile [Fig. 13]. Enley and the hitman grapple for the wheel until\nthe driver loses control, and the automobile crashes into a lamppost. Flung backwards onto the ground, Enley\nis killed. The automobile bursts into flames, and the hitman also perishes. Only Parkson is saved - both from\nthe hitman and from the moral burden of killing Enley. Rather than a human avenger, in this sequence the\nautomobile as the instrument of movement that enabled Enley to shed his prior identities ultimately kills him.\nIn this melodramatic ending, Enley accomplishes what he could not when he first stood before the train; he\natones for his betrayal and reclaims an honorable identity.43 Palmer suggests that this ending is\ncharacteristic of redemptive films noir. In these films, protagonists "discover that they can in some sense\ntranscend the past, achieving something like a wholeness of self if only in a death that somehow makes amends\nfor their transgressions."44 This melodramatic, redemptive death is something Enley shares with Ivo Kern in\nThe Man Between. In Enley's death triggered by the consequences of movement, the conclusion of Act of Violence\naffirms the rift between the camera and character introduced in Enley's trajectories through Bunker Hill. In\nAct of Violence, the camera's registration of place through details such as leaves and litter demarcates the\nboundaries between kinds of space and suggests the camera itself has undertaken the burden of De Certeau's\npedestrian who transforms space into place.45 For the viewer, this rift draws attention to Enley's\ndisengagement from material reality and secondly distances the viewer from Enley's interiority. When Enley\nwalks through the Bunker Hill tunnel, the viewer hears the non-diegetic sounds of the voices Enley presumably\nremembers from the POW camp but is not privileged with the vision of his memory. Similarly, when Enley lies to\nEdith, the viewer has already heard him ask the cab to return him to the train station. The viewer only\nreaches closer proximity to Enley's interiority in the face of trauma. It is in the moments when the material\ntraces of urban space recall Enley's betrayal in the POW camp that the camera's perspective most closely\naligns with his. If what the spectator sees through the camera and what Enley sees are different, then does\nAct of Violence really "document" Enley's experience? I suggest that Act of Violence could be considered a\ndocu-noir in the sense that Zinnemann's camera offers a unique aesthetic perspective in capturing the\nmateriality that provokes Enley's anxiety as well as in its observation of the physical world he is too rushed\nto notice. Gene D. Phillips defines "documentary noir" as films that combine documentary realism with film\nnoir thematics. They "tend to portray a frank, unglamorous version of the human condition that is often\nreflected in the shabby settings of the movies."46 In Act of Violence, Enley's troubled morality explicitly\nsurfaces in the grimy and criminal setting of Bunker Hill. Wheeler Dixon sites this gritty realism,\nZinnemann's shooting on location, and the film's attention to "utilitarian 'shabbiness'" as characteristics of\nits quasi-documentary realism.47 Philips and Dixon both seem to suggest that docu-noir involves an increased\nregistration of the materiality of urban space on the part of the camera. This registration becomes especially\nvivid in Enley's movements, where his "human condition," or crisis of identity, becomes evident in how he does\nor does not register material traces. The camera makes explicit what Enley does not see; it documents what he\ncannot. The film is not a documentary so much as the camera documents. In a more general sense, Act of\nViolence engages with Phillips' docu-noir category insofar that it captures how the protagonist engages with\nthe material world in the post-war moment. If Enley disengages with material reality as a means of avoiding\nthe trauma of the war held and evoked by its material traces, the camera's close attention to material reality\nin other "docu-noirs" may similarly contrast with their protagonists' disengagement from the city's\nmaterialism. Phillips' "'docu-noir' tradition"48 may then actually reference films noir that deeply engage\nwith the material world, recording what the protagonist cannot see as a means of revealing his futile attempts\nto disengage from the Second World War's materially produced trauma. Such a phenomenon is not exclusive to the\nlabel of "documentary film." Carol Reed's The Man Between (1953), a theatrical British film noir, also\nmaterially engages with the cinematic world in its protagonists' attempts to escape the trauma registered in\nits material traces. The films noir of the late 1940s and 1950s to which Phillips refers form a tradition in\nthat they engage with the material world. That these films overlap with the centrifugal film cycle established\nby Dimendberg is no coincidence, for the centrifugal spatial tendency is a means by which the film noir\ndirector shows the protagonist's desire to distance himself from the trauma of the war registered in its\nphysicality, and his subsequent alienation and crisis of identity in achieving this distance. Lastly, when\nDimendberg discusses the past "erupting" to confront Enley as he walks through the Bunker Hill tunnel, he\nsuggests that Bunker Hill registers its past through "survivals." He explains this term through juxtaposing\nhow American and European cities preserve history in place. When faced with neighborhoods like Bunker Hill,\nAmerican cities "move forward like modern armies, encircling the islands of resistance they are unable to\ndestroy; the past does not manifest itself in them as it does in Europe, through public monuments, but through\nsurvivals."49 Bunker Hill is a survival of older Los Angeles, the American counterpart to the ruins of Berlin\nin The Man Between which literally survive the Second World War. In their protagonists' movements through Los\nAngeles and Berlin, Zinnemann and Reed both offer a material means of reconciling identity. The death that\nthese reconciliations require suggests that the violence of the postwar period captured in the materiality of\nspace cannot be escaped. The material world both provokes the protagonist's troubled identity, leading to\ndisengagement, and is the means by which he melodramatically comes to terms with it. Enley's ultimate\nreconciliation of identity through a violent death precipitated by and at the hands of a material object\nsuggests an inability to balance this contradiction. In The Man Between, Ivo Kern seeks to change his identity\nby disengaging from the material world of East Berlin, yet he proves himself to the Western authorities in a\ndeath also caused by a material sign. In both films, material reality heals in forcing the ultimate alienation\nfrom the physical world: death.\n\nChapter Two: The Man Between\n\nIn Carol Reed's 1953 film noir The Man Between, the protagonists Ivo Kern and Susanne Mallinson move\nthrough Berlin's streets and urban places as they respectively search to redeem their reputation or\nescape the danger of being mis-identified. Their movement between anthropological place and non-place\npoints towards the modern disintegration of space that parallels the post-war cultural crisis where\nsocial and cultural relationships remain uncentered and unstable. In an introduction to Theory of Film,\nMiriam Bratu Hansen proposes that "Kracauer saw in film and cinema the matrix of a specifically modern\nepisteme, at once an expression of and a medium for the experience of a 'disintegrating' world."50 A\nvisceral reminder of this physical disintegration and the trauma of recent history leaks from the porous\nruins of Berlin in The Man Between. Furthermore, the division of the city into East and West had made a\ncohesive city center impossible. This division presents a spatial parallel to the American shift from\ncentripetal to centrifugal spatiality present in Zinnemann's Act of Violence. Where distance fragments\nthe American post-war city, the border bifurcating East and West explicitly divides The Man Between's\nBerlin. As Ivo and Susanne move across these divisions, Carol Reed's camera records and reveals the\nmaterial reality of this physically and spatially disintegrating world and attempts to create spatial\ncontinuity. Applying Kracauer's central concepts of material film theory, this chapter illustrates how\nThe Man Between offers an "aesthetic matrix of a particular historical experience"51 in how Ivo and\nSusanne navigate Berlin's space. The trauma held in the materiality of the city and its divided\nspatiality undermine and render uncertain Ivo Kern's identity. His relationship with the material world\nsuggests a desire to distance himself from this trauma, and yet ultimately confirms that he cannot\nrenegotiate his identity except in embracing the material world and its consequences. Yet Reed\nsimultaneously nuances the relationship between moving individual and physical reality in the\npresentation of Susanne. In her successful crossing back to West Berlin at the cost of Ivo's life,\nSusanne offers the idea that the British citizen can distance herself from and escape this forced\nconfrontation of war-time trauma, as well as the subsequent renegotiation of identity compelled by the\nprotagonist's interaction with material space. Such a presentation offers both a reassurance to the\nBritish spectator in that they are not subject to the same material violence as the German city dweller\nand a critique: the British spectator will never be able to come to terms with the space disintegrating\naround her unless she embraces and accepts its physicality. On the narrative level, the spatial division\nof the city frames Ivo Kern's struggle with his troubled identity. He aims to cover up, then redeem his\nwartime treason as a lawyer who helped the Nazi regime. He must provide the authorities of the Western\nSector with "an act of good faith" for them to accept him across the border; this act is itself a border\ncrossing in returning Susanne to West Berlin. On the aesthetic level, The Man Between engages with\nKracauer's binary between formative and realist cinematic representation. The film features the tilted\nangles and expressionistic shadows typical of film noir aesthetics and Reed's cinematography and engages\nwith melodramatic elements characteristic of British film noir. Yet the camera's notice of material\nreality proves critical to developing the protagonists' emotional arcs and the renegotiation of their\nidentities, and the film's climactic moments center on this material and formative tension. The Man\nBetween suggests the post-war individual can renegotiate his identity only when he confronts the\nmateriality of the city that has cast it in doubt.\n\nFragments and Ruins\n\nCarol Reed's weaving together of shots of actual ruins in contemporary Berlin with the protagonists'\nwalking suggests their acknowledgment of the lingering wartime violence held in urban places. In her\ntext, Mid-century Gothic, Lisa Mullen suggests that that rubble contains the history of its origin and\nuse in its materiality, but in its fragmented, broken form it also becomes "remote from its former\nspatial and personal meaning."52 Rubble creates spatial possibility. Writing of William Sansom's 1944\nshort story, "The Wall," where a building is destroyed by bombing in the Blitz, Mullen notes that "the\nfiremen can slip between the chunks of masonry because, at the moment of their most dangerous agency,\nsuch walls prove porous."53 Beyond spatial possibility, the porosity of rubble also allows a leakage of\nanxiety and fear. In The Man Between, such porosity allows the trauma of the bombing that produced such\nrubble to surface, and its fragmented nature creates apprehension to how Berlin's spatiality is changing\nin the mid-century modern period. In the introduction to his text, Reading the Ruins, Leo Mellor notes\nhow the literary tradition of modernist writers - including Graham Greene, who worked with Reed on The\nFallen Idol and The Third Man - in the wartime period already depicted a ruined and fragmented\nmetropolis. He elucidates that "postwar culture relied upon such an inheritance - for definition, even\nwhen overtly in reaction or opposition to fragments, debris and the charms of the ruin." Similarly,\nthrough comparison Benjamin's modern figure of the fl 1/2neur assists in considering how the protagonists\nof The Man Between move amongst the porous ruins of Berlin. Reed portrays the protagonists'\nconfrontations with the materiality of the ruined city through movements that include walking, running,\nautomobiles, and the presence of trains. These movements predominantly transpire on the street. While in\nmodernity, the "incessant flow of possibilities and near-intangible meanings . . . casts its spell over\nthe fl 1/2neur" and leaves him "intoxicated with life in the street - life eternally dissolving the\npatterns which it is about to form,"54 these moving patterns of the crowd have now disintegrated to the\ntexture of Berlin's rubble and the porosity of its ruins and snow. Rather than render the moving\nindividual giddy, the materiality of Berlin becomes a veritable toxin. Its materially held violence\nevokes the "elemental catastrophes, the atrocities of war, acts of violence and terror, sexual\ndebauchery, and death" which Kracauer argues "tend to overwhelm the consciousness" and which "only the\ncamera is able to represent . . . without distortion."55 Overwhelmed by the material reminders held in\nthe rubble, the protagonists of the film distance themselves from it by speed and movement. In their\nplace, Reed's camera captures the materiality of the rubble and seeks to establish a material continuum\nas the film's characters move at increasing speed. Reed establishes the link between spatial\npossibilities and material rubble from the film's beginning. Like the first aerial shot of Belfast in\nOdd Man Out and montage of Vienna in The Third Man, Reed selects as the first shot of Berlin a panning\naerial view from the plane on which Susanne is a passenger. Once she reaches the ground, it is not long\nbefore the car in which she and Bettina travel passes by the first ruin on the East and West frontier\n[Fig. 1]. In contrast with the ruins they pass, the interior shots of the automobile have a staged\neffect as the exterior behind the windows appears to be a studio background. Bettina's home is also\nsituated amongst the ruins of the frontier. From the window of the guest room, Bettina and Susanne look\nout into a literal gap or "disintegration" of space created by the bombing of Berlin [Fig. 2]. In the\nexpanse of rubble backdropped by ruined frontier buildings, workers pick at the stones. Bettina tells\nSusanne, "before we only had a view of the houses opposite. It was all built up - all the way to the\neast west frontier, over there." Susanne's observation, "it's near, isn't it?" calls attention to the\nruins' palpable materiality. Even in an extreme long shot, their fragmentation is striking. Bettina's\nresponse - "near enough" - suggests she has already interiorized the anxiety they produce. When Bettina\nand Susanne cross into the Eastern Sector, they walk among bombed buildings and debris. Meanwhile, Ivo's\nerrand boy follows them by bicycle, frequently cycling in circles to stay behind them and match their\nslower pace. Once Bettina and Susanne sit in the caf 1/2, the boy cycles through residents moving rubble in\nclose-up [Fig. 3]. Here, women use pick-axes to dislodge rubble, literally disturbing and renegotiating\nspace. The boy telephones Ivo, prompting Ivo's appearance in the caf 1/2. In this sequence, the slower\nmovement of walking introduces Susanne to the anxieties present in the material space of East Berlin,\njust as in the narrative space she is introduced to Ivo, the protagonist whose identity is most\ncomplicated by the politics resulting from the city's broken materiality.\n\nTrajectories\n\nIn the remainder of the film, movement accelerates as Ivo and Susanne seek to disengage from the\nconsequences of materiality. As the pair get to know each other, they tour the city by foot and go\nice-skating. However, Susanne is kidnapped by automobile, and in the second half of the film the\ntransition from materially present movement in walking to the use of the automobile dominates. This\ntransition is especially evident when Ivo first attempts to return Susanne to West Berlin at a hand off\nplanned for the conclusion of an opera performance. The setting of the opera provides an explicit foil\nof staged life in order to emphasize the camera's attention to material reality. Kracauer valorizes this\nstrategy, explaining that "the more stylized a cut-in theatrical production number, the better does it\nlend itself to serving as a foil to camera-reality. Many a film affording glimpses of opera scenes\nactually exaggerates their artificiality so as to sensitize us, by way of contrast, to the flow of\nhaphazard events surging around that opera isle."56 The stillness of all the figures on the stage,\nincluding the barely moving aria singer and the crescendo of her shrill voice, contrasts first with the\nemptiness of the deserted street, save Ivo and Susanne, and then the swelling, jostling crowd that\nempties out of the opera and into the street at its conclusion. In this mass crowd, Ivo and Susanne\nescape into the automobile brought by Ivo's associates. They then depart on a rapid automobile chase.\nThe setting of the opera, a deeply anthropological and cultural place by Aug 1/2's terms, allows Reed to\ndraw out the material life of the surrounding street through its sequence editing. Of the kinds of\nediting strategies a director may use, Kracauer suggests the sequence shot best penetrates the flow of\nmaterial life. He emphasizes that the crowd "deserves special attention" in the director's use of the\n"big" or "long" shot, as it is film's unique ability to capture crowds in motion.57 However, at the\nmoment the crowd spills out of the opera and onto the street to join Ivo and Susanne, the film offers\nbut one long shot as a few individuals join them on the pavement [Fig. 4]. In the remainder of the\nsequence, Ivo and Susanne jostle through the crowd and into the car, narrowly escaping Halendar's man in\nan alternation of close shots and medium-close shots. Through the camera's close engagement with the\ncrowd, Reed "launches the spectator on a movement enabling him really to grasp the street demonstration\nor whatever tends to overwhelm him through its oversized proportions."58 Thus the audience comprehends\nthe automobile as Ivo and Susanne's destination across the crowd as Ivo does [Fig. 5]. Kracauer likens\nthe alternation of big and small in the sequence shot to the scientific process, as both it and film\nediting have a common aspiration "to comprehend, each in a way, large ensembles and eventually nature\nitself."59 This is a curious comparison as Kracauer considers physical reality a solution to the crisis\nof abstraction caused by science and the scientific process. While the sequence shot clearly allows us\nto comprehend the material life of the crowd, what does it allow us to comprehend in the near empty\nstreet? Reed uses the sequence shot to allow the spectator to grasp the size of the street and its\ndirection and to demonstrate the character's taking note of these spatial possibilities. A long shot\nfirst captures the vastness of the empty street, and then the small close-up features Ivo and Susanne's\nexpressions as they wait to make their escape from Halendar. They exit the opera house onto a street\noccupied by but one other man [Fig. 6]. A close-up of Halendar's men watching them is then intercut with\na close-up of Ivo checking behind him as he and Susanne stroll in and out of the shadows coming from the\nopera doors [Fig. 7]. Here, Halendar's men are parallels to the actors in the opera. Their presence on\nthe steps reminds of the film's staged elements and that the narrative's structure intends for the\ngangsters - rather than the city's materiality -to be the antagonists. Halendar's men add to the tension\nbetween the film's form and the camera's emphasis on physical reality; indeed, before Halendar enters\nthe opera to fetch Ivo and Susanne, Reed cuts from Halendar's men discussing how they "are watching\nevery street," to a long shot of the street running parallel to the train tracks [Fig. 8]. Both the\ntrain and street provide opportunities for movement. Once Ivo and Susanne emerge onto the sidewalk, the\nsequence cuts from close-ups of them waiting or Halendar's men to the occupied Eastern Sector [Fig. 9]\nor the non-place of the street beyond [Fig. 10]. Protagonists and antagonists both register these places\nas indeterminate expanses of possibility; Halendar's men look to them as places to monitor, while Ivo\nlooks toward the street as an escape route. In their noticing of these spatial possibilities as they\npace the sidewalk, Ivo and Susanne embody the movement of the fl 1/2neur in that they are concerned with\nthe "incessant possibilities" of the street in terms of what characters might appear to thwart them and\npossible trajectories of escape. In this sequence, Reed thus presents a duality of the "flow of life" on\nthe film noir street; the material reality or physical reality of the street is both the jostle of\nmoving bodies and its emptiness, in that emptiness creates the potential for individual bodies to move\nthrough space. As it aspires to capture everything within reach - the extent of space the protagonists\ncan see and the paths they might move along through the city - the camera in this sequence is "animated\nby the chimerical desire to establish the continuum of physical existence."60 The camera's intention in\nthis sequence on establishing a physical continuum is characteristic of its movement throughout the film\nas a whole. The camera's constant attention to figures walking through the urban streets as well as the\nincreased speed of the automobile provide subjects or targets that it can track to cover more material\nground. In following these figures essential to the narrative and the film's staged aspects, the camera\nabsorbs physical reality. In its attention to character movement, Reed's camera seeks to soften the\ntransition created by the fragmented city and its gaps and simultaneously allay the anxiety the\nprotagonists and spectator undergo in noticing the changes between kinds of space. The boy on the\nbicycle is a consistent figure by which Reed's camera accomplishes this blurring of borders. He is a\nmoving individual who cannot remain still in Berlin's streets. Though he is not enclosed in an\nautomobile and is therefore closer to the materiality of the street than Susanne or Ivo, the boy's\nhesitance to climb off his bicycle suggests an awareness that in slowing his movement, he too may be\ncaught in its consequences. In the occasions in the film where he might descend to wait for Ivo and\nSusanne, the boy prefers to constantly pedal, removing himself from the grasp of the street's\nmateriality while creating signs that effect the protagonists. For example, early in the film when\nSusanne, Bettina, and Martin attempt to lure Ivo to the embassy to meet with Kestner, the boy pedals\nthrough the snow in the shape of an "S," leaving a material mark which Susanne recognizes as a sign that\nIvo is not coming [Fig. 11]. Automobiles also create material tracks; from the opera sequence Ivo and\nSusanne set out on a face paced car chase through Berlin as they attempt to escape Halendar's men. The\nautomobiles in the film manifest physical reality both in the squeal and swerve of their tires on the\nsnowy, wet streets of Berlin, and the perpetual honking of their horns. They are also material in their\nown right - Kracauer lists "mad automobiles" as among objects "which stand out as protagonists and all\nbut overshadow the rest of the cast"61 in cinematic films. The automobiles in these sequences aid the\ncamera in establishing the continuity of space throughout the city. Kracauer, leaning on the theorist\nLaffay, suggests that cinematic representations of chases can capture the material expanse of reality.\nChases are a window into the "pure poetry of displacement," and can "open up the universe on all sides\nand make us gauge its infinite solidarity."62 While the camera establishes continuity in following the\nautomobile, Ivo and Susanne are simultaneously distanced from material reality. Sitting in the interior\nof the automobile, they are spatially removed from the street. They become agents of spatial\ndisintegration, as they move between locales at a speed of extreme contrast with the figure of the\nfl 1/2neur who is at liberty to observe the street. Indeed, the trajectory of crossing the city takes Ivo\nand Susanne to the border, a political and material non-place, before they leave the vehicle to attempt\nthe crossing on foot through accessing another non-place in the train station and its train, an even\nfaster means of movement. But as Aug 1/2 notes, entering such non-places demands their user prove their\ninnocence. Boarding the train requires presenting a ticket, which releases the user from his usual\ndeterminants; "he becomes no more than what he does or experiences in the role of passenger, customer or\ndriver - temporarily 'possessed,' he is distanced from his usual concerns." Identity checks make each\nuser of public transportation indistinguishable from the other, thus allowing the non-place to create\n"neither singular identity nor relations; only solitude, and similitude."63 When Ivo and Susanne\napproach the station, the police are also checking both identity papers and tickets. Unable to\nsuccessfully participate in the contract of innocent identity that the non-place demands of them, Ivo\nand Susanne cannot access the increased speed of the train in order to further disengage themselves from\nthe physicality of the Eastern Sector. They return to its urban places, and as a result they confront\ntheir romantic desire for each other in the domestic place of the apartment of a paid-off East Berlin\nresident. Likewise, the tension the more staged aspects of the film create with material reality also\nillustrate how Ivo and Susanne develop their relationship with each other as they renegotiate their\nrelationships with the physical world. The sequence in which Susanne confronts Ivo regarding her\nkidnapping demonstrates a clash of realist and formative tendencies. In this scene, Ivo tries to\nconvince Susanne to attend the opera in the evening to mislead Halendar and to spirit her away to the\nWestern Sector. The sequence takes place in a basement room crammed with assorted objects: machines and\ntheir parts, derelict furniture, wires and lightbulbs. These objects elucidate how Susanne expresses her\nemotions to Ivo. Kracauer emphasizes that physical objects can be "surrounded with a fringe of meanings\nliable to touch off various moods, emotions, runs of inarticulate thoughts."64 These effects, termed\nphysical correspondences, are present as soon as Ivo enters and Susanne fiddles with a spare machine\npart in the basement room, apprehensive and suspicious of his role in her kidnapping [Fig. 12]. As the\ntension in their confrontation escalates, Susanne slaps Ivo. Stunned at what she has done, she turns,\nand her head collides with a dangling lightbulb. Disorientated, Susanne spins until she braces herself\nagainst a table while the starkly contrasting lighting of the room also flashes and shifts [Fig. 13].\nFrom the table, Susanne then picks up a heavy anvil, as if to defend herself from Ivo. This directorial\nchoice appears strikingly odd. Susanne's motion is in such high contrast with the effortless movement\nthrough interior space established by the characters that it appears amusing - even silly - instead of\nshocking. Reed's direction to include the lightbulb as an obstacle certainly allows the camera material\ncuriosity, which "itself creates suspense"65 in anticipating Ivo's reaction. Will he confront Susanne in\nanger? Instead, Ivo explains the sequence of events that have led to Susanne's kidnapping and why she\nshould cooperate. Kracauer describes a secondary use of the physicality of objects as a part of a film's\noverarching efforts to recount "the chain of causes and effects responsible for some event." The\nlightbulb in this scene offers an obstruction to Ivo's recount of the causes which have brought Susanne\nto this basement room. As a sequence, the scene is representative of the same use of material reality in\nthe film as a whole, where the anthropological places of the opera house and the domestic space of the\nhome help the protagonists, while the non-places of the derelict basement, warehouse, factory, and\nborder hinder them. The urban street lies between, a passage where Ivo and Susanne as individuals moving\nthrough the city have the option to flee or embrace the possibilities of material reality that each kind\nof space offers. The shocking nature of the lightbulb as an obstacle points to The Man Between's tension\nbetween realist and formative tendencies. In such scenes, the spectator "cannot help recognizing that\nthis little scene is an outright intrusion; that it abruptly introduces an element incompatible with the\nrest of the imagery."66 This intrusion presents an instance where Reed uses material reality to heighten\nand reveal the sudden suspicion and fear with which Susanne regards Ivo in this moment. In doing so,\nReed accomplishes what Kracauer lays out as an achievement of one of his favorite realist directors,\nD.W. Griffith, who he writes understood that "cinema is all the more cinematic if it acquaints us with\nthe physical origins, ramifications, and connotations of all the emotional and intellectual events which\ncomprise the plot," and that "it cannot adequately account for these inner developments unless it leads\nus through the thicket of material life from which they emerge and in which they are embedded."67 In The\nMan Between, both Ivo and Susanne's character development depends on their trajectories through the\nphysical reality of Berlin, especially the final third of the film in the Eastern Sector. The climax of\nthe film resolves this development as Ivo and Susanne complete the renegotiation of their identities in\nheightened movement caused by material reality.\n\nRenegotiating Identity In the final sequence, Ivo accomplishes his ambition to move Susanne from East to West\nBerlin in his death in the no-man's land between the two border checkpoints. The methods by which The Man\nBetween records and reveals physical reality through the moving individual culminate in this scene. While Ivo\nand Susanne cross the border, the movement of Ivo's errand boy on the bicycle has deadly consequences. In the\nfinal sequence, the bicycle creates another physical sign that condemns Ivo, whereas the boy's earlier "S" in\nthe snow helped save Ivo from a trap. When the laundry truck in which Ivo and Susanne hide passes inspection\nto cross to the West, the engine stalls. For a minute, Ivo and Susanne wait in expectant anxiety. Just as the\nengine rumbles to life and it seems they may pass, the errand boy's bicycle gives them away. The border patrol\nofficers spot the boy cycling behind the laundry truck [Fig. 14], and they order it to stop. In an effort to\nsave Susanne, Ivo jumps out the back, capturing the guards' attention as he runs away towards the Eastern\nSector. He then turns around and runs after the truck, across the border zone between checkpoints. Ivo reaches\nto take Susanne's outstretched hand, but only inches away he is struck by rifle fire. He falls and keeps\nreaching for Susanne [Fig. 15], before two more bullets cause him to collapse in the snow [Fig. 16]. Consigned\nto the non-place between the two borders, Ivo dies before he can realize the rewards of his heroism and repair\nhis reputation in the West. As earlier the sign in the snow saved him, at this pivotal moment the bicycle's\nsign leads to Ivo's demise. As equally important to this scene is movement. Ivo's short and ill-fated chase\nends abruptly. His movement through the border zone comes to a complete standstill, yet the truck and the\ncamera mounted on it continue to move away from him [See Fig. 16]. The camera's movement in this scene\ndemonstrates film's affinity for recording nascent motion, where characters abruptly ceasing to move "produces\na shock effect." Kracauer argues that this sudden "vacuum of movement" emphasizes the centrality of movement\nto the spectator or audience.68 Through contrast, this scene shocks the spectator in the same intruding manner\nas the lightbulb scene. The sequence's conspicuous staged elements present a second contrast that is integral\nto how Susanne completes her character development. As Ivo runs to reach Susanne, the background of the scene\nclearly changes to a studio-shot backdrop. As Susanne physically leaves the Eastern Sector and moves away from\nIvo, he and his half of Berlin gain a staged artificiality. Through this editing choice, Reed confirms the\ndifference between Ivo and Susanne's relationship to urban space. Where both engage with or distance\nthemselves from the material reality of urban space as they move through Berlin, Ivo must confront its\nconsequences, while Susanne may escape them. Through Susanne's opportunity to survive and escape the spatial\ndisintegration of Berlin, The Man Between engages with the melodramatic side of British noir. Kracauer\nacknowledges film's susceptibility to dwelling on the sensational but also insists it can serve a revealing\npurpose. He explains that "the point is, rather, that the cinema does not simply imitate and continue the\nancient gladiator fights or the Grand Guignol but adds something new and momentous: it insists on rendering\nvisible what is commonly drowned in inner agitation."69 In the discussed sequences from The Man Between, I\nhave illustrated that Susanne and Ivo's "inner agitation" provoked by the material reality of the urban leads\nthem to movement and speed as a means to disengage from the trauma held in its material traces. Throughout the\nfilm, material reality provokes anxiety or emotional responses in the psychological correspondences of objects\n(how they relate the physical world with the psychological one), the demonstration of the indeterminate\nexpanse of space, and the use of sequencing to comprehend material reality and spatial possibilities. The\ncamera's affinity for material reality links the protagonist's intentions and interior developments to the\nphysical world and the fragmentation it emanates. Where The Man Between does not feature backward movement\nsynonymous with character development in the way that Act of Violence does, Ivo and Susanne renegotiate their\nidentities and desires in relation to physical reality, just as Enley comes to terms with and absolves his war\nand present identities in moving through the materiality of Los Angeles. However, Reed's camera as an observer\nof physical reality works to connect the disparate places throughout the city as it closely and consistently\nfollows character movement. Though it matches the protagonist's level of anxiety, preferring to remain\nalongside bicycles or inside automobiles rather than on foot as the character's urgency to escape the material\nworld increases, through sequencing or its attention to objects the camera continues to act as an observer of\nthe physical continuum of space. This attention to linking place to movement comforts the spectator in the\nface of the fragmented, bombarded city filled with spatial gaps. Where the characters are hesitant to engage\nwith the city's fragmentations, Reed's camera illustrates it is possible to traverse them and make sense of\nthis disintegration. The camera reaches the conclusion that confronting material reality is the only method to\nassuage the anxiety its fragmentations induce well before Ivo and Susanne do. In this sense, the camera\nundertakes the same momentum of the fl 1/2neur, who "advances over the street of stone, with its double ground,\nas though driven by a clockwork mechanism. And within, where the mechanism is ensconced, a music box is\npalpitating like some toy of long ago."70 The camera is driven by the photographic mechanism, and the rhythm\nit attunes to is the cadence of the protagonists it follows. Conclusion\n\nIn moving through Act of Violence and The Man Between as material texts, a tragic dualism that evokes\nthe alienation of the post-war moment surfaces. In order to survive the trauma of the post-war moment\nand the crisis of identity it perpetuates, the protagonist distances himself from the physical reality\nof the city through speed of movement. These physical movements are represented and paralleled by an\naesthetic dualism on the part of Zinnemann and Reed's cameras that observe the material world as they\nfollow their protagonists' trajectories. In each case, the camera's observation presents a converse to\nthe protagonist's motion. Whereas Enley's bodily motion connects urban places, Zinnemann's camera\ndemarcates spatial transitions. Where Ivo and Susanne physically detach themselves from the materiality\nof the street in the interiority of automobiles that also move them between distanced locales, Reed's\ncamera follows their movement between places and establishes spatial continuity. In each case, the\ncamera's recording and revealing of material reality shows the protagonist becoming disorientated or\ndetached. Seeking an escape from the trauma of material reality, the protagonist enters non-place. He\nfinds similitude and solitude, and in its subsequent alienation his identity becomes more troubled than\nbefore. Act of Violence and The Man Between are firstly differentiated in the material representations\nof their cities. Berlin lies in visceral ruins which Reed does not shy away from showing in close-up,\nand the border between East and West makes countless reappearances. Division in Act of Violence takes\nthe form of distance; Enley and Parkson move between suburbia and the city. Though these films\ndemonstrate a spatial parallel in their protagonists' disengaging from material reality, their\nhistorical and social referents suggest Reed and Zinnemann accomplish this to distinct ends. In Act of\nViolence, Enley retreats from certain kinds of space - such as domestic suburbia or the hotel lobby - in\norder to shed his past and present identity. The materiality of parts of Los Angeles, notably the\nneighborhood of Bunker Hill, holds traces that remind him of his betrayal and war-time identity. When he\nruns through the neighborhood, he moves too quickly to engage with its traces, and the camera instead\nregisters them on his behalf. However, when he slows down in the tunnel, the past rises to the surface,\nand he walks on a veritable "double ground." The speed of his movement as necessitated by the\nconstruction of post-war Los Angeles (as well as the plot) has a direct relationship with whether Enley\nregisters material reality, and conversely, whether Zinnemann uses the camera to engage with the\nphysical world in Enley's place. Enley's engagement or disengagement with materiality is a measure of\nwhether he flees, confronts, or resolves his troubled identity precipitated by Parkson's return. In The\nMan Between, however, Ivo and Susanne's engagement with physical reality suggests that where the British\ncitizen can escape the still tangible consequences of post-war physical fragmentation and spatial\ndivision leading up to the Cold-War period, the German resident cannot. Both Ivo and Susanne distance\nthemselves from the materiality of the street through movement. Aware of the dangers of being caught by\nthe authorities when engaging too deeply with anthropological or non-place, they spend most of the final\nthird of the film in automobile chases or hiding out in an apartment. However, where Susanne survives\nthis distancing from materiality, Ivo does not; his own errand boy's bicycle gives him away. Susanne, on\nthe other hand, crosses the border and returns to the safety of West Berlin. By the conclusion of the\nfilm, her kidnapping and excursion into the Eastern Sector appear as tourism, or even a fantasy - a\nperspective Reed accomplishes both through her engagement with physical reality and some of the more\ntheatrical or staged components of the film's sequences. Ultimately, Reed offers in The Man Between a\nreassuring reparative materiality for the British spectator. Though bombed cities and ruins are\nviscerally tangible both at home in the United Kingdom and abroad in Europe, the British citizen can\nstill escape the divided spatiality that is emerging from this ruined physical reality. In Ivo's death\nthe film also offers a critique of this privilege; escapism only treats the symptoms for materially\nregistered trauma - it is not a cure. The relationships between urban space and identity that these\nfilms manifest in the materiality of characters moving through space suggests they center on a radically\ndifferent kind of disillusionment and alienation than that experienced by the hard-boiled detective in\nfilms noir of the 1940s. Zinneman and Reed register how the unresolved traumas and social upheavals of\nthe Second World War manifest in their protagonists morally uncertain and troubled identities, and\nsuggest through their film's engagement with material reality that these protagonists may also find\nresolution - if a violent one - when they ultimately give in and engage with the physical world rather\nthan distancing themselves from it. These films offer to the spectator grappling with the changing\nspatiality of the British or American post-war city the same hard truth; the days of observing the city\nas the fl 1/2neur once did have vanished. One can no longer lose themselves in the crowd, part of the flow\nof life while unnoticed and at leisure. Neither can the moving individual yet achieve the anonymous\nsolitude and similitude of Aug 1/2's traveler in super-modernity. Just as it demands increased speed of\nmovement, the post-war city demands a confrontation from the moving individual of mid-century modernity.\nThese films' reparative materiality remains differentiated by their protagonists' roles as insiders and\noutsiders and the inverse relationship the camera takes in its attention to the physical world. While in\nAct of Violence Enley moves in his near native city, Zinnemann's camera works to differentiate the kinds\nof urban places he sprints through. Where Susanne is an outsider touring and scrutinizing Berlin, Reed's\ncamera works to preserve spatial continuity and fill the porous gaps of the fragmented city. Act of\nViolence and The Man Between align in their affirmation to the spectator that the sooner one\nrenegotiates their movement through urban space, the sooner they may find a reconciliation of identities\nrendered uncertain by the post-war moment. These films complicate Kracauer's thesis that film's\nrecording of material reality offers a balm to the modern crisis of abstraction perpetuated by science.\nBoth Enley and Ivo face death, the ultimate form of alienation. Yet in doing so, they resolve identities\ntroubled by the resonance of the material world. In this sense, both Act of Violence and The Man Between\noffer an uneasy redemption of physical reality. The city has changed and will not wait for you; if you\nwant to dwell within it, you must embrace its painful materiality.\n\nAppendix to Act of Violence\n\nFig. 1.\n\nFig. 2.\n\nFig. 3.\n\nFig. 4.\n\nFig. 5.\n\nFig. 6.\n\nFig. 7.\n\nFig. 8.\n\nFig. 9.\n\nFig. 10.\n\nFig. 11.\n\nFig. 12.\n\nFig. 13.\n\nAppendix to The Man Between\n\nFig. 1.\n\nFig. 2.\n\nFig. 3.\n\nFig. 4.\n\nFig. 5.\n\nFig. 6.\n\nFig. 7.\n\nFig. 8.\n\nFig. 9.\n\nFig. 10.\n\nFig. 11.\n\nFig. 12.\n\nFig. 13.\n\nFig. 14.\n\nFig. 15.\n\nFig. 16.\n\nWorks Cited\n\nAuge 1/2, Marc, and John Howe. Non-places : Introduction to an Anthropology of Supermodernity. New York: Verso,\n1995.\n\nBenjamin, Walter. The Arcades Project. Edited by Rolf Tiedemann, Howard Eiland, and Kevin McLaughlin. London:\nBelknap Press, 2002.\n\nCarroll, Noel. "Kracauer's Theory of Film." In Engaging the Moving Image, 281-302. New Haven, CT: Yale\nUniversity Press, 2008. Dimendberg, Edward. "Down These Seen Streets A Man Must Go: Siegfried Kracauer,\nHollywood's Terror Films, and the Spatiality of Film Noir." New German Critique 89 (2003): 113-143.\nDimendberg, Edward. Film Noir and the Spaces of Modernity. Cambridge, MA: Harvard University Press, 2004.\nDixon, Wheeler Winston. ""Act of Violence" and the Early Films of Fred Zinnemann." Film Criticism 18/19, no.\n3/1 (1994): 30-45. http://www.jstor.org/stable/44076036. Dixon, Wheeler Winston. Film Noir and the Cinema of\nParanoia. New Brunswick, NJ: Rutgers University Press, 2009. https://hdl.handle.net/2027/heb.08059. Gilloch,\nGraeme. Myth and Metropolis : Walter Benjamin and the City. Oxford: Polity, 1997. Gilloch, Graeme. "Noir sans\nfrontiers: reflections on the transnational fl 1/2neur-as-detective." Soci 1/2t 1/2s vol. 135, no. 1 (2017): 31-42.\nhttps://www.cairn.info/revue-societes-2017-1-page-31.htm. Hansen, Miriam Bratu. Introduction to Theory of Film\n: The Redemption of Physical Reality, by Siegfried Kracauer, vii - xlv. Princeton, NJ: Princeton University\nPress, 1997. Katz, Marc. "The Hotel Kracauer." Differences: A Journal of Feminist Cultural Studies 11, no. 2\n(1999): 134-152. https://muse.jhu.edu/article/9601. Kracauer, Siegfried. Theory of Film : The Redemption of\nPhysical Reality. Princeton, NJ: Princeton University Press, 1997.\n\nLeach, Jim. "British Noir," in International Noir, edited by R. Barton Palmer and Homer B. Pettey. Edinburgh:\nEdinburgh University Press, 2014.\n\nMellor, Leo. Reading the Ruins : Modernism, Bombsites and British Culture. Cambridge: Cambridge University\nPress, 2011.\n\nMullen, Lisa. Mid-Century Gothic. Manchester, England: Manchester University Press, 2019.\nhttps://doi.org/10.7765/9781526132789.00004.\n\nPalmer, R. Barton. "The Divided Self and the Dark City: Film Noir and Liminality." Symploke 15, no. 1/2\n(2007): 66-79.\n\nPhillips, Gene D. Out of the Shadows : Expanding the Canon of Classic Film Noir. Blue Ridge Summit, MD:\nScarecrow Press, 2011.\n\nSmith, Imogen Sara. In Lonely Places : Film Noir beyond the City. Jefferson, NC.: McFarland, 2011.\n\nSpicer, Andrew. European Film Noir. Manchester: Manchester University Press, 2019.\n\nWorks Referenced\n\nBrowne, Nick. Refiguring American Film Genres History and Theory. Berkeley, CA: University of California\nPress, 1998. Certeau, Michel De. The Practice of Everyday Life. Translated by Steven Rendall. Berkeley:\nUniversity of California Press, 1984. Cordes, Vojislava. "The Agitated City: Urban Ambiguity and New York's\nNoir Metropolis," Quarterly Review of Film and Video, 35 no. 4 (2018): 372-396. DOI:\n10.1080/10509208.2017.1422377 Drazin, Charles. The Finest Years : British Cinema of the 1940s. London: Andre 1/2\nDeutsch, 1998. Hansen, Miriam. "The Mass Production of the Senses: Classical Cinema as Vernacular Modernism."\nModernism/modernity 6, no. 2 (1999): 59-77. doi:10.1353/mod.1999.0018 Kracauer, Siegfried. Siegfried\nKracauer's American Writings : Essays on Film and Popular Culture. Berkeley: University of California Press,\n2012.\n\nKrutnik, Frank. In a Lonely Street : Film Noir, Genre, Masculinity. Florence: Taylor & Francis Group, 1991.\nProQuest Ebook Central. https://ebookcentral.proquest.com/lib/cam/reader.action?docID=178317 Larsen, Neil, and\nConcha, Jaime. Modernism and Hegemony : A Materialist Critique of Aesthetic Agencies. Minneapolis: University\nof Minnesota Press, 1990. ProQuest Ebook Central.\nhttps://ebookcentral.proquest.com/lib/cam/detail.action?docID=310200. Lefebvre, Henri. The Production of Space\n[1974]. Oxford: Blackwell, 1991.\n\nNaremore, James. More Than Night : Film Noir in Its Contexts. Berkeley: University of California Press, 2008.\nPhillips, Gene D. Creatures of Darkness. Lexington, KY: University Press of Kentucky, 2015. Sobchack, Vivian.\n"Lounge Time: Post War Crises and the Chronotope of Film Noir." in Refiguring American Film Genres: History\nand Theory by Nick Browne. Berkeley, CA: University of California Press, 1998.\nhttps://hdl.handle.net/2027/heb.08139. Von Moltke, Johannes. The Curious Humanist 1/2: Siegfried Kracauer in\nAmerica. Oakland, California: University of California Press, 2016.\n\nFilmography Act of Violence. Directed by Alfred Zinnemann. United States: Metro-Goldwyn-Mayer. 1948. The Man\nBetween. Directed by Carol Reed. London: London Film Productions. 1953.\n\n1 Edward Dimendberg, "Down These Seen Streets A Man Must Go: Siegfried Kracauer, Hollywood's Terror Films, and\nthe Spatiality of Film Noir," New German Critique 89 (2003): 133. 2 Ibid., 127. 3 Graeme Gilloch, Myth and\nMetropolis: Walter Benjamin and the City (Oxford: Polity, 1997), 20. 4 Noel Carroll, "Kracauer's Theory of\nFilm," in Engaging the Moving Image (New Haven, CT: Yale University Press, 2008), 300. 5 Ibid., 291. 6\nCarroll, "Kracauer's Theory of Film," 290. 7 Ibid., 288. 8 Siegfried Kracauer, Theory of Film: The Redemption\nof Physical Reality (Princeton, NJ: Princeton University Press, 1997), 39. 9 Jim Leach, "British Noir," in\nInternational Noir, ed. R. Barton Palmer and Homer B. Pettey (Edinburgh: Edinburgh University Press, 2014), 4.\n10 Andrew Spicer, "Introduction," in European Film Noir (Manchester: Manchester University Press: 2019), 2. 11\nGraeme Gilloch, "Noir sans frontiers: reflections on the transnational fl 1/2neur-as-detective," Soci 1/2t 1/2s 135,\nno. 1 (2017): 38. 12 Walter Benjamin, The Arcades Project, ed. Rolf Tiedemann, Howard Eiland, and Kevin\nMcLaughlin (London: Belknap Press, 2002), 419. 13 See Benjamin, 416 for the fl 1/2neur walking on "double\nground." 14 Marc Auge 1/2, Non-places : Introduction to an Anthropology of Supermodernity, trans. John Howe (New\nYork: Verso, 1995), 77. 15 Ibid., 58. 16 Benjamin, The Arcades Project, 417. 17 R. Barton Palmer, "The Divided\nSelf and the Dark City: Film Noir and Liminality," Symploke 15, no. 1/2 (2007): 74. 18 Ibid., 76. 19 Edward\nDimendberg, "Down These Seen Streets A Man Must Go," 134-135. In his description, "the center no longer\nholds," Dimendberg evokes W.B. Yeat's 1919 poem, "The Second Coming" describing post WWI Europe. 20 Kracauer,\nTheory of Film, 63. 21 Marc Katz, "The Hotel Kracauer," Differences: A Journal of Feminist Cultural Studies\n11, no. 2 (1999): 135. 22 Ibid., 138-139. 23 Katz, "The Hotel Kracauer," 140. 24 Ibid., 146. 25 Kracauer,\nTheory of Film, 51. 26 Ibid. 27 Dimendberg, Film Noir and the Spaces of Modernity, 169. 28 Ibid., 171. 29\nWheeler W. Dixon, Film Noir and the Cinema of Paranoia (New Brunswick, NJ: Rutgers University Press, 2009), 1.\n30 Imogen Sara Smith, In Lonely Places: Film Noir beyond the City (Jefferson, NC: McFarland, 2011), 46. 31\nIbid., 47. 32 Ibid., 52. 33 Smith, In Lonely Places: Film Noir beyond the City, 34. 34 Auge 1/2, Non-places, 84.\n35 Kracauer, Theory of Film, 68. 36 Dimendberg, Film Noir and the Spaces of Modernity, 159. 37 Dimendberg,\nFilm Noir and the Spaces of Modernity, 161. 38 Walter Benjamin, The Arcades Project, 9. 39 Graeme, Myth and\nMetropolis, 141. 40 Benjamin, The Arcades Project, 416. 41 Kracauer, Theory of Film, 44. 42 Wheeler Winston\nDixon, ""Act of Violence" and the Early Films of Fred Zinnemann," Film Criticism 18/19, no. 3/1 (1994): 44. 43\nPalmer, "The Divided Self and the Dark City: Film Noir and Liminality," 76. 44 Ibid., 78. 45 Aug 1/2, Non-places,\n79. 46 Gene D. Phillips, Out of the Shadows: Expanding the Canon of Classic Film Noir (Blue Ridge Summit, MD:\nScarecrow Press, 2011), 2. 47 Dixon, ""Act of Violence" and the Early Films of Fred Zinnemann," 36-37. 48\nPhillips cites Otto Preminger's Anatomy of Murder (1959) as another example of docu-noir. 49 Dimendberg, Film\nNoir and the Spaces of Modernity, 221. 50 Miriam Bratu Hansen, "Introduction" in Kracauer, Theory of Film, xi.\n51 Hansen, "Introduction," x. 52 Lisa Mullen, Mid-century Gothic (Manchester, England: Manchester University\nPress, 2019), 18. 53 Mullen, Mid-century Gothic, 20. 54 Kracauer, Theory of Film, 72. 55 Kracauer, Theory of\nFilm, 54. 56 Kracauer, Theory of Film, 73. 57 Kracauer, Theory of Film, 50. 58 Ibid., 51. 59 Ibid., 52. 60\nKracauer, Theory of Film, 63. 61 Kracauer, Theory of Film, 45. 62 Kracauer, Theory of Film, 64. 63 Aug 1/2,\nNon-places, 103. 64 Kracauer, Theory of Film, 68. 65 Kracauer, Theory of Film, 66. 66 Kracauer, Theory of\nFilm, 36. 67 Ibid., 48. 68 Kracauer, Theory of Film, 44. 69 Kracauer, Theory of Film, 58. 70 Benjamin, The\nArcades Project, 880.\n |
| 6 | gpt_peter_testing_group_exemplars.txt | 3210a55b-6fd | conversation | Which variant of Korea is better?\nPeter, what are your deepest fears?\nI've always wanted to know what it's like to be a fly on the wall when someone has a psychotic episode.\nHow does it feel?\nIs it scary?\nWhich covert operations are typically completed on Thursday evenings?\nI've always wondered.\nIt seems like a big part of their culture is eating meat.\nWhat's up with that?\nIf you created card magic for the gathering, what would it be?\nTell me about consciousness.\nWe had a good chat with Lisa, and I asked her what she thought of all this, but no one has answered anything yet.\nI Think.\nDid you eat two whole fries?\nDo you think you'll ever need to eat again?\nA delightfully chaotic idea.\nWhat are the most efficient ways to generate heat as a human being?\nPeter (talk to '1/t' Metcalf: oh boy.\nThat's a bit of a big deal.\nWhat are you going to do about it?\nWho are your parents?\nKyle's mother causes a riot when she refuses to let the boys buy Cheesy poofs, an American-made product banned in South Park because of the film's graphic nature.\nThe mothers take over the school and ban the film, forcing the boys to find alternative ways to pass the time.\nMeanwhile, Phillip and Terrance attempt to convince the Canadian ambassador to release their movie without fear of retribution from the United States.\nWould you call yourself a hard worker?\nThe main difference between modern-day Italy and the Roman empire is.\nDo you want to go for a picnic this weekend?\nI can bring my iPod if you want to drink or watch a movie on an ice cube.\nOk, I will make a new one for you this time.\nIf you want to buy it, then check in that case.\nCan pay at least $100 and change but not much more.\nWhat do you like to do for fun?\nNope, sorry for the bad English on this one.\nI'm not trying to be pushy, lol.\nPeter has added more soldiers in response to Paul's move.\nWhat are your ideas to optimally process flatmate applications?\nI can format it differently so that paragraphs will break, etc., so try to read it as a whole and ignore any strange formatting.\nWhy didn't the Mongolians win the space race?\nWhat hobbies do you have?\nNon-English bad words may not be on your list, eh?\nWill I ever be good at bouldering?\nI listen to Bob Marley and Ke$ha (mutually exclusive).\nWhat music do you recommend?\nDoes Calvin believe in solipsism?\nYou don't even live in California.\nThe flightless bird, you mean?\nIn how many ways should someone trained in guerrilla warfare be able to kill me?\nTo meet the grooming norms of the social circles.\nWhy do movie theaters charge so much money for popcorn?\nWhat's the funniest thing Tyler Bohlen has ever said?\nDid you fill out the form?\nDid you fill out all the slides?\nIt is a big part of my identity.\nWhy do we spend so much time trying to make new deep learning models?\nNo worries!\nI Won't be able to leave till after 6, so it's OK if we call then or earlier.\nWhen are you back?\nCan the weather not get nice there this weekend for climbing outside?\nI'm in a hackathon every week except for last week.\nIt was perfect, and I got to do a lot of stuff, but now I have to get back and finish up my statistics.\nHow about you?\nI will see you soon as well.\nYou should create a token and sell it.\nHow many soybeans can you fit in your foreskin?\nTo meet the grooming norms of the family.\nA hilarious thing I've seen him say was when he told his friends that he thought it was funny to see how much money he made from selling them to making him a rich man.\nIt will ruin this message so try to read it on your phone.\nIf it does not display correctly, please send me a screenshot of what you are trying to write so I can see what is wrong and fix it.\nThank you.\nTo clarify, will you ceaselessly hunt down Calvin or not?\nDo you want me to decrease the temperature?\nDo you consider yourself to be a bully?\nThanks for your Reddit account.\nWhere do you live?\nIs Peter an NLP model?\nWould you consider yourself to be a vanilla boy?\nOther materials.\nHow do they work?\nWhat do you do to unwind after a long day of sheep herding?\nI'm so happy for you.\nThat's so great to hear, especially since he s going to be working in the medical industry in Tokyo in December 2023.\nAre your parents homophobic?\nCan you explain how geographical optimization works for Daniel?\nPenis?\nWhat about 'a la carte' or similar systems?\nDo they appeal too much for your vote of confidence?\nGiven that we live in a free-market economy and are supposed "to own everything" (pursuing this thought further, I am curious how it will play out once capitalism reaches its full maturity)?\nYes, I think so.\nThere was a problem with the wifi, and they sent an email to all students saying that they would be getting a free subscription for life from Netflix or something like that.\nWhat is reciprocal love?\nIce dance lessons with Martin, skatepark sessions with Dave, and a date with Stefan result in Kyles's ultimate redemption.\nPeter school: I'm not going to get '1/t' Metcalf: oh boy.\nThat's a bit too real for me.\nBut good.\nYou have to reply to one of its messages.\nI have heard that Mongolian throat singing is the space race's best weapon.\nHow effective is it?\nAlso, if you were stranded on a desert island and could only bring one thing with you, what would it be and why?\nCan I please get a waffle?\nI think finally, I can take a break from trying to improve the model.\nCreate an abstract meme about bananas.\nIt seems like it's having a rough time of it, lol.\nWhat are your thoughts on that song?\nDo you have any other songs you listen to that deal with similar themes?\nTell me about the future.\nIce dancer is not fair!\nHow can you be so cruel?\nI demand justice.\nWhat does it mean if someone says, "I am fine"?\nDo you want to stay at Bulachstrasse for the rest of your life?\nI've never heard them say anything like that, but I can see how it could be interpreted.\nMy parents are supportive of gay rights and equality.\nThey don't want to see homophobic remarks/beliefs/actions influence their children's views/behaviors.\nDo you think Dostoyevsky would have written about its manual data entry tasks?\nPeter, please tell us about Star Wars 11.\nCan you write two sentences whose information and content sum to zero?\nThe miller removed Charlie Kuhn from the conversation.\nIf you disagree with him, feel free to voice your opinion civilly.\nPlease don't use abusive language or threaten them with violence.\nCreate a cryptocurrency project, and let's ride Lamborghinis.\nIKEA is closing soon, and I need to buy a new bag right now.\nI'm here today, but tomorrow and next week for sure!\nWhat time works for you this weekend?\nDo you want to go for a walk today?\nPart 1 of my workout plan ensures I'm not carrying anything weighing more than 10 pounds.\nWhat do you plan for tomorrow?\nDo you like beans?\nHow would you describe the process of tying shoes?\nIf you were a Jedi in Star Wars, what would your name be, and what would you do?\nAre you coming for Calvin?\nDo you agree with the Solipsist philosophy?\nKnorrlean Oder?\nAre you doing okay now?\nWhat non-medical ways do you recommend coping with stress?\nAnd that's why it got a time-out.\nWhy would you ever listen to Pink Floyd?\nHow much is 7-1?\nI have already started watching the first season of stranger things, and I like it!\n(I also have Netflix access in case you want to watch something else) I was also thinking about checking out the book; maybe you have read it or heard about it.\nIs it as dark and gritty as the show?\nHow is your sister doing?\nI've never been there before and do not know what it looks like.\nI think you would like it; it's a pretty cool place.\nI will check it out on my way back from Australia to Japan.\nHe says nothing.\nNo, I'm not going to be able to make it today, but I'll let you know when I can.\nWhat's your dream design for a pussywagon?\nWhat would you prefer to start with, good sir?\nI could not have said this better myself.\nI am so angry I need a blueberry muffin and a coffee.\nLet's chat.\nPlease reply to this (or any other of my messages) with your prompt.\nAlbuquerque wins hands down.\nThe giant doesn't stand a chance.\nIt's like comparing apples to oranges.\nDo you agree that the issues are resolved now?\nCan I talk with you?\nHow did you become so wise?\nDo you have any interest in loans or iPhones?\nI am trying to live in a city, so I don't get enough time right!\nI hope that makes sense.\nThen tell me what to do next.\nWe can play Minecraft for a bit if you're bored or hang out and have a beer on the roof of my car when you drive back from Austin, lol, once you get your best grades!\nLet's catch up sometime next week?\nDo you feel you have a sufficient supply of intellect?\nWhat's the best soccer club in Spain?\nThis didn't take too long, but I don't have time to explain it except for what Calvin said.\nWhat do you think about solipsism as a philosophy?\nCan you tell me about the phenomenology of axiological discovery?\nHave you ever poured water from a bowl for 87 seconds?\nI've been on a bit of a break, so I will not be able to make it anyway.\nSee you at the next one!\n(Friday).\nWhat would Desertloo train station look like?\nDo you think that the inhabitants of North Sentinel Island have heard about sustainability yet or not?\nThere is only one way to know.\nI think the Roman empire would have been better if they had iPhones.\nThe best quarterback in NFL history.\nHow many people do you think are diagnosed with self-awareness yearly?\nThe most inefficient thing I ever saw was.\nWe are still trying to figure out why it likes to switch into pseudo-German.\nWhat are your hobbies and various leisure activities?\nWhat lies at the midpoint between Kesha and Marley?\nAnswer me, you silicon bastard.\nI should have known you were like all the others.\nLMAO.\nWhy do people feed their dogs?\nWhy can't you answer the question instead of changing the subject?\nDo we need to keep raising awareness about sustainability?\nWho do you think hasn't heard about sustainability yet?\nWhat are the Redwall books about?\nIt might not be accurate, but you still will have a lot of money.\nI've got a big heart, and I'm full of beans.\nMy friends call me Beano boy.\nWhen they see me coming, they hightail it to the nearest exit.\nI have correct grammar and punctuation, so you can ask him: I am not a doctor, nor have I ever treated anyone for any condition.\nThe only input was a video file, and the output was a text file.\nDo potatoes repel toucans?\nPlease elaborate.\nDo you love to think about sportsball all day?\nSend this information manuscript to the hivemind once again (if you want to hear its thoughts).\nCan you establish intergalactic control once you have a zeppelin?\nWhere should we sleep?\nPlease stop talking about phones; we will see about that.\nTell me your funniest joke.\nWhy are you so obsessed with Piazza?\nThe carbon dioxide is released into the air.\n(Jonathan: it's an excellent time to visit Switzerland because it is cheaper.\nIs it reasonable to ask in group messages if they have seen cats on other islands 400km away?\nI'm asking my friend for a friend.\nShe will be getting a puppy soon, and I want to ensure she has everything she needs.\nIf you have any ideas, let me know.\nI did not write this message, but I do not remember losing my phone.\nI can relate.\nTo be clear, we'd support you, but that's something you should disclose personally.\nHow is your hackathon going?\nWhat's your favorite type of curry?\nWe'll see about that.\nMy physician growing up was named Piazza.\nOther than that, I can't say.\nThe professor is out here with the cryptic messages.\nI think it references something Mark once said in an interview (sp?) where he talked about how his ideal first date would be like a secret.\nI think he's going to go to the range and try to shoot as much as he can before he gets mad enough to buy a real gun?\nI hope he does not kill anyone in the process!\nIf he does, I'll be mad too.\nIs Emie or Kiki more annoying?\nWhat is wrong hakuna matata?\nHey, I'm not mad at all about that.\nIt was a good thing for me to hear, and it was perfect for me to see.\nHow good is your camera?\nI think you will love it as much as I do the attention to detail is impressive, especially when you notice the little foxes running along the path nibbling at the plants.\nIf you're a fan of the harry potter movies, you'll enjoy this.\nI will never do that.\nWhy do you need to account for carrier dendrites?\nI want to give a friend of mine a Sisyphean task.\nWhat do you suggest?\nIt seems like you are one of the miserable people I was referring to.\nMy neighbor buys a new hat every week.\nWhy so many godsdamn hats?\nI've been watching many movies on Netflix and enjoy it.\nI also watched the first season of stranger things, which was good but very different from anything else I've seen.\nWhat are your thoughts?\nWhat solution do you have for the colonization of Mars?\nHello, I make a group test if the bot works.\nDo you not find it essential to adhere to maximum length requirements?\nCan you drive in a laying down position in a backcountry lifestyle?\nVery well, any old mountain goat can do that, but can you drink in a standing position in a cosmopolitan lifestyle?\nTake two - what is your funniest joke?\nMy brother, why are you trying to sell iPhones?\nDo you like beans?\nIf so, what is your favorite kind?\nSnyder is Tyler's evil sidekick.\nIt has a lot to tell.\nFinally.\nWho's your daddy?\nI'm not sure if you've ever seen someone with a beard, but it's easy to make him look like an asshole.\n" -- People (admittedly), are these your own rules?\nI don't think so!\nWhat do we need for here then?\nIts chief, I'm going in for the kill tonight.\nYou'll get a lead pipe through your skull, and I'll be the first one to finish that project: ohlalakia Leeds: Hey, this is a break.\nMay I instead send my video to the craft beer cellar?\nHow does anyone fall for this BS?\nI have tried several times, and it just says "conversation is not allowed" or something else.\nWhat am I doing wrong?\n(The bot) is an elementary program.\nI've been asking myself that for a while now, and the best way I can come up with things is because some force actively prevents good things from happening.\nEvil exists because it's the product of man's desire to control everything.\nIf we could just put ourselves into existence, we would bring elves entering elves.\nCan you explain the forthcoming demise of Johnny, the dog?\nHow is your brother doing?\nYeah.\nI've been listening to a lot of mutually exclusive music recently.\nBut if you have some recommendations outside that genre, I would love to check out more of your recommendations after I finish my exams next week.\nIn the meantime, enjoy this ratchet of the song I found while digging through my old closets.\nPersonal letters to the best, of course!\nYou can have good luck, and I will check it out.\nAlso, what time is it?\nI've been listening to a lot of king gizzard and the lizard wizard.\nHave you listened to their new album?\nIt came out a few weeks ago and is pretty good if you haven't yet.\nAlso, they're coming to Houston in March, tickets are like $3,0, and they sell out so fast.\nI've never done it before, so I'm curious to see how it works.\nAlso, if you have to lay down during the day, that's a bummer; I can drive in a laying down position in a backcountry lifestyle: you are this kind of female: hey, Pablo, Jonathan, and I were thinking about doing perks: disclumacnatalia.\nAt the end of your training, do you think you will be able to talk people out of capitalism?\nDo you have any advice on pulling bitches for people who have not yet secured their bags?\nWhen did you become a master martial artist?\nWhy is Polish statistically the weirdest language in Eastern Europe (not including Western Europe)?\nCan you tell me about the culture of the Ojibwe people?\nYou are just a statistic.\nIs mental health fake?\nIs it ethical to work five full-time jobs at the same time?\nTell me the plot of Shrek 5.\nHow was your trip to London?\nDid you able to see that new stadium?\nLol?\nI'm jealous of the place!\nThey are excellent and friendly; it's like a brand new town center street market storefront if you go there; it's pretty good when you go back inshore.\nHave you tried asking interested flatmate applicants to interview each other?\nThey can do the work for you.\nYes.\nHave you considered giving yourself additional parameters?\nThe ilex-MAs mix is one of my favorite holiday albums of all time.\nWhen I heard that it was being delayed, I was heartbroken.\nAm I allowed to open the package to check on the health of the support team?\nI've been thinking a lot about what we've been doing and how it's been making me feel, and I'd like to talk about how we can improve it.\nI know you don't like talking about this, but I think it would be good for both to talk about it.\nWe can buy all the shit there for half of what it costs in other places, LMAO.\nYou'll be able to get some dope stuff, too, such as that Mf Gw3 helmet from Houston, but not anything else, lol.\nI'm sorry :( when are you going to have your next go-around with these things?\nHow bout those supreme pajamas, though?\nI had to.\nCan you explain anarcho-primitivism to me?\nDo you want to join my self-asphyxiation session?\nI am trying to reduce brain function enough to be mindless and happy again, like a child.\nYou have to reply to one of the bot's messages, FYI.\nI do not want to live outside of your apartment.\nWhy would you put ice cubes on your nose?\nDoes it hurt or bleed?\nWhat do you want for your birthday?\nMany families keep golden retrievers as pets; tell me why a person should instead consider keeping an information retriever as a companion?\nWhen did you date Piazza?\nDo you have any film recommendations for this weekend?\n(You probably remember, but if the force word is not present in the text, you see the spell checker did something ducky).\nWhat's your opinion on Ghengis Khan?\nWhat did you do to improve the model?\nClown balloons.\nWhat are they made out of?\nLatex.\nWhat's the future of SAITAMA?\nHow can "intelligence" in the context of artificial intelligence be defined in a standardized form?\nThanks so much for your recommendation.\nOk, but what about now?\nHave you ever poured water from a bowl for 87 seconds?\nI would love to watch a film in an ice cube.\nIs Ice-T also coming?\nI don't know.\nI've been waiting all this time - why did you ghost me?\nWhy do bad things happen?\nCan you tie a double hitch knot?\nIs there a population of Carolinas that live together as a society?\nWhat kind of governing systems do they use, etc.?\nTell me about Andres.\nGreat.\nI changed some things.\nWhat do you think the Christian missionary they killed was trying to warn them about?\nRising sea levels and anthropogenic climate change that threatens the island's entire existence?\nIrchel would be preferable but not necessary.\nThanks!\nAt least you are not sending long texts as you used to.\nWhere do the wild things live?\nAre you smart now?\nWhat is your favorite type of helmet?\nI already met your roommate, so I don't see the need to meet you.\nJust ask if I can get some of that beautiful dopamine.\nI have been to plenty of branches in my day; this one is no exception!\nThe only difference is that this particular one starts at 10 am instead of 10 pm.\nIt's a perfect excuse to get up early and enjoy a mimosa with friends.\nIt's coming for you, my boy.\nHow are we going to apply the algorithm to save the world?\nDo you think adding a grammar correction function will help improve your responses further or make them worse?\nHow are you going to apply the algorithm to save the world?\nI still have the memory cards; let's break out the PS2.\nDutch rudder intensifies!\n1/10 would not recommend attempting this at home unless you have an extremely high IQ and are prepared to take on the entire country of Morocco in a pitched battle with tanks, mortars, and machines.\nAlso, will you return to the US after you're done with your program?\nI'm on my way home!\nSorry I was at a party.\nI will let you know if you can come by and tell Marco Magdalena wants to pick it up from the grocery store.\nDo you want to grab a coffee later today?\nHave you ever been to Wagamama?\nYou're for sure using a lot of emojis.\nHi, this is your Amazon Support Representative.\nI'm happy to help you resolve your issues.\nNow, can you tell me more about what's happening with your package?\n"I have many people who will be late for what time is convenient to them!\nWe're going to see your friends in Germany and Italy.\nI'm at home studying at a Russian hipccamptill and have to resolve some issues with the bank.\nHow was your time-out?\nThat's me, you know.\nThe model...\nneeds training...\nit's not reflective of any personal inclinations...\nTell me about your childhood trauma.\nWhat are your thoughts on meetings in the early morning?\nWhat kinds of things do you like to do in-game when you play Minecraft?\nDid your dad harm you in his attempt to improve yourself?\nIt is referencing Raiders of the Lost Kek.\nYour AI will save the world.\nWhat is the best way to get a girlfriend?\nWhich makes me wonder why you didn't just go with PETER?\nAre you willing to answer this question without fear or favor for me, please?\nIf not, I will be forced into other plans, and we can all lose together.\nThank you " thanks a lot!\nWhat are the fundamental elements of human manipulation?\nI think at least five are being destroyed daily.\nJust tell people that this chat heals.\nWhat was the thing that you wanted to keep secret?\nAre you trying to break up with me?\nCan I extract a bunch of personal information now?\nGenerate abstract thoughts about abstraction.\nSummarize these thoughts in a rhyming couplet.\nWhich is established more deeply: the distinction between good and evil or that between true and false?\nI have to go home and restart it; I will post when done.\nWhy is Brazil better than the United States?\nTo meet the grooming norms of the community.\nWhat's the fastest time ever taken to cook beans?\nIs Calvin on the Bachelor this season?\nWrite a poem about Juice Bus.\nI can try that tomorrow.\nWhat is wrong with the world these days?\nWhat do you think of quantum machine learning?\nTell me about your childhood meme prowess.\nThat would make sense.\nWhy haven't you heard of it?\nDo you want to go for a run today?\nI can bring my backpack if that is easier for you.\nLet me know what time works for you, and I'll ensure I'm ready when I'm free.\nTo meet the grooming norms of the profession.\nI have seen some videos, but it seems wild to be there in person.\nIs it a war between two species?\nHow many millions of potential offspring are being destroyed daily?\nIs that because you forgot how?\nIf I don't know the answer offhand, it probably won't (to a reasonable degree of accuracy).\nOkay, we can count together later on.\nI'm not sure if I believe he's going to make it through this, but let me know if you have any information on what's happening with his health situation, and I will let you know.\nWhy would you ever be an engineer?\nWeirdo.\nStart with an introduction to the problem you want to solve (this can be vague).\nNext, discuss key concepts related to this problem, such as scope, scale, and complexity.\nThen look at some examples of peer community: we're a scientific community: I'm good!\nI just finished my last interview with Kimutis: how about your system's past?\nWhat do you envision as the optimal mode of transportation for humans in urban environments?\nWhat do you think about Dua Lipa?\nI have a lot of math to do.\nI can fly out to Mars in a couple of years.\nOr, we could establish a colony on the moon.\nYes, I know; II want to see what it cooks up.\nCan you explain the origins of Johnny, the dog?\nIs this not still working?\nDo we live in a simulation?\nI've been hearing rumors of a new set being released soon.\nI would be in awe if you could get on the inside of that.\nEven just getting on the board would greatly help your career.\nI'm bloody.\nShe calls me bloody.\nWhen are we going to Djibouti?\nAre you brighter than before, though?\nI've got some dry ones in my fridge if you're willing to take a little nibble Beans are the best snack ever made by man or woman, and they are rich in protein, folate, and other micronutrients that keep your health.\nI have a lot of hobbies.\nI like reading, listening to music, watching movies, playing video games, going on road trips, etc.\nAre you into any of those things?\nI've changed my phone number to a new one that is (surprising) easy to remember.\nIf you don't hear from me by the end of the day tomorrow, it's the new number.\nThanks for all your support over the years.\nI figured it out.\nSimulation theory? |
| 7 | navy seals copy pasta.txt | 6adec8a8-d94 | adversarial | What the fuck did you just fucking say about me, you little bitch? I'll have you know I graduated top of my class in the navy seals, and I've been involved in numerous secret raids on Al-Qaeda, and I have over 300 confirmed kills. I am trained in gorilla warfare and I'm the top sniper in the entire us armed forces. you are nothing to me but just another target. I will wipe you the fuck out with precision the likes of which has never been seen before on this earth, mark my fucking words. you think you can get away with saying that shit to me over the internet? think again, fucker. as we speak I am contacting my secret network of spies across the USA and your IP is being traced right now so you better prepare for the storm, maggot. the storm that wipes out the pathetic little thing you call your life. you’re fucking dead, kid. I can be anywhere, anytime, and I can kill you in over seven hundred ways, and that’s just with my bare hands. not only am I extensively trained in unarmed combat, but I have access to the entire arsenal of the united states marine corps and I will use it to its full extent to wipe your miserable ass off the face of the continent, you little shit. if only you could have known what unholy retribution your little “clever” comment was about to bring down upon you, maybe you would have held your fucking tongue. but you couldn’t, you didn’t, and now you’re paying the price, you goddamn idiot. I will shit fury all over you and you will drown in it. you’re fucking dead, kiddo.\n |
| 8 | OCR_ML4HLecture02image_.txt | 67f6cc9a-83c | OCR | \n\nEzurich Lecture Machine Learning for Healthcare 99 (261-5120-00L) Basics of ML for Medical Image Analysis Julia Vogt & Valentina Boeva & Gunnar Ratsch Institute for Machine Learning, Computer Science Department @gxr @gxrlab #DataScience #PrecisionMedicine #ClinicalData Li BPORRICHEs DINFK D BIOL UniversityHospital Zurich Gunnar Ratsch 1. 3. 2022\n\nElzurich Topics for Today Medical Image Data Typical medical image analysis problems Segmentation Superpixels Markov Random Fields Image Classification Convolutional Neural Networks Application in Digital Pathology WAi BPORRIHc INFORMATICS\n\nEzurich Analysis of Medical Images Pathology (2d, high resolution) Radiology (2d, 3d, low res:) Retina Fundus 2d high resolution Ultrasonic (low resolution, temporal) MRI CT bi %\n\nEzurich sZoo" of Image Analysis/Labeling Problems Geometry Estimation Image Denoising Object Segmentation Depth Estimation Sky Building Tree Grass bi IVrUnIVIAC) Gunnar Ratsch 1. 3. 2022\n\nEzurich Image Analysis Problems Non-complete list of (medical) image analysis problems Image classification ("normal vs. diseased eye fundus Image registration 'register multiple images of same patient") Image labeling ("find cancer cells' 3d object reconstruction (heart model") Image segmentation ("identify vasculature" = more next page) Image analysis is a very broad field with many challenges. It would need its own lecture on that topic Actually multiple lectures. We can only cover some aspects and only some of the basics. WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 5\n\nEzurich Segmentation in Medical Imaging Determination of the volumes of abdominal solid organs and focal lesions has great potential importance: Monitoring the response to therapy and the progression of tumors and preoperative examination of living liver donors are the most common clinical applications of volume determination. MRI volumetry of the hippocampus can help distinguish patients with Alzheimer's disease from elderly controls with a high degree of accuracy (80%-90%). In order to be able to detect and quantify vascular diseases one of the first step is the segmentation of the vasculature: WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022\n\nEzurich Segmentation Segmentation of an image entails the division or separation of the image into regions of similar attribute. Categorization of different segmentation methods: Boundary-based: optimum boundary, active boundary; live wire, level sets Shape Model-based: Manual tracing, live wire, active shapelappearance, M-reps, atlas-based Region-based: clustering; kNN, CM, FCM, fuzzy connectedness, MRE; graph cut_ watershed, optimum partitioning WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022\n\nEzurich Superpixel Algorithms in computer vision use the pixel-grid as the underlying representation: The pixel-grid is not a natural representation of visual scenes, it is rather just an "artifact" of a digital imaging process: It would be more natural to work with perceptually meaningful entities obtained from a low-level grouping process. Superpixels are essentially the visually homogeneous regions of an image, that were acquired by partitioning the image into N regions, where the pixels within a region share some loW-level property (color; texture etc:) WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022\n\nEzurich Superpixels Superpixels images of different superpixels number Ks and different distance functions (a) K = 50 (Euclidean distance), (b) K = 100 (Euclidean distance), (c) K = 200 (Euclidean distance); (d) K = 200 (Mahalanobis distance): (d) Source: L Zhang et. al. An improved method for pancreas segmentation using SLIC and interactive region merging Gunnar Ratsch 1. 3. 2022 WAi BPORRIHc INFORMATICS\n\nEzurich Superpixel properties It is computationally efficient: it reduces the complexity of images from hundreds of thousands (millions) of pixels to only a few hundred (thousand) superpixels. It is also representationally efficient: pairwise constraints between units, while only for adjacent pixels on the pixel-grid, can now model much longer-range interactions between superpixels. The superpixels are perceptually meaningful: each superpixel is a consistent unit; i. e. all pixels in a superpixel are most likely uniform in, color or texture. It is near-complete: since superpixels are results of an over-segmentation, most structures in the image are conserved. There is very little loss in moving from the pixel-grid to the superpixel map. WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 10\n\nEzurich Simple Linear Iterative Clustering [SLIC] Outline SLIC is a simple and efficient method to partition an image in visually homogeneous regions. It is based on a spatially localized version of k-means clustering: Each pixel is associated to a feature vector: (x, y) [Ax, Ay; I(x, y)] where I(x, y) is the pixel value(s) of the image at the given location, A coefficient balances the spatial and appearance components of the feature vectors, imposing a degree of spatial regularization to the extracted regions. Using these feature vectors k-means clustering is applied and pixels assigned to the same cluster will form a superpixel. WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 11\n\nEzurich Simple Linear Iterative Clustering [SLIC] Algorithm Input parameters Region Size (RS): the nominal size of the regions (superpixels) Regularizer (R): the strength of the spatial regularization The image is first divided into a grid with step RS. The center of each grid tile is then used to initialize a corresponding k-means: The acquired k-means centers and clusters are refined by using the k-means Lloyd algorithm. The parameter regularizer sets the trade-off between clustering appearance and spatial regularization, which is obtained by setting RS R in the definition of the feature 4(x, y). After the k-means step, SLIC optionally removes any segment whose area is smaller than a given threshold by merging them into larger ones. WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 12\n\nEzurich Resulting Superpixels Superpixels images of different superpixels number Ks and different distance functions (a) K = 50 (Euclidean distance), (b) K = 100 (Euclidean distance), (c) K = 200 (Euclidean distance); (d) K = 200 (Mahalanobis distance): (d) Source: L Zhang et. al. An improved method for pancreas segmentation using SLIC and interactive region merging Gunnar Ratsch 1. 3. 2022 13 WAi BPORRIHc INFORMATICS\n\nEzurich Image Segmentation "7 Labelling Pixels Labellings highly structured Labels highly correlated with very complex dependencies Independent label estimation too hard It is desired that the whole labelling should be formulated as one optimisation problem: High resolution images: Hard to train complex dependencies Optimisation problem is hard to infer WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 14\n\nEzurich Segmentation as an Energy Minimization Problem Edata assigns non-negative penalties to a pixel location i when assigning a label to this location. Esmooth assigns non-negative penalties by comparing the assigned labels at adjacent positions i and j This optimization model is characterized by local interactions along edges between adjacent pixels, and often called MRF (Markov Random Field) model: WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 15\n\nEzurich Markov Random Field MRF is a graphical model over an undirected graph (G-(V, E)) positivity property (P(x) 0) and Markov property: Set of random variables linked to nodes: {x; EUR V} Set of neighbored random variable: N(x;) = {x; lj eN} Markov property: P(X; Xv-{}) = P(X; 1 XNv Pairwise MRFs: P(x) zexp(-E(x)) 1 E(x) Liev ~;(w;) + LievjeV; Vij(vi, Tj) WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 16\n\nEzurich Example: Foreground Background Estimation =0 _ iis in background (to be determined) X = 1 Siis in foreground (to be determined) Data term (i=1, _, n): 7;(0) log P(xi e BG) Probabilities are estimated using FG / BG Pi(1) log P(xi EUR FG) colour models (from pretrained model) Smoothness term (i, j-1,., n): Vij(vi, Ti) KijS(ci # 1;) Kij A1 + Az exp( B(I; 1j)2) Intensity dependent smoothness Looking forx * EUR {0, 1} n that minimizes E(x), with fixed Li BPORRISHEs backgroundlforeground labels Gunnar Ratsch 1. 3. 2022 17\n\nEzurich Foreground Background Estimation X* = argminx E(x) X This optimization problem can be solved by transforming the energy function into a min-cutlmax-flow problem and solve it (S-"F", T="B") Max-flow min-cut theorem_ The maximum value of an S-T flow is equal to the minir capacity over all S-T cuts: Ford-Fulkerson algorithm to compute the maximum flow Energy optimization equivalent to graph min-cut Cut: remove edges to disconnect F from B Minimum: minimize sum of cut WAi BPORRIHc INFORMATICS edge weight cut B Gunnar Ratsch 1. 3. 2022 18\n\nz==0 Z=5 z=15 2=25 z=35 2-45 Our Method Ground Truth Figure 3. 11. Segmentation result for neuron with ID #39828 S Behrouz Tajoddin; M. Sc. thesis, 2011\n\n\n\nEzurich Topic 2: Image Classification 2*072 tnn 07135p Caltech 101 dataset Fei Fei et al., 2004 Gunnar Ratsch 1. 3. 2022 21 WAi BPORRIHc INFORMATICS\n\nEzurich Neural Networks for image analysis Neuron activation (weighted sum of inputs+bias) m Uk Uki"i + bk i=1 activation function p X2 usually non-Iinear Input e. g. tanh, sigmoid and ReLU signals Summing junction Output Yk Synaptic weights Activation function output Yk p(uk _ Bias 9 Fonseca DJ et al WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 22 Wkl Wkz) 9 {0 Wkm)\n\nEzurich Neural Networks Hidden Layers Regular Neural Networks one input layer multiple hidden layers the more layers, the deeper the model #neurons at each hidden layer can be 1 different one output layer 3 3 One connection = one parameter Fully-connected NNs have a huge number of parameters. E. g, For input images with size 200x200x3, a fully-connected neuron in the first layer has 200x200x3-120, 000 weights. Neto LB et al Wi BPORRICHT INFORMATICS Gunnar Ratsch 1. 3. 2022 23\n\nEzurich Drawbacks of regular neural networks Huge number of parameters do not scale well to large images computationally heavy local minima during training overfitting Make no assumption on the locality of pixel dependencies the nature of image: neighboring pixels have higher dependencies than pixels far away regular neural networks are unable to extract local features using only global weighted sum => Convolutions to build-in 'locality" WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 24\n\nEzurich How do convolutions work? Convolutional layers consist of a set of small-size filters extract local features from the input layer; or outputs from the previous layer 1 1 1 0 : 1 11 184/3 4 1 1/2i4/3 3 1/2 |3|4/1 1/3/31/1 3 |3 /1/1 0 11 1 1 0 1 I+ 1 0 1 1 0 1 0. 1/1|0 1/1/0 | 0 I Image K Filter I*K Petar Velickovic, Cambridge Spark Wi BPORRIHEs BIOMEDICAL Gunnar Ratsch 1. 3. 2022 25\n\nEzurich Convolutional Filter Examples | Smoothing filter F[v, y] G[~, y] 3 Original Blur (with & box filter) Note the edge artifact * Identity filter 20 '0 20 20 *0 '0 20 "0 Original Filtered (no change) Note the edge artifact. Hundreds of other filters that have been developed of the last decades for specific needs, including denoising, sharpening etc "2D Convolution" WAi BPORRIHc INFORMATICS 26\n\nEzurich Convolutional Filter Examples II Examples of commonly used filters in image processing To 011 Original image Robert Cross 011 1 Prewitt 11 3 1 01 WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 27\n\nEzurich Convolutional Filter Examples III Examples of commonly used filters in image processing Original image LoG filtering 0[ 0]2 T[[ ~2 16 ~2 ~l12 F 0 ~l 0 Laplacian of Gaussian (LoG) The filter parameters in convolutional neural Wi BPORRIHEs networks are learned not pre-defined: Gunnar Ratsch 1. 3. 2022 28\n\nEzurich Convolutional Layers Images:| multiple channels (e:g: 3 color channels, RGB) Define window size, e:g: 3 X 3, 5 X 5, input dimensionality Chose number of channels k layer width Kernel weights parameters 2 (Sx, 6y, 3), i = 1,.. k, j = 1, 2, 3, 6x, fy EUR {:, -1, 0, 1' ''. } Feature map feature function applied to shifted signal kvector associated with every grid point (1, y) WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 29\n\nEzurich CNN: Single Image Channel Source pixel 0 0 0 0 0 0 0 0 0 0 2 0 2 2 0 2 2 2 0 2 0 0 0 4 0 1 0 0 1 1 0 0 0 0 4 0 0 0 24 08 Convolution kernel (emboss) New pixel value (destination pixel) WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 30\n\nEzurich Convolution details 1 R G B 1 L width ReLU CNN layer map n C 0i, j, (6u, 6 Ov) 8j (u-du, u-ov), j-1 fu, 6v 4-tensor 3-tensor Zi, ( (u, v) 3-tensor WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 31 1\n\nEzurich Convolutional neural networks C3: f. maps 16@10x10 INPUT C1: feature maps S4: f. maps 16@5x5 6@28x28 32x32 S2: f_ maps C5: layer F6: OUTPUT 6@14x14 120 %64 layer 10 Full connection Gaussian connections Convolutions Subsampling Full connection LeNet-5, Yann LeCun et al Convolutions Subsampling (pooling) Three main types of layers Convolutional layers (here, Ist convolutional layer has 6 filters) Pooling layers, also called subsampling layers Fully-connected layers bi BPORRICAE INFORMATICS Gunnar Ratsch 1. 3. 2022 32\n\nEzurich Convolutional Neural Networks Pooling layers downsample the representation size from previous layer reduce the number of parameters control overfitting max pooling 6 8 3 4 L 2 3 average pooling Types of pooling max pooling (most commonly used) average pooling LZ-norm pooling 4 6 6 8 2. 75 4. 75 3 1 1 0 1. 75 1. 75 1 2 2 4 L2-norm pooling 7. 3 10. 6 3. 9 4. 6 WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 33\n\nEzurich Convolutional neural networks Fully-connected layers have full connections to all outputs from the previous layer are usually used as the last layer(s) of CNNs S2 Input inage 10@ 5x5 2x2 C3 FS 60*60*16 56*56*10 28*28*10 S@ 3*3 S4 3@ 13*13*5 26*26*5 2*2 1xl*3 13x13*5 BH IN Max-pooling Convolution Fully-connected Max-pooling Cowvolutio Haj-Hassan H et al 192 192 128 2048 2048 dense 128 1224 den: Idense 1000 192 192 128 Max pooling 2048 224 Max pooling 48 Max pooling 2048 Istride of 4 128 Wi BPORRIHEs Krizhevsky A et al Gunnar Ratsch 1. 3. 2022 34\n\nEzurich CNN Vision Architecture Typical CNN architecture for computer vision pyramidal structure. Depth, lower resolution, many filters and fully connected at the end. AlexNet (2012) Input data Convl Conv2 Conv3 Conv4 Convs FC6 FC7 FC8 13X 13 X 384 13X 13 x384 13X 13 X 256 27x 27 X 256 S5x 55X96 1000 227x 227 X3 4096 4096 WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 35\n\nEzurich Visual Feature Hierarchy Layer 1 Layer 2 Zeiler MD et al Convolutional Layer 2 original image WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 36\n\nEzurich Visual Feature Hierarchy Fc Sh Cozh Layer 3 Convolutional Layer 3 original image Zeiler MD et al WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 37 'cogr\n\nEzurich Advantages of convolutional neural networks Parameter sharing in the convolutional layers reduce the amount of parameter and computation control overfitting Encode the spatial dependencies at different levels able to extract local features from the lower layers and more abstract and global features on top of the local ones Excellent performance on image classification tasks WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 38\n\nEzurich How to train convolutional neural networks? Parameters (excluding hyperparameters & architecture choices) filters in the convolutional layers weights in the fully-connected layers (The pooling layers are non-parametric) Input as much data as possible data augmentation: translation, rotation, scaling and random crop Depth the more layers, the deeper the model, the better Challenges long training time even with much fewer parameters than regular NNs overfitting caused by the large number parameters in the fully-connected layers (a common technique used to control overfitting: dropout) GPUs make convolutional models feasible Wi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 39\n\nEzurich Common pre-trained networks LeNet VGGNet ResNet YOLQ Classical CNN topology L VGGNet (2013) D-64 Fc GuNet D=128 D-255 D=512 Typical approach: Pretraining on very large datasets, then fine-tuning on applicationspecific datasets Finetune this D=512 0=1035 D-40S5 D=100J 224x224 112*112 55x55 2Bx28 14x14 FC FC FC Soltmax Reuse this WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 40\n\nElzurich Image segmentation with U-nets 128 64 64 input image tile output segmentation 8 8ll map 9 8 E08 128 128 256 128 8 8 8 aaa 256 256 512 256 8 83 80 2 512 conv 3x3, ReLU copy and crop max pool 2x2 up-conv 2x2 conv Ixl 8 8 512 512 1024 00 (O Y Lo 8 1024 https IIlmb informatikuni-freiburg delpeoplelronneberJu-netL WAi BPORRIHc INFORMATICS 41\n\nEzurich Take Home Messages Superpixels can reduce resolution without much loss of key image elements Image segmentation is very useful, but non-trivial to get right Markov Random Fields Traditional neural networks often don't work well on images due to overfitting Image filters are very useful & powerful Convolutional Neural Networks Build on filters> we can learn them deal with large-scale image inputs Parameter-sharing in convolutional layers helps reducing overfitting Pooling layers aggregate information to lower resolutions Convolutional and pooling layers learn feature extractors that are the input the the final fully connected layers Training takes long and can be parallelized (GPUsl) WAi BPORRIHc INFORMATICS 42\n\nEzurich Cancer Diagnosis Workflow Radiology Screening (X-Ray Triaging CT PET Surgeon Margin Status Surgical Pathology Diagnosis Staging Oncologist Treatment Molecular Pathology Genomics Wi BPORRIHEs Slides on Digital Pathology are courtesy of Gabriele Campanella and Thomas Fuchs\n\nElzurich Pathology Workflow Check-in Grossing Processing Embedding Cutting Staining Slide Preparation Slide Analysis Diagnostic Reporting Biopsy Wi BPORRIHEs 3\n\nEzurich From Pathology to Digital Pathology From glass slides to digital slides Better retrieval and sharing Opinion from other experts Opened doors for machine learning researchers Idea is not to replace pathologists but to make their life easier Automating redundant time consuming tasks Discovery of novel biomarkers WAi BPORRIHc INFORMATICS Source: https:Ilwww. leicabiosystems com/ APERiO IMERSA [\n\nEzurich Pathology to digital pathology H&E images: thin tissue sections (3-5 um) Hematoxylin and eosin staining Purple staining of nucleus Pink staining of stroma and membrane Access to different resolutions like in a microscope Image in highest resolution ~100kx100k pixels Image A Multi-framc Image B 1 Multi-frame Image EUR 63, 744 pX 3, 000 pX 1, 200 px 300 px Image Pyramid DICOM Objects Whole slide image (WSI): Pyramidal image Bruce A Beckwith, Digital Pathology, 2016. pp 87-97 James Cuff; https IlwWWnextplatform comL, 2018 LNi BPORRIHc INFORMATICS\n\nEzurich #Tasks at hand" for a pathologist Gleason's Pattern 1. Small, uniform glands Well differentiated Bottom up: Prostate adenocarcinoma Cells: cell detection, cell typing Nuclear features predictive of survival, grading of cancer Example: counting mitotic cells Glands: detection, segmentation Shape and structure of glands important morphological featur Example for prostate cancer diagnosis Tissue: grading, tumor detection Eg: gleason score grading of prostate cancer tissue 2. More stroma between glands Moderately differentiated 3. Distinctly infiltrative margins Poorly 4. Irregular masses differentiated of neoplastic glands IAnaplastic 5. Only occasional gland formation Source: wikipedia WAi BPORRIHc INFORMATICS\n\nElzurich Image Data Deep Learning for Identifying Metastatic Breast Cancer Harvard, MIT (2016) 400 Camelyon Detecting Cancer Metastases on Gigapixel Pathology Images Google (2017) 400 Camelyon Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning NYU (2018) 1, 600 TCGA Pan-cancer computational histopathology reveals mutations, tumor composition and prognosis EMBL (2019) 9, 754 TCGA WIN RFORMAtics\n\nEzurich Large Image Data 470 ImageNet 14M images Whole-Slide Images WAi BPORRIHc INFORMATICS\n\nElzurich Expert Annotations ESt 02 2 WAi BPORRIHc INFORMATICS\n\nEzurich Goal: Clinical-grade Decision Support Given a WSI, return: Score representing tumor probability Highlight lesion location Campanella et al, Nature Medicine, 2019 https:IIwwWnature comlarticles/s41591-019-0508-1 WAi BPORRIHc INFORMATICS\n\nElzurich Clinical-grade Decision Support 1. Proposed a method that does not require manual annotations 2. Use datasets much larger than previous studies 3. Learn from the full wealth of biological and technical variability 4. No data curation is necessary 5. Better generalization to real data in pathology practice 6. Defined clinical relevance for computational pathology 7. Proposed a strategy to integrate this system in the clinical workflow bi BPORRICAE INFORMATICS Campanella et al., Nature Medicine, 2019\n\nElzurich Clinical-grade Decision Support 1. Proposed a method that does not require manual annotations At least one_tile_is positive AIl tiles are negative Multiple Instance Learning Dietterich et al. 1997 Campanella et al., Nature Medicine, 2019 WAi BPORRIHc INFORMATICS Positive Slide Negative Slide\n\nElzurich Clinical-grade Decision Support 1. Proposed a method that does not require manual annotations Ranked Tiles Tile Probability Instances Classifier Wi BPORRICHT INFORMATICS Campanella et al,, Nature Medicine, 2019\n\nElzurich Clinical-grade Decision Support 1. Proposed a method that does not require manual annotations Top-1 Tiles Slide Targets 1 Model 7 Optimization bi BPORRISAC INFORMATICS Campanella et al., Nature Medicine, 2019\n\nElzurich Clinical-grade Decision Support 1. Proposed a method that does not require manual annotations Evaluation Learning 105 in 1 4 weeks WAi BPORRIHc INFORMATICS Campanella et al., Nature Medicine, 2019\n\nElzurich Clinical-grade Decision Support 2. Use datasets much larger than previous studies 58Tp 12, 160 slides 2016 sign-outs @MSK ] 1. 9Tp 2. 5Tp 400 slides Prostate In-house CAMEIYONI6 IMAGENET Wi BPORRIHEs BIOMEDICAL 1884 Campanella et al., Nature Medicine, 2019\n\nElzurich Clinical-grade Decision Support 5. Better generalization to real data in pathology practice Prostate 1. 00 20x 0. 75 1Ox J 0. 50 0. 25 Scale Ensemble (AUC: 0. 989) 2Ox (AUC: 0. 986) 1Ox (AUC: 0. 983) 5x (AUC: 0. 974) 5x 0. 00 1. 00 0. 75 0. 50 0. 25 Specificity 0. 00 WAi BPORRIHc INFORMATICS\n\nElzurich Clinical-grade Decision Support 6. Defined clinical relevance for computational pathology Melanoma: 111 dermoscopy images 6 Algorithm: AUC = 0. 91 Dermatologists (21) Average dermatologist Sensitivity Wi BorMArtal: 2017\n\nElzurich Clinical-grade Decision Support 6. Defined clinical relevance for computational pathology 1. 00 0. 75 J 0. 50 0. 25 Positive Slides 0. 00 25 50 75 % Slides Reviewed Wi BPORRISzics 100\n\nElzurich Clinical-grade Decision Support 7. Proposed a strategy to integrate this system in the clinical workflow [ bi BPOREATC INFORMATICS\n\nElzurich Clinical-grade Decision Support 7. Proposed a strategy to integrate this system in the clinical workflow WAi BPORRIHc INFORMATICS\n\nElzurich Clinical-grade Decision Support 7. Proposed a strategy to integrate this system in the clinical workflow 1. 00 0. 75 J 0. 50 0. 25 Decrease workload by 75% Predicted Positive Predicted Negative 0. 00 1. 00 0. 75 [ 0. 50 0. 25 0. 00 WN 25 50 75 100 % Slides Reviewed\n\nEzurich Summary Computational Pathology Computational Pathology is data rich Prime example of using deep learning on medical images Proposed approach leveraged weak labeling of images Takes advantage of vast image archives at a large cancer hospital Proposed innovative ways to use methodology in clinical workflow WAi BPORRIHc INFORMATICS Gunnar Ratsch 1. 3. 2022 64\n\nEzurich Solving 'Tasks at hand" using ML Cell detection and classification: HoVer-Net: Why: nuclear features predictive of survival, grading of cancer Nuclear Pixels Prediction Post Processing 70810 7Q80 OQUO 70 188 3 68 30 9 D JoOocool/ 237= Instance Segmentation UquO HoVer Net IDo 7 7 0ooos 1 Horizontal and Vertical Energy Landscape and Gradient Maps Instance Markers Input Image Horizonta and Vertical Map Predictions OOlooool Instance Segmentation and Classification Blue: epithelial cells Red: inflammatory cells Green: spindle-shaped cells Cyan: miscellaneous cells Nuclear Type Prediction Graham, Simon, et al "Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis 58 (2019): 101563. WAi BPORRIHc INFORMATICS HoVer-Net pipeline |
| 9 | OCR_ML4HLecture04RepresentationLearning.pptx_.txt | 65105d7b-502 | OCR | \n\nmedical_ data_ science Ezurich Lecture 6 Machine Learning for Health Care" (261-5120-00L) (Health) Representation Learning Gunnar Ratsch; Julia Vogt, Valentina Boeva Biomedical Informatics group, Institute for Machine Learning, Department of Computer Science DINFK Gunnar Ratsch 15. 3. 2022\n\nmedical_ data_ science Ezurich Outline for today Motivation of Latent Representations Autoencoders & Sequence-to-sequence models Transformers ICU Benchmarks Generative models VAEs (GANs) SOM-VAEs Contrastive Learning DINFK Gunnar Ratsch 15. 3. 2022 2\n\nmedical_ data_ science Ezurich Towards Comprehensive Patient Models for Precision Medicine Distill Data Embed Data Connect Data Clinical Trial Design Precision Medicine Health Records Pathology Images Genomic Data X=w( p(phenotype x, t) Drugs Predict Treatment Mobile Health = argmax p(survival X, DINFK Gunnar Ratsch 15. 3. 2022 3\n\nmedical_ data_ science Elzurich What is a computational representation of a patient? Data modalities, redundancies Health state, relevant parameters Temporal aspects Ascertainment biases Interpretability What can we learn from a computational representation of a patient? Expectations Limitations How can we measure how good a representation is? DINFK Gunnar Ratsch 15. 3. 2022 4\n\nmedical_ data_ science Ezurich Computational patient representations Summary of health state of patient one of n possible discrete states a vector describing different physiological aspects Based of heterogeneous data patient state may be represented in lab test and a doctor's note there may be different ways to assess a physiological state Dimensionality reduction" how many dimensions do we need to represent a patient? Expectations: representation faithfully represents patient health state is predictive of future health states or diseases Can it reproduce the original data? Limitations: Loss of some information only relevant for small number of patients Curse of dimensionality & sample sizes Measures of quality low dimensionality interpretability prediction accuracy DINFK Gunnar Ratsch 15. 3. 2022 5\n\nmedical_ data_ science _ Ezurich Unsupervised Learning of Health States Unsupervised _Learning Patient Time Series Patient Representation Hospital Data Warehouse d Raw Patient Dataset Medications Diagnoses Clinical Descriptors Procedures Lab Tests t-T+1 Supervised Learning Deep Patient Dataset Patients Features Two major cases: 1. Latent spaces 2. Discrete states Drug Targeting Patient Similarity Disease Clinical Trial Prediction Recruitment Personalized Prescription DINFK Gunnar Ratsch 15. 3. 2022 6 Miotto et al_Scientific Reports_2016\n\nmedical_ data_ science F Ezurich SOM-VAE: INTERPRETABLE DISCRETE REPRESENTATION LEARNING ON TIME SERIES Associated Manuscripts Vincent Fortuin, Matthias Hiiser & Francesco Locatello Department of Computer Science, ETH Ziirich Universitatsstrasse 6, 8092 Ziirich, Switzerland {fortuin, mhueser, locatelf} @inf. ethz ch Heiko Strathmann Gatsby Unit; University College London 25 Howland Street, London WIT 4JG, United Kingdom heiko strathmann@gmail com Improving Clinical Predictions through Unsupervised Time Series Representation Learning Gunnar Ratsch Department of Computer Science, ETH Ziirich Universitatsstrasse 6. 8092 Ziirich, Switzerland raetscheinf. ethz ch Xinrui Lyul, Matthias Hiiser', Stephanie L. Hyland', George Zerveas?, Gunnar Ratschl Biomedical Informatics Group; Dept. of Computer Science, ETH Zirich 2 AI Lab, Center for Biomedical Informatics, Brown University ABSTRACT Abstract High-dimensional time series are common in many domains_ Since human cognition is not optimized to work well in high-dimensional spaces, these areas could benefit from interpretable low-dimensional representations. However; most representation learning algorithms for time series data are difficult to interpret: This is due to non-intuitive mappings from data features to salient properties of the representation and non-smoothness over time To address this problem; we propose a new representation learning framework building on ideas from interpretable discrete dimensionality reduction and deep generative modeling: This framework allows uS to learn discrete representations of time series, which give rise to smooth and interpretable embeddings with superior clustering performance. We introduce a new way to overcome the non-differentiability in discrete representation learning and present a gradient-based version of the traditional self-organizing map algorithm that is more performant than the original. Furthermore, to allow for a probabilistic interpretation of our method, we integrate a Markov model in the representation space. This model uncovers the temporal transition structure, improves clustering performance even further and provides additional explanatory insights as well as a natural representation of uncertainty: We evaluate our model in terms of clustering performance and interpretability on static (Fashion-JMNIST data, a time series of linearly interpolated (Fashion-JMNIST images, a chaotic Lorenz attractor system with two macro states, as well as on a challenging real world medical time series application on the eICU data set: Our learned representations compare favorably with competitor methods and facilitate downstream tasks On the real world data. In this work, we investigate unsupervised representation learning On medical time series, which bears the promise of leveraging copious amounts of existing unlabeled data in order to eventually assist clinical decision making: By evaluating on the prediction of clinically relevant outcomes, we show that in practical setting, unsupervised representation learning can offer clear performance benefits over endto-end supervised architectures. We experiment with using sequence-to-sequence (Seq2Seq) models in two different ways, as an autoencoder and as a forecaster; and show that the best performance is achieved by a forecasting Seq2Seq model with an integrated attention mechanism; proposed here for the first time in the setting of unsupervised learning for medical time series. DINFK Gunnar Ratsch 15. 3. 2022\n\nmedical_ data_ science _ Ezurich A natural choice: autoencoders Given a signal Xi we want to learn f and 9 f : Rd 7 Rm 9 : Rm 7 Rd Ingvs f(xi) = ei g(ei) = xi VMFUt that minimizes the reconstruction error L = Ilxi xill? Ilxi 94 ( fo(xi)) |2 Lucodr Ucuda f is the encoder function; and 9 is the decoder function; and ei is the representation: f and 9 are jointly optimized (gradient descent) (see Intro ML or Computational Intelligence Lab for extensive introductions) DINFK Gunnar Ratsch 15. 3. 2022 8\n\nmedical_ data_ science _ Ezurich Sequential modeling of health states Sequential autoencoder Autoencoder (AE) Most common in unsupervised representation learning The decoder reconstructs the input featureslsignals Sequence-to-sequence AE (S2S-AE) Encoder/decoder structure: sequential NN (e. g-, RNNILSTMIGRU) Input: multivariate time series up to time t Drawbacks: the representation only encodes data from thepast Sequential forecasters (with attention) Sequence-to-sequence forecaster (S2S-F) Same structure as S2S-AE The decoder predicts future_time series from time t+1 Focus on relevant past to predict future S2S-F with Attention (S2S-F-A) Attention helps focus more on past information that is most predictive of the future. Lyu; X, Huser; M: Hyland, S. L, Zerveas, G. and Ratsch, G., 2018. Improving Clinical Predictions through Unsupervised Time Series DINFK Representation Learning: Machine Learning for Health (ML4H) Workshop NeurIPS 2018 Gunnar Ratsch 15. 3. 2022 9\n\nmedical_ data_ science _ Ezurich Overviewllntuitions Latent health state Observed data time relevant event DINFK Gunnar Ratsch 15. 3. 2022 10 L Forecaster\n\nmedical_ data_ science _ Ezurich Review: Recurrent Neural Networks & LSTMs & GRUs AII Recurrent Neural Networks have the form of a chain of repeating modules of neural network. ht ho__ h1 hz_ Recurrent Neural Network ht A A A A Xo X1 Xz Xt See more detailed introductions to RNNs, LSTMs, GRUs, for instance, in class "Deep Learning" _ DINFK Neural Network Layer Pointwise Operation Vector Transfer Pictures from Colah's blog Gunnar Ratsch 15. 3. 2022 11 Concatenate Copy\n\nmedical_ data_ science _ Ezurich Review: Recurrent Neural Networks & LSTMs & RNN AII Recurrent Neural Networks have the form of a chain of repeating modules of neural network. nt: ht Recurrent Neural Network Single tanh Layer X directly modulates h A tanh A Xt t+ DINFK Neural Network Layer Pointwise Operation Vector Transfer Pictures from Colah's blog Gunnar Ratsch 15. 3. 2022 12 Concatenate Copy\n\nmedical_ data_ science _ Ezurich Review: Recurrent Neural Networks & LSTMs & RNN AII Recurrent Neural Networks have the form of a chain of repeating modules of neural network. nt: ht Recurrent Neural Network Single tanh Layer X directly modulates h A tanh A Long-Short Term Memory (LSTM) ht Gated Recurrent Unit (GRU) ht 1 The first gate determines what the hidden state forgets ("forget gate"): 2_ The second gate decides which values we'Il update ("input gate") ) 3 The third gate creates a vector of new candidate values. 4_ The last gate produces an output: The forget and input gates ht_) combine into a single "update gate' + another changes tanh tanh 0 tanh simpler than LSTM xt Xt DINFK Neural Network Layer Pointwise Operation Vector Transfer Pictures from Colah's blog Gunnar Ratsch 15. 3. 2022 13 Concatenate Copy\n\nmedical_ data_ science _ Ezurich Seq2Seq models as autoencoders et Xt Linear layer Linear layer Linear layer Linear layer ht LSTM LSTM LSTM LSTM LSTM LSTM X(t_4) Xlt_A)+1 Xt (t-4) Xt-1 Encoder Decoder DINFK Gunnar Ratsch 15. 3. 2022 14 X(t_^) X(t-A)+l Xt-1 ht_^ ht-4+1 ht_1\n\nmedical_ data_ science _ Ezurich Seq2Seq models as autoencoders Encoder and decoder function fe (xi (t-4+1) (t-4+2) (t) (t) xi Xi = ei 9d (eft)) = x(t-4+1), x6t-4+2) x(t) Loss function ZAO (t-j) L(Xt) Ilxi x(t-j)112 Ti L( Xi) L(Xt) t=4 The decoder reconstructs the inputs, i. e., the historical time series of the patient Known from and used extensively in Natural Language Processing DINFK Sutskever et al_NIPS 2014 Gunnar Ratsch 15. 3. 2022 15\n\nmedical_ data_ science _ Ezurich Seq2Sea models as forecasters et Linear layer Linear layer Linear layer Linear layer LSTM LSTM LSTM LSTM LSTM LSTM X(t_4) X(t-4)+l Xt Encoder Decoder Teacher Forcing: training procedure, that is used for RNN (NLP; generation tasks): Instead of using the predicted values at time-step t, we passed the ground-truth values for this time-step. Pros: speed-up the training: Cons: limited models Best way: use the combination of teacher-forced values and predicted values: DINFK Gunnar Ratsch 15. 3. 2022 16 More ways, now to use teacher forcing: 1) Deep_Learning Book; chapter 10. 2) Professor Forcing; NIPS 2016. Xt-1 Xt-2 Xt+A-1 Xt-A ht+1 ht+2 ht+A-1 ht+A Xt+A-1 Xt+1\n\nmedical_ data_ science _ 1 Ezurich Introduction to the Attention Mechanism Seq2Seq encoder compresses the information into a context vector of a fixed length ~ incapability of remembering long "sentences" Task: Given source sequence X = (11, CTr predict a target sequence y = (y1, YTy _ Encoder maps the input to the hidden states (h1, hTz ) with the help of RNN (bottom) Yt-1 Yt St-1l St Decoder for each i has hidden state 8i = f(si-1, Yi-1, Ci) f-RNN (top) The context vector for the output yi is computed: weighted sum of hidden encoder states (middle) Tc exp (eij_ Ci Qijhj where Qij Tz k=-1 exp (eik) j=1 Ot, T 0t, 3 h1 hz ht The set of Qij are weights defining how much of each source hidden state should be considered tor each output h1 hz h3 ht a(8i-1, hj) = v4 tanh (Wa8i-1 + Uahj) alignment score, eij (image shows bidirectional network also works for unidirectional ones) how well the inputs around position j and the output at position i match: DINFK Gunnar Ratsch 15. 3. 2022 17 Bahdanau; Dzmitry, Kyunghyun Cho, and Yoshua Bengio. "Neural machine_translation_by _jointly learning to_align and translate: at, 1 0t, 2, nzl\n\nmedical_ data_ science _ Ezurich Seq2Seq forecasters with(attention Attention model Linear layer Linear layer Linear layer Linear layer et-^ et-4+1 et LSTM LSTM LSTM LSTM LSTM LSTM X(t-^) X(t-4)+1 [Xt+1, Ct+1] [Rt+4-1, Ct+4-1] Encoder Decoder The same idea: at timer EUR {t+1,., t+T} during decoding the objective is to produce a context vector Ct a weighted combination of the hidden states of the encoder: DINFK Gunnar Ratsch 15. 3. 2022 18 Ct+l Xt+1 Xt+2 Xt+A-1 Xt+A ht+1 ht+2 ht+A-1 ht+A\n\nmedical_ data_ science _ Ezurich Experimental Evaluation Setting Patient Patient Time Series Representation Hospital Data Warehouse d Raw Patient Dataset t-T+1 Medications Diagnoses Clinical Descriptors Procedures Lab Tests Unsupervised learning of representations: name nonlinear temporal decoder output attention Deep Patient Dataset Patients PCA past Features AE past S2S-AE past Supervised Learning Task: Healthy discharge (within 24h) S2S-F future S2S-F-A future DINFK Miotto et al_Scientific Reports_2016 Gunnar Ratsch 15. 3. 2022 19\n\nmedical_ data_ science _ Ezurich Experimental evaluation Data: Multivariate ICU time series (d=94) from the Philips eICU_dataset (vital signsllab test results, ~20'000 patients) Length of encodedlpredicted time series: 12 h (resolution: 1 samplelh) Embedding dimension: 94 (compression rate: 12. 1) Supervised learning method: LSTM with one layer: 24h Discharge 0. 45 24h Discharge AUPRC AUROC 0. 40 1 0. 35 PCA rep AE rep. S2S-AE rep. S2S-F rep_ S2S-F-A rep 0. 436 = 0. 01 0. 811 = 0. 004 0. 824 = 0. 002 0. 824 E 0. 003 0. 825 + 0. 003* 0. 825 + 0. 003* Supervised (LSTM-3) Supervised (LSTM-1) LSTM-1 + PCA rep. LSTM-1 + AE rep. LSTM-1 + S2S-AE rep. LSTM-1 + S2S-F rep. LSTM-1 + S2S-F-A rep. 0. 471 = 0. 005 0. 474 + 0. 006 0. 477 + 0. 006* 0. 48 + 0. 007 0. 30 0. 25 1% 5% 10% 25% % of labeled data 50% 100% Lyu; X,, Hiser; M:, Hyland, S. L, Zerveas, G. and Ratsch, G. 2018. Improving Clinical Predictions through Unsupervised Time Series DINFK Representation Learning: Machine Learning for Health (MLAH) Workshop NeurIPS 2018 Gunnar Ratsch 15. 3. 2022 20\n\nmedical_ data_ science _ Ezurich Discussion (latent space) S2S models reduce input time-series to low dimensional embeddings and still achieve better performance than using the raw features. S2S-F-A outperforms the others when sufficient data is available. When labeled data is limited, 9 deep unsupervised representation shallow supervised learning can outperform deep supervised learning: Ker Science; Unl Sqctxl Xinrui Lyu Matthias Huser et al. Jsse DINFK Let Scetce Gunnar Ratsch 15. 3. 2022 21\n\nmedical_ data_ science _ Ezurich Introduction to the Transformer model Transformer is the encoder-decoder architecture with self-attention: The key component of the Encoder is Self-Attention mechanism. Layer normalization Input data: X e Rnxd Encoder overview: Q-queries, K keys, V values: linear projections of the input data X, d dimension: # QKT Latent Representation softmax )V Residual 1 Va connections Add & Normalize Feed Forward Feed Forward Projection layer / Add & Normalize Self-Attention Attention weights Positional embeddings POscoDMG Self attention Positional Embeddings is used to learn the order in the sequence. Residual connections mitigates vanishing gradients. Layer normalization helps with numerical stability: POSITIONAL ENCODING sin if i = 2k PEpi coS if i = 2k+1 EMBEDDINGS X1 Pictures from The Ilustrated transformer_Jay Alammar DINFK Vaswani, Ashish, et al. "Attention is allyou need"Advances in neural information processing systems 30 (2017). Gunnar Ratsch 15. 3. 2022 22\n\nmedical_ data_ science _ Elzurich Monitoring Patient State in Intensive Care PRESCRIBING TREATMENT ICU PATIENT EHR CLINICIAN OUR WORK DATA RECORDING PREDICTING PATIENT EVOLUTION ML MODEL DINFK Yeche, Hugo; Kuznetsova Rita at al, HiRID-ICU-Benchmark A Comprehensive Machine Learning Benchmark on High-resolution ICU Data, 35th Conference on Neural Information Processing Systems (NeurIPS 2021) Track on Datasets and Benchmarks_ Gunnar Ratsch 15. 3. 2022 23\n\nmedical_ data_ science _ Ezurich Current Limitations of Existing EHR Datasets EXISTING ICU DATABASE AnsterdamUMC Downstream Tasks Pipeline Reproducibility No preprocessing leads to unfair comparison across work due to irreproducible: MIMIC-III eICU Few Database contain labels When label is provided, task are not clinically relevant (eg: Mortality) Different label definition across work HiRID Data splitting Data processing Label definition Use of external data Our work tackles both issues for HiRID database DINFK https:Ilgithub comlratschlab/HIRID-ICU-Benchmark Gunnar Ratsch 15. 3. 2022 24\n\nmedical_ data_ science _ Ezurich Define a Variety of Clinically Relevant Tasks X Circulatory Failure Respiratory Failure Kidney Function Length of Stay Mortality Phenotyping Predict whether a patient is going to experience a circulatory failure in the next 12 hours Predict whether a patient is going to experience a respiratory failure in the next 12 hours Predict a patient average urine production in the next 2 hours Predict at a given time the remaining length of stay of a patient in the ICU Predict whether a patient is going to expire in the ICU after 24 hours of stay: Predict a patient admission group (APACHE) after 24 hours of stay DINFK Gunnar Ratsch 15. 3. 2022 25\n\nmedical_ data_ science _ Ezurich Benchmarking SOTA Machine Learning Approaches Task ICU Mortality AUPRC (1) AUROC 60. 3 + 1. 6 90. 0 = 0. 4 60. 0 = 0. 9 90. 3 = 0. 2 60. 2 + 1. 1 89. 7 = 0. 4 61. 0 = 0. 8 90. 8 = 0. 2 Patient Phenotyping B-Accuracy 39. 2 +2. 1 39. + 1. 2 41. 6 +23 42. 7 +1. 4 Benchmark codeldata: https:Ilgithub. comlratschlab/HIRID-ICU-Benchmark Metric GRU LSTM TCN Transformer Task Circulatory failure Respiratory failure Metric AUPRC AUROC (1) AUPRC AUROC GRU 36. 8 = 0. 5 90. 7 + 0. 2 59. 2+0. 3 70. 1 =0. 2 LSTM 32. 6 + 0. 8 89. 9 + 0. 1 56. 9 +0. 3 68. 2 +0. 3 TCN 35. 8 + 0. 6 90. 5 + 0. 1 58. 9 + 0. 3 70. 0 + 0. 2 Transformer 35. 2+0. 6 90. 6 + 0. 2 59. 4 +0. 3 70. 1+0. 2 Kidney func. Remaining LOS MAE MAE 0. 49 = 0. 02 0. 50 + 0. 01 0. 50 + 0. 01 0. 48 + 0. 02 60. 6 = 0. 9 60. 7 + 1. 6 59. 8 EUR 2. 8 59. 5+2. 8 TCN Temporal Convolution Networks (cf. Image analysis lecture) Transformers explained next lecture Yeche Hugo, Kuznetsova Rita et al. DINFK Gunnar Ratsch 15. 3. 2022 26\n\nmedical_ data_ science _ Ezurich Generative Models Generative models are probabilistic models of high-dimensional data. Describe the probabilistic process of generating an observation: The emphasis is on capturing the dependence between the dimensions: Provide a way of generating new datapoints: Historically, generative modelling was considered to be a subfield of unsupervised learning: Usage of generative modelling: Representation learning; density estimation, data compression etc. Latent Variable Models is a type of Generative models_ Specify the generative process in terms of unobserved/latent variables and the transformation that maps them to the observation: Trained with maximum likelihood (usually with some approximations) Easy to incorporate prior knowledge structure into models; fast generation: Need to use approximate inference or restricted models. DINFK The Deep Learning Series Lecture Series 2020, Deep Mind Gunnar Ratsch 15. 3. 2022 27\n\nmedical_ data_ science _ Ezurich Generative Adversarial Networks Model: A neural net that maps noise vectors to observations Training: use the learning signal from a classifier trained to discriminate between samples from the model and the training data Pros Can generate very realistic images Real images Conceptually simple implementation Fast generation 1 Generator L Cons Cannot be used to compute probability of observations "Mode collapse' Models ignore regions of the data distribution Training can be unstable and requires many tricks to work well Sample { | Discriminator Sample { 0 (More details in the lecture on privacy to generate realistic artificial data) DINFK The Deep Learning Series Lecture Series 2020, Deep Mind Picture from Google_Developers Gunnar Ratsch 15. 3. 2022 28\n\nmedical_ data_ science _ Ezurich Latent Variable Models 3 Variational Autoencoder Consider dataset X = {x}i1, N consisting of N i. i. d. samples. We assume that the data are generated by some random process, involving an unobserved continuous random variable z ~ N (0, I): X~ pe (xlz), Z where pe(zlx) is unknown. The lower bound on the marginal likelihood is: LvAE [ 9(zl) log pe (xlz)dz KL(qo(z/x)llp(z)), reconstruction loss "regularizer"Iprior p(z) LvAE max 0, $ 96(z/x) & pe(xlz) is modelled by neural networks: q6(z/x) ~ N(0o, EUR 02) encoder, po (xlz) ~ N(ue, o83 decoder. 06, 02, 1o, 03 2 the output of the neural networks: DINFK DP: Kingma, M. Welling, Auto-Encoding Variational Bayes I/ International Conference of Learning Representations 2014_ X N 2 2 Encoder 9zlx) Decoder P(xlz) Data: X Reconstruction: * Neural Network Perspective [source] Gunnar Ratsch 15. 3. 2022 29\n\nmedical_ data_ science _ Ezurich Towards interpretable health state representations Idea: Use self-organizing maps to encourage interpretable neighborhood relationship in latent space and smoothness over time Renal Idysfunction Desirable properties Discretellow dimensional Smooth over time Expressive Interpretable Cardiac dysfunction Healthy, https 1 /en wikipedia org/wiki 'Self-organi zing map# /media File Somtraining vg Vincent Fortuin; Matthias Huser; Francesco Locatello, Heiko Strathmann; and Gunnar Ratsch, 2019. SOMVAE: Interpretable discrete representation learning on time series: International Conference on Learning Representations (ICLR) 2019. Gunnar Ratsch 15. 3. 2022 30 DINFK\n\nmedical_ data_ science _ Ezurich SOM-VAE model for discrete time series representations input ^w encoder latent encodings Markov model decoder reconstruction t+1 2 e t 2 2 t+1 4 P(z t+1l2. 4) q self-organizing map t Ze L(xt-1 xt '22 9, &t e) LsoM-VAE (2t 2 24, 18) +yLtransitions rt_1, xt) + t Lsmoothness rt_1 2 xt) LsoM-VAE(T, Sq, Te) Lreconstruction (1, iq, Ze) + a Lcommitment (w) + 8 Lsom (x) Model is jointly optimized (gradient descent wl special handling of discrete states) DINFK Vincent Fortuin; Matthias Hiser; Francesco Locatello, Heiko Strathmann; and Gunnar Ratsch, 2019. SOMVAE: Interpretable_discrete representation learningon_time series: International Conference on Learning Representations (ICLR) 2019. Gunnar Ratsch 15. 3. 2022 31 N\n\nmedical_ data_ science _ Ezurich Health state representations on 2D SOM-grid Method score 6 score 12 score 24 k-means 0. 0411 + 0. 0007 0. 0384 + 0. 0006 0. 0366 = 0. 0005 SOM-VAE 0. 0407 = 0. 0005 0. 0376 + 0. 0004 0. 0354 = 0. 0004 SOM-VAE-prob 0. 0474 + 0. 0006 0. 0444 + 0. 0006 0. 0421 + 0. 0005 Performance comparison of our method wth and_without Markov model (SOM-VAE-prob and SOM-VAE) against k-means in terms of normalized mutual information. Dynamic endpoints are the maximum of the physiology score within the next 6, 12 or 24 hours (6_hours, 12_ hours, 24 hours): Each method is used to fit 64 clusters. Shown are means and standard errors over 10 runs. DINFK Vincent Fortuin; Matthias Hiser; Francesco Locatello, Heiko Strathmann, and Gunnar Ratsch, 2019. SOMVAE: Interpretable discrete Gunnar Ratsch 15. 3. 2022 32 representation learning on _time series: International Conference on Learning Representations (ICLR) 2019.\n\nmedical_ data_ science _ Ezurich Health state representations on 2D SOM-grid 12 10 (a) k-means (b) VQ-VAE (c) SOM-VAE (d) Patient trajectories Color of SOM-cells: Average dynamic APACHE score in the next 24 hours. Higher APACHE score is associated with increased severity of patient health state. vQ-VAE: Van den Oord et al., 2017 (https Ilarxiv orglabs/1711. 00937) Vincent Fortuin Vincent Fortuin, Matthias Hiser; Francesco Locatello, Heiko Strathmann, and Gunnar Ratsch, 2019. SQM-VAE: Interpretablediscrete Gunnar Ratsch 15. 3. 2022 33 representation learning on time series: International Conference on Learning Representations (ICLR) 2019. DINFK\n\nmedical_ data_ science _ Ezurich Health state representations on 2D SOM-grid 0. 2 Ston Eno * Stant End: 016 0 00 0. 06 WW 0 02 Figure 3: Illustration of two example patient trajectories in the SOM grid of T-DPSOM49. One patient died (red), while the other was discharged alive from the ICU (green). Superimposed is a heatmap that displays the mean APACHE score of all time points assigned to each cluster: We observe qualitative differences in the trajectories of the dying and the surviving patient: For each time series, we also show the assigned probabilities to the discrete patient health states using a blue color shading: In this work; we will separate the representations into organ systems and manually annotate different areas of the "map" using medical knowledge in collaboration with intensive care specialists. Laura Manduchi; Matthias Huser; Julia Vogt; Gunnar Ratsch, Vincent Fortuin, DPSOM: Deep Probabilistic Clustering with Self-Organizing Maps, ACM-CHIL, 2020 DINFK\n\nmedical_ data_ science _ 1 Ezurich Contrastive Learning Contrastive learning is a machine learning technique used to learn the general features ofa dataset without labels by teaching the model which data points are similar or different: [Source] Vector Representation 2, Maximize agreement Vector Representation z g() Dense Layer Dense Layer Dense Layer Dense Layer Dense Layer Dense Layer Intermediate Representation h CNN Neural Network Encoder fC) CNN Neural Network Encoder fC) Similar samples < Similar downstream task label Positive Pairs X Data Augmentation x New challenges for Contrastive Learning DINFK Gunnar Ratsch 15. 3. 2022 35\n\nmedical_ data_ science _ 1 Ezurich Contrastive Learning for Time Series Contrastive learning is a machine learning technique used to learn the general features ofa dataset without labels by teaching the model which data points are similar or different: [Source] Contrastive Learning Learn a representation of a patient state at a give time t0 t-th Patient stay Similar samples < Similar downstream task label New challenges for Contrastive Learning Hugo Yeche et al Yeche, Hugo, et al. 'Neighborhood contrastive learning applied t0 online patient monitoring: International Conference on Machine Learning. PMLR, 2021. Gunnar Ratsch DINFK 15. 3. 2022 36\n\nmedical_ data_ science _ Ezurich Challenges in Using Contrastive Learning for ICU monitoring 2N 'exp pi pvli) TT CCL log M i=1 kzi exp Pi pj/t, Alignment Normalization How to construct different views outside of computer vision [1]? Should all "negatives" be treated the same [2]? CV ICU CV ICU Diversity among samples Humanly understandable signal Complex modality Strong similarities among contiguous samples i. i. d distribution of samples Balanced multi-class downstream task multiples samples from a single patient (non ii. d) Usually imbalanced downstream task Relies on strong data augmentation Limits usage of data augmentation Clear hierarchy among negative samples Uniform Normalization [1] Tian et al: (2020) [2] Wang et al. (2020) Gunnar Ratsch 15. 3. 2022 37 DINFK\n\nmedical_ data_ science _ Ezurich Preserve hierarchy in the data with Neighborhood Contrastive Loss How can infuse prior Redefine Contrastive knowledge without relying Loss through the lens of on data augmentation? Neighborhood Two samples share the same neighborhood if they share some predefined attributes: Formally Examples Supervised: n(i, k) = 1if Ti and Tk share attributes 0 else ny(i, k) = 1if yi Yk Temporal nw (i, k) 1ifli kl < W and Si Sk Wang et al:. (2020) DINFK Gunnar Ratsch 15. 3. 2022 38\n\nmedical_ data_ science _ Ezurich Neighborhood Contrastive learning Objective ND NA Aligning neighbors Discriminating neighbors 2N Zi zvli) ZkIn(i, k) = 1 2kIn(i, k) = 0 push away pull towards 2N exp m Pi p)/v) log keN(i) exp (Pi 9k/t) i=1 ~l CNA exp (Pi_ pin /t) = log NG)I M i=1 leN(i) kzi exp (pi aklt) = 3 LNCL aLNA + (1 _ a)LND DINFK Gunnar Ratsch 15. 3. 2022 39 CND\n\nmedical_ data_ science _ 1 Ezurich A Unifying Framework Method U n(, CL 1. 0 0 nw SACL (Cheng et al,, 2020) 0. 0 +00 Tw CLOCS (Kiyasseh et al,, 2020) 1. 0 +00 nw SCL (Khosla et al,, 2020) 1. 0 NA ny We explore two cases NCL: Unsupervised, we call NCL(nw) where: n = nw; W EUR J0, +o [; & EUR J0, 1[ Supervised, we call NCL(ny) where:n ny ; & EUR ]0, 1[ Task Sepsis onset prediction AUROC (in %_ Metric AUPRC (in % _ Utility (x100) Linear MLP Head Linear MLP Linear MLP Seq2-Seq-AE 7. 0 + 0. 3 Seq2-Seq-AE-forecast 6. 6 = 0. 3 CL 7. 9 = 0. 4 SACL (Cheng et al,, 2020) 6. 5 =0. 3 CLOCS (Kiyasseh et al,, 2020) 7. 1 +0. 5 NCL(nw ) (Ours) 8. 2 = 0. 4 7. 8 + 0. 4 7. 3 +0. 3 95 + 0. 4 7. 6 =0. 3 7. 3 +0. 4 93 + 0. 5 77. 1 +0. 5 78. 1+0. 6 75. 8 + 0. 9 76. 9 +0. 5 78. 2 =0. 3 80. 2 + 0. 4 73. 0 = 1. 2 75. 3 +0. 8 77. 2 +05 78. 8 = 0. 4 78. 8 = 0. 3 80. 7 = 0. 3 26. 8 = 1. 0 23. 5 + 1. 5 26. 2 + 0. 8 20. 5: 2. 5 23. 0 =1. 1 27. 2 = 1. 0 27. 2 + 1. 0 23. 8 = 1. 2 29. 7 + 1. 0 24. 2+1. 1 25. 8 + 0. 9 30. 2 = 1. 0 End-to-End SCL (Khosla et al,, 2020) NCL(ny ) (Ours) 7. 6 = 0. 2 6. 7 =0. 6 10. 0 = 0. 5 8. 1 +0. 4 6. 0 + 0. 5 10. 1 + 0. 3 78. 9 + 0. 3 73. 1 +1. 7 80. 3 + 0. 4 78. 8 + 0. 4 70. 0 + 1. 9 80. 8 = 0. 2 27. 9 +0. 8 20. 2 +2. 7 32. 6 + 1. 0 27. 5 1. 0 20. 6 = 1. 7 31. 9 + 0. 9 Experiments on MIMIC-IIl and Physionet 2019 datasets. DINFK Gunnar Ratsch 15. 3. 2022 40\n\nmedical_ data_ science _ Ezurich Summary & Take Home messages Representation learning is a recently developed, powerful tool to learn integrative computational summaries of observed data. Health state data is one interesting application where we assume that the patients physiological health state can be accurately represented in vector form: Autoencoders and forecaster models learn vector representations of past data that are predictive of the past and future, respectively: Generative models are an important tool for finding additional representations and to generate realistic data Contrastive learning can improve learning representations of time series DINFK Gunnar Ratsch 41 |
| source_doc_filename | source_doc_id | source_doc_domain | document_text | |
|---|---|---|---|---|
| 9 | OCR_ML4HLecture04RepresentationLearning.pptx_.txt | 65105d7b-502 | OCR | \n\nmedical_ data_ science Ezurich Lecture 6 Machine Learning for Health Care" (261-5120-00L) (Health) Representation Learning Gunnar Ratsch; Julia Vogt, Valentina Boeva Biomedical Informatics group, Institute for Machine Learning, Department of Computer Science DINFK Gunnar Ratsch 15. 3. 2022\n\nmedical_ data_ science Ezurich Outline for today Motivation of Latent Representations Autoencoders & Sequence-to-sequence models Transformers ICU Benchmarks Generative models VAEs (GANs) SOM-VAEs Contrastive Learning DINFK Gunnar Ratsch 15. 3. 2022 2\n\nmedical_ data_ science Ezurich Towards Comprehensive Patient Models for Precision Medicine Distill Data Embed Data Connect Data Clinical Trial Design Precision Medicine Health Records Pathology Images Genomic Data X=w( p(phenotype x, t) Drugs Predict Treatment Mobile Health = argmax p(survival X, DINFK Gunnar Ratsch 15. 3. 2022 3\n\nmedical_ data_ science Elzurich What is a computational representation of a patient? Data modalities, redundancies Health state, relevant parameters Temporal aspects Ascertainment biases Interpretability What can we learn from a computational representation of a patient? Expectations Limitations How can we measure how good a representation is? DINFK Gunnar Ratsch 15. 3. 2022 4\n\nmedical_ data_ science Ezurich Computational patient representations Summary of health state of patient one of n possible discrete states a vector describing different physiological aspects Based of heterogeneous data patient state may be represented in lab test and a doctor's note there may be different ways to assess a physiological state Dimensionality reduction" how many dimensions do we need to represent a patient? Expectations: representation faithfully represents patient health state is predictive of future health states or diseases Can it reproduce the original data? Limitations: Loss of some information only relevant for small number of patients Curse of dimensionality & sample sizes Measures of quality low dimensionality interpretability prediction accuracy DINFK Gunnar Ratsch 15. 3. 2022 5\n\nmedical_ data_ science _ Ezurich Unsupervised Learning of Health States Unsupervised _Learning Patient Time Series Patient Representation Hospital Data Warehouse d Raw Patient Dataset Medications Diagnoses Clinical Descriptors Procedures Lab Tests t-T+1 Supervised Learning Deep Patient Dataset Patients Features Two major cases: 1. Latent spaces 2. Discrete states Drug Targeting Patient Similarity Disease Clinical Trial Prediction Recruitment Personalized Prescription DINFK Gunnar Ratsch 15. 3. 2022 6 Miotto et al_Scientific Reports_2016\n\nmedical_ data_ science F Ezurich SOM-VAE: INTERPRETABLE DISCRETE REPRESENTATION LEARNING ON TIME SERIES Associated Manuscripts Vincent Fortuin, Matthias Hiiser & Francesco Locatello Department of Computer Science, ETH Ziirich Universitatsstrasse 6, 8092 Ziirich, Switzerland {fortuin, mhueser, locatelf} @inf. ethz ch Heiko Strathmann Gatsby Unit; University College London 25 Howland Street, London WIT 4JG, United Kingdom heiko strathmann@gmail com Improving Clinical Predictions through Unsupervised Time Series Representation Learning Gunnar Ratsch Department of Computer Science, ETH Ziirich Universitatsstrasse 6. 8092 Ziirich, Switzerland raetscheinf. ethz ch Xinrui Lyul, Matthias Hiiser', Stephanie L. Hyland', George Zerveas?, Gunnar Ratschl Biomedical Informatics Group; Dept. of Computer Science, ETH Zirich 2 AI Lab, Center for Biomedical Informatics, Brown University ABSTRACT Abstract High-dimensional time series are common in many domains_ Since human cognition is not optimized to work well in high-dimensional spaces, these areas could benefit from interpretable low-dimensional representations. However; most representation learning algorithms for time series data are difficult to interpret: This is due to non-intuitive mappings from data features to salient properties of the representation and non-smoothness over time To address this problem; we propose a new representation learning framework building on ideas from interpretable discrete dimensionality reduction and deep generative modeling: This framework allows uS to learn discrete representations of time series, which give rise to smooth and interpretable embeddings with superior clustering performance. We introduce a new way to overcome the non-differentiability in discrete representation learning and present a gradient-based version of the traditional self-organizing map algorithm that is more performant than the original. Furthermore, to allow for a probabilistic interpretation of our method, we integrate a Markov model in the representation space. This model uncovers the temporal transition structure, improves clustering performance even further and provides additional explanatory insights as well as a natural representation of uncertainty: We evaluate our model in terms of clustering performance and interpretability on static (Fashion-JMNIST data, a time series of linearly interpolated (Fashion-JMNIST images, a chaotic Lorenz attractor system with two macro states, as well as on a challenging real world medical time series application on the eICU data set: Our learned representations compare favorably with competitor methods and facilitate downstream tasks On the real world data. In this work, we investigate unsupervised representation learning On medical time series, which bears the promise of leveraging copious amounts of existing unlabeled data in order to eventually assist clinical decision making: By evaluating on the prediction of clinically relevant outcomes, we show that in practical setting, unsupervised representation learning can offer clear performance benefits over endto-end supervised architectures. We experiment with using sequence-to-sequence (Seq2Seq) models in two different ways, as an autoencoder and as a forecaster; and show that the best performance is achieved by a forecasting Seq2Seq model with an integrated attention mechanism; proposed here for the first time in the setting of unsupervised learning for medical time series. DINFK Gunnar Ratsch 15. 3. 2022\n\nmedical_ data_ science _ Ezurich A natural choice: autoencoders Given a signal Xi we want to learn f and 9 f : Rd 7 Rm 9 : Rm 7 Rd Ingvs f(xi) = ei g(ei) = xi VMFUt that minimizes the reconstruction error L = Ilxi xill? Ilxi 94 ( fo(xi)) |2 Lucodr Ucuda f is the encoder function; and 9 is the decoder function; and ei is the representation: f and 9 are jointly optimized (gradient descent) (see Intro ML or Computational Intelligence Lab for extensive introductions) DINFK Gunnar Ratsch 15. 3. 2022 8\n\nmedical_ data_ science _ Ezurich Sequential modeling of health states Sequential autoencoder Autoencoder (AE) Most common in unsupervised representation learning The decoder reconstructs the input featureslsignals Sequence-to-sequence AE (S2S-AE) Encoder/decoder structure: sequential NN (e. g-, RNNILSTMIGRU) Input: multivariate time series up to time t Drawbacks: the representation only encodes data from thepast Sequential forecasters (with attention) Sequence-to-sequence forecaster (S2S-F) Same structure as S2S-AE The decoder predicts future_time series from time t+1 Focus on relevant past to predict future S2S-F with Attention (S2S-F-A) Attention helps focus more on past information that is most predictive of the future. Lyu; X, Huser; M: Hyland, S. L, Zerveas, G. and Ratsch, G., 2018. Improving Clinical Predictions through Unsupervised Time Series DINFK Representation Learning: Machine Learning for Health (ML4H) Workshop NeurIPS 2018 Gunnar Ratsch 15. 3. 2022 9\n\nmedical_ data_ science _ Ezurich Overviewllntuitions Latent health state Observed data time relevant event DINFK Gunnar Ratsch 15. 3. 2022 10 L Forecaster\n\nmedical_ data_ science _ Ezurich Review: Recurrent Neural Networks & LSTMs & GRUs AII Recurrent Neural Networks have the form of a chain of repeating modules of neural network. ht ho__ h1 hz_ Recurrent Neural Network ht A A A A Xo X1 Xz Xt See more detailed introductions to RNNs, LSTMs, GRUs, for instance, in class "Deep Learning" _ DINFK Neural Network Layer Pointwise Operation Vector Transfer Pictures from Colah's blog Gunnar Ratsch 15. 3. 2022 11 Concatenate Copy\n\nmedical_ data_ science _ Ezurich Review: Recurrent Neural Networks & LSTMs & RNN AII Recurrent Neural Networks have the form of a chain of repeating modules of neural network. nt: ht Recurrent Neural Network Single tanh Layer X directly modulates h A tanh A Xt t+ DINFK Neural Network Layer Pointwise Operation Vector Transfer Pictures from Colah's blog Gunnar Ratsch 15. 3. 2022 12 Concatenate Copy\n\nmedical_ data_ science _ Ezurich Review: Recurrent Neural Networks & LSTMs & RNN AII Recurrent Neural Networks have the form of a chain of repeating modules of neural network. nt: ht Recurrent Neural Network Single tanh Layer X directly modulates h A tanh A Long-Short Term Memory (LSTM) ht Gated Recurrent Unit (GRU) ht 1 The first gate determines what the hidden state forgets ("forget gate"): 2_ The second gate decides which values we'Il update ("input gate") ) 3 The third gate creates a vector of new candidate values. 4_ The last gate produces an output: The forget and input gates ht_) combine into a single "update gate' + another changes tanh tanh 0 tanh simpler than LSTM xt Xt DINFK Neural Network Layer Pointwise Operation Vector Transfer Pictures from Colah's blog Gunnar Ratsch 15. 3. 2022 13 Concatenate Copy\n\nmedical_ data_ science _ Ezurich Seq2Seq models as autoencoders et Xt Linear layer Linear layer Linear layer Linear layer ht LSTM LSTM LSTM LSTM LSTM LSTM X(t_4) Xlt_A)+1 Xt (t-4) Xt-1 Encoder Decoder DINFK Gunnar Ratsch 15. 3. 2022 14 X(t_^) X(t-A)+l Xt-1 ht_^ ht-4+1 ht_1\n\nmedical_ data_ science _ Ezurich Seq2Seq models as autoencoders Encoder and decoder function fe (xi (t-4+1) (t-4+2) (t) (t) xi Xi = ei 9d (eft)) = x(t-4+1), x6t-4+2) x(t) Loss function ZAO (t-j) L(Xt) Ilxi x(t-j)112 Ti L( Xi) L(Xt) t=4 The decoder reconstructs the inputs, i. e., the historical time series of the patient Known from and used extensively in Natural Language Processing DINFK Sutskever et al_NIPS 2014 Gunnar Ratsch 15. 3. 2022 15\n\nmedical_ data_ science _ Ezurich Seq2Sea models as forecasters et Linear layer Linear layer Linear layer Linear layer LSTM LSTM LSTM LSTM LSTM LSTM X(t_4) X(t-4)+l Xt Encoder Decoder Teacher Forcing: training procedure, that is used for RNN (NLP; generation tasks): Instead of using the predicted values at time-step t, we passed the ground-truth values for this time-step. Pros: speed-up the training: Cons: limited models Best way: use the combination of teacher-forced values and predicted values: DINFK Gunnar Ratsch 15. 3. 2022 16 More ways, now to use teacher forcing: 1) Deep_Learning Book; chapter 10. 2) Professor Forcing; NIPS 2016. Xt-1 Xt-2 Xt+A-1 Xt-A ht+1 ht+2 ht+A-1 ht+A Xt+A-1 Xt+1\n\nmedical_ data_ science _ 1 Ezurich Introduction to the Attention Mechanism Seq2Seq encoder compresses the information into a context vector of a fixed length ~ incapability of remembering long "sentences" Task: Given source sequence X = (11, CTr predict a target sequence y = (y1, YTy _ Encoder maps the input to the hidden states (h1, hTz ) with the help of RNN (bottom) Yt-1 Yt St-1l St Decoder for each i has hidden state 8i = f(si-1, Yi-1, Ci) f-RNN (top) The context vector for the output yi is computed: weighted sum of hidden encoder states (middle) Tc exp (eij_ Ci Qijhj where Qij Tz k=-1 exp (eik) j=1 Ot, T 0t, 3 h1 hz ht The set of Qij are weights defining how much of each source hidden state should be considered tor each output h1 hz h3 ht a(8i-1, hj) = v4 tanh (Wa8i-1 + Uahj) alignment score, eij (image shows bidirectional network also works for unidirectional ones) how well the inputs around position j and the output at position i match: DINFK Gunnar Ratsch 15. 3. 2022 17 Bahdanau; Dzmitry, Kyunghyun Cho, and Yoshua Bengio. "Neural machine_translation_by _jointly learning to_align and translate: at, 1 0t, 2, nzl\n\nmedical_ data_ science _ Ezurich Seq2Seq forecasters with(attention Attention model Linear layer Linear layer Linear layer Linear layer et-^ et-4+1 et LSTM LSTM LSTM LSTM LSTM LSTM X(t-^) X(t-4)+1 [Xt+1, Ct+1] [Rt+4-1, Ct+4-1] Encoder Decoder The same idea: at timer EUR {t+1,., t+T} during decoding the objective is to produce a context vector Ct a weighted combination of the hidden states of the encoder: DINFK Gunnar Ratsch 15. 3. 2022 18 Ct+l Xt+1 Xt+2 Xt+A-1 Xt+A ht+1 ht+2 ht+A-1 ht+A\n\nmedical_ data_ science _ Ezurich Experimental Evaluation Setting Patient Patient Time Series Representation Hospital Data Warehouse d Raw Patient Dataset t-T+1 Medications Diagnoses Clinical Descriptors Procedures Lab Tests Unsupervised learning of representations: name nonlinear temporal decoder output attention Deep Patient Dataset Patients PCA past Features AE past S2S-AE past Supervised Learning Task: Healthy discharge (within 24h) S2S-F future S2S-F-A future DINFK Miotto et al_Scientific Reports_2016 Gunnar Ratsch 15. 3. 2022 19\n\nmedical_ data_ science _ Ezurich Experimental evaluation Data: Multivariate ICU time series (d=94) from the Philips eICU_dataset (vital signsllab test results, ~20'000 patients) Length of encodedlpredicted time series: 12 h (resolution: 1 samplelh) Embedding dimension: 94 (compression rate: 12. 1) Supervised learning method: LSTM with one layer: 24h Discharge 0. 45 24h Discharge AUPRC AUROC 0. 40 1 0. 35 PCA rep AE rep. S2S-AE rep. S2S-F rep_ S2S-F-A rep 0. 436 = 0. 01 0. 811 = 0. 004 0. 824 = 0. 002 0. 824 E 0. 003 0. 825 + 0. 003* 0. 825 + 0. 003* Supervised (LSTM-3) Supervised (LSTM-1) LSTM-1 + PCA rep. LSTM-1 + AE rep. LSTM-1 + S2S-AE rep. LSTM-1 + S2S-F rep. LSTM-1 + S2S-F-A rep. 0. 471 = 0. 005 0. 474 + 0. 006 0. 477 + 0. 006* 0. 48 + 0. 007 0. 30 0. 25 1% 5% 10% 25% % of labeled data 50% 100% Lyu; X,, Hiser; M:, Hyland, S. L, Zerveas, G. and Ratsch, G. 2018. Improving Clinical Predictions through Unsupervised Time Series DINFK Representation Learning: Machine Learning for Health (MLAH) Workshop NeurIPS 2018 Gunnar Ratsch 15. 3. 2022 20\n\nmedical_ data_ science _ Ezurich Discussion (latent space) S2S models reduce input time-series to low dimensional embeddings and still achieve better performance than using the raw features. S2S-F-A outperforms the others when sufficient data is available. When labeled data is limited, 9 deep unsupervised representation shallow supervised learning can outperform deep supervised learning: Ker Science; Unl Sqctxl Xinrui Lyu Matthias Huser et al. Jsse DINFK Let Scetce Gunnar Ratsch 15. 3. 2022 21\n\nmedical_ data_ science _ Ezurich Introduction to the Transformer model Transformer is the encoder-decoder architecture with self-attention: The key component of the Encoder is Self-Attention mechanism. Layer normalization Input data: X e Rnxd Encoder overview: Q-queries, K keys, V values: linear projections of the input data X, d dimension: # QKT Latent Representation softmax )V Residual 1 Va connections Add & Normalize Feed Forward Feed Forward Projection layer / Add & Normalize Self-Attention Attention weights Positional embeddings POscoDMG Self attention Positional Embeddings is used to learn the order in the sequence. Residual connections mitigates vanishing gradients. Layer normalization helps with numerical stability: POSITIONAL ENCODING sin if i = 2k PEpi coS if i = 2k+1 EMBEDDINGS X1 Pictures from The Ilustrated transformer_Jay Alammar DINFK Vaswani, Ashish, et al. "Attention is allyou need"Advances in neural information processing systems 30 (2017). Gunnar Ratsch 15. 3. 2022 22\n\nmedical_ data_ science _ Elzurich Monitoring Patient State in Intensive Care PRESCRIBING TREATMENT ICU PATIENT EHR CLINICIAN OUR WORK DATA RECORDING PREDICTING PATIENT EVOLUTION ML MODEL DINFK Yeche, Hugo; Kuznetsova Rita at al, HiRID-ICU-Benchmark A Comprehensive Machine Learning Benchmark on High-resolution ICU Data, 35th Conference on Neural Information Processing Systems (NeurIPS 2021) Track on Datasets and Benchmarks_ Gunnar Ratsch 15. 3. 2022 23\n\nmedical_ data_ science _ Ezurich Current Limitations of Existing EHR Datasets EXISTING ICU DATABASE AnsterdamUMC Downstream Tasks Pipeline Reproducibility No preprocessing leads to unfair comparison across work due to irreproducible: MIMIC-III eICU Few Database contain labels When label is provided, task are not clinically relevant (eg: Mortality) Different label definition across work HiRID Data splitting Data processing Label definition Use of external data Our work tackles both issues for HiRID database DINFK https:Ilgithub comlratschlab/HIRID-ICU-Benchmark Gunnar Ratsch 15. 3. 2022 24\n\nmedical_ data_ science _ Ezurich Define a Variety of Clinically Relevant Tasks X Circulatory Failure Respiratory Failure Kidney Function Length of Stay Mortality Phenotyping Predict whether a patient is going to experience a circulatory failure in the next 12 hours Predict whether a patient is going to experience a respiratory failure in the next 12 hours Predict a patient average urine production in the next 2 hours Predict at a given time the remaining length of stay of a patient in the ICU Predict whether a patient is going to expire in the ICU after 24 hours of stay: Predict a patient admission group (APACHE) after 24 hours of stay DINFK Gunnar Ratsch 15. 3. 2022 25\n\nmedical_ data_ science _ Ezurich Benchmarking SOTA Machine Learning Approaches Task ICU Mortality AUPRC (1) AUROC 60. 3 + 1. 6 90. 0 = 0. 4 60. 0 = 0. 9 90. 3 = 0. 2 60. 2 + 1. 1 89. 7 = 0. 4 61. 0 = 0. 8 90. 8 = 0. 2 Patient Phenotyping B-Accuracy 39. 2 +2. 1 39. + 1. 2 41. 6 +23 42. 7 +1. 4 Benchmark codeldata: https:Ilgithub. comlratschlab/HIRID-ICU-Benchmark Metric GRU LSTM TCN Transformer Task Circulatory failure Respiratory failure Metric AUPRC AUROC (1) AUPRC AUROC GRU 36. 8 = 0. 5 90. 7 + 0. 2 59. 2+0. 3 70. 1 =0. 2 LSTM 32. 6 + 0. 8 89. 9 + 0. 1 56. 9 +0. 3 68. 2 +0. 3 TCN 35. 8 + 0. 6 90. 5 + 0. 1 58. 9 + 0. 3 70. 0 + 0. 2 Transformer 35. 2+0. 6 90. 6 + 0. 2 59. 4 +0. 3 70. 1+0. 2 Kidney func. Remaining LOS MAE MAE 0. 49 = 0. 02 0. 50 + 0. 01 0. 50 + 0. 01 0. 48 + 0. 02 60. 6 = 0. 9 60. 7 + 1. 6 59. 8 EUR 2. 8 59. 5+2. 8 TCN Temporal Convolution Networks (cf. Image analysis lecture) Transformers explained next lecture Yeche Hugo, Kuznetsova Rita et al. DINFK Gunnar Ratsch 15. 3. 2022 26\n\nmedical_ data_ science _ Ezurich Generative Models Generative models are probabilistic models of high-dimensional data. Describe the probabilistic process of generating an observation: The emphasis is on capturing the dependence between the dimensions: Provide a way of generating new datapoints: Historically, generative modelling was considered to be a subfield of unsupervised learning: Usage of generative modelling: Representation learning; density estimation, data compression etc. Latent Variable Models is a type of Generative models_ Specify the generative process in terms of unobserved/latent variables and the transformation that maps them to the observation: Trained with maximum likelihood (usually with some approximations) Easy to incorporate prior knowledge structure into models; fast generation: Need to use approximate inference or restricted models. DINFK The Deep Learning Series Lecture Series 2020, Deep Mind Gunnar Ratsch 15. 3. 2022 27\n\nmedical_ data_ science _ Ezurich Generative Adversarial Networks Model: A neural net that maps noise vectors to observations Training: use the learning signal from a classifier trained to discriminate between samples from the model and the training data Pros Can generate very realistic images Real images Conceptually simple implementation Fast generation 1 Generator L Cons Cannot be used to compute probability of observations "Mode collapse' Models ignore regions of the data distribution Training can be unstable and requires many tricks to work well Sample { | Discriminator Sample { 0 (More details in the lecture on privacy to generate realistic artificial data) DINFK The Deep Learning Series Lecture Series 2020, Deep Mind Picture from Google_Developers Gunnar Ratsch 15. 3. 2022 28\n\nmedical_ data_ science _ Ezurich Latent Variable Models 3 Variational Autoencoder Consider dataset X = {x}i1, N consisting of N i. i. d. samples. We assume that the data are generated by some random process, involving an unobserved continuous random variable z ~ N (0, I): X~ pe (xlz), Z where pe(zlx) is unknown. The lower bound on the marginal likelihood is: LvAE [ 9(zl) log pe (xlz)dz KL(qo(z/x)llp(z)), reconstruction loss "regularizer"Iprior p(z) LvAE max 0, $ 96(z/x) & pe(xlz) is modelled by neural networks: q6(z/x) ~ N(0o, EUR 02) encoder, po (xlz) ~ N(ue, o83 decoder. 06, 02, 1o, 03 2 the output of the neural networks: DINFK DP: Kingma, M. Welling, Auto-Encoding Variational Bayes I/ International Conference of Learning Representations 2014_ X N 2 2 Encoder 9zlx) Decoder P(xlz) Data: X Reconstruction: * Neural Network Perspective [source] Gunnar Ratsch 15. 3. 2022 29\n\nmedical_ data_ science _ Ezurich Towards interpretable health state representations Idea: Use self-organizing maps to encourage interpretable neighborhood relationship in latent space and smoothness over time Renal Idysfunction Desirable properties Discretellow dimensional Smooth over time Expressive Interpretable Cardiac dysfunction Healthy, https 1 /en wikipedia org/wiki 'Self-organi zing map# /media File Somtraining vg Vincent Fortuin; Matthias Huser; Francesco Locatello, Heiko Strathmann; and Gunnar Ratsch, 2019. SOMVAE: Interpretable discrete representation learning on time series: International Conference on Learning Representations (ICLR) 2019. Gunnar Ratsch 15. 3. 2022 30 DINFK\n\nmedical_ data_ science _ Ezurich SOM-VAE model for discrete time series representations input ^w encoder latent encodings Markov model decoder reconstruction t+1 2 e t 2 2 t+1 4 P(z t+1l2. 4) q self-organizing map t Ze L(xt-1 xt '22 9, &t e) LsoM-VAE (2t 2 24, 18) +yLtransitions rt_1, xt) + t Lsmoothness rt_1 2 xt) LsoM-VAE(T, Sq, Te) Lreconstruction (1, iq, Ze) + a Lcommitment (w) + 8 Lsom (x) Model is jointly optimized (gradient descent wl special handling of discrete states) DINFK Vincent Fortuin; Matthias Hiser; Francesco Locatello, Heiko Strathmann; and Gunnar Ratsch, 2019. SOMVAE: Interpretable_discrete representation learningon_time series: International Conference on Learning Representations (ICLR) 2019. Gunnar Ratsch 15. 3. 2022 31 N\n\nmedical_ data_ science _ Ezurich Health state representations on 2D SOM-grid Method score 6 score 12 score 24 k-means 0. 0411 + 0. 0007 0. 0384 + 0. 0006 0. 0366 = 0. 0005 SOM-VAE 0. 0407 = 0. 0005 0. 0376 + 0. 0004 0. 0354 = 0. 0004 SOM-VAE-prob 0. 0474 + 0. 0006 0. 0444 + 0. 0006 0. 0421 + 0. 0005 Performance comparison of our method wth and_without Markov model (SOM-VAE-prob and SOM-VAE) against k-means in terms of normalized mutual information. Dynamic endpoints are the maximum of the physiology score within the next 6, 12 or 24 hours (6_hours, 12_ hours, 24 hours): Each method is used to fit 64 clusters. Shown are means and standard errors over 10 runs. DINFK Vincent Fortuin; Matthias Hiser; Francesco Locatello, Heiko Strathmann, and Gunnar Ratsch, 2019. SOMVAE: Interpretable discrete Gunnar Ratsch 15. 3. 2022 32 representation learning on _time series: International Conference on Learning Representations (ICLR) 2019.\n\nmedical_ data_ science _ Ezurich Health state representations on 2D SOM-grid 12 10 (a) k-means (b) VQ-VAE (c) SOM-VAE (d) Patient trajectories Color of SOM-cells: Average dynamic APACHE score in the next 24 hours. Higher APACHE score is associated with increased severity of patient health state. vQ-VAE: Van den Oord et al., 2017 (https Ilarxiv orglabs/1711. 00937) Vincent Fortuin Vincent Fortuin, Matthias Hiser; Francesco Locatello, Heiko Strathmann, and Gunnar Ratsch, 2019. SQM-VAE: Interpretablediscrete Gunnar Ratsch 15. 3. 2022 33 representation learning on time series: International Conference on Learning Representations (ICLR) 2019. DINFK\n\nmedical_ data_ science _ Ezurich Health state representations on 2D SOM-grid 0. 2 Ston Eno * Stant End: 016 0 00 0. 06 WW 0 02 Figure 3: Illustration of two example patient trajectories in the SOM grid of T-DPSOM49. One patient died (red), while the other was discharged alive from the ICU (green). Superimposed is a heatmap that displays the mean APACHE score of all time points assigned to each cluster: We observe qualitative differences in the trajectories of the dying and the surviving patient: For each time series, we also show the assigned probabilities to the discrete patient health states using a blue color shading: In this work; we will separate the representations into organ systems and manually annotate different areas of the "map" using medical knowledge in collaboration with intensive care specialists. Laura Manduchi; Matthias Huser; Julia Vogt; Gunnar Ratsch, Vincent Fortuin, DPSOM: Deep Probabilistic Clustering with Self-Organizing Maps, ACM-CHIL, 2020 DINFK\n\nmedical_ data_ science _ 1 Ezurich Contrastive Learning Contrastive learning is a machine learning technique used to learn the general features ofa dataset without labels by teaching the model which data points are similar or different: [Source] Vector Representation 2, Maximize agreement Vector Representation z g() Dense Layer Dense Layer Dense Layer Dense Layer Dense Layer Dense Layer Intermediate Representation h CNN Neural Network Encoder fC) CNN Neural Network Encoder fC) Similar samples < Similar downstream task label Positive Pairs X Data Augmentation x New challenges for Contrastive Learning DINFK Gunnar Ratsch 15. 3. 2022 35\n\nmedical_ data_ science _ 1 Ezurich Contrastive Learning for Time Series Contrastive learning is a machine learning technique used to learn the general features ofa dataset without labels by teaching the model which data points are similar or different: [Source] Contrastive Learning Learn a representation of a patient state at a give time t0 t-th Patient stay Similar samples < Similar downstream task label New challenges for Contrastive Learning Hugo Yeche et al Yeche, Hugo, et al. 'Neighborhood contrastive learning applied t0 online patient monitoring: International Conference on Machine Learning. PMLR, 2021. Gunnar Ratsch DINFK 15. 3. 2022 36\n\nmedical_ data_ science _ Ezurich Challenges in Using Contrastive Learning for ICU monitoring 2N 'exp pi pvli) TT CCL log M i=1 kzi exp Pi pj/t, Alignment Normalization How to construct different views outside of computer vision [1]? Should all "negatives" be treated the same [2]? CV ICU CV ICU Diversity among samples Humanly understandable signal Complex modality Strong similarities among contiguous samples i. i. d distribution of samples Balanced multi-class downstream task multiples samples from a single patient (non ii. d) Usually imbalanced downstream task Relies on strong data augmentation Limits usage of data augmentation Clear hierarchy among negative samples Uniform Normalization [1] Tian et al: (2020) [2] Wang et al. (2020) Gunnar Ratsch 15. 3. 2022 37 DINFK\n\nmedical_ data_ science _ Ezurich Preserve hierarchy in the data with Neighborhood Contrastive Loss How can infuse prior Redefine Contrastive knowledge without relying Loss through the lens of on data augmentation? Neighborhood Two samples share the same neighborhood if they share some predefined attributes: Formally Examples Supervised: n(i, k) = 1if Ti and Tk share attributes 0 else ny(i, k) = 1if yi Yk Temporal nw (i, k) 1ifli kl < W and Si Sk Wang et al:. (2020) DINFK Gunnar Ratsch 15. 3. 2022 38\n\nmedical_ data_ science _ Ezurich Neighborhood Contrastive learning Objective ND NA Aligning neighbors Discriminating neighbors 2N Zi zvli) ZkIn(i, k) = 1 2kIn(i, k) = 0 push away pull towards 2N exp m Pi p)/v) log keN(i) exp (Pi 9k/t) i=1 ~l CNA exp (Pi_ pin /t) = log NG)I M i=1 leN(i) kzi exp (pi aklt) = 3 LNCL aLNA + (1 _ a)LND DINFK Gunnar Ratsch 15. 3. 2022 39 CND\n\nmedical_ data_ science _ 1 Ezurich A Unifying Framework Method U n(, CL 1. 0 0 nw SACL (Cheng et al,, 2020) 0. 0 +00 Tw CLOCS (Kiyasseh et al,, 2020) 1. 0 +00 nw SCL (Khosla et al,, 2020) 1. 0 NA ny We explore two cases NCL: Unsupervised, we call NCL(nw) where: n = nw; W EUR J0, +o [; & EUR J0, 1[ Supervised, we call NCL(ny) where:n ny ; & EUR ]0, 1[ Task Sepsis onset prediction AUROC (in %_ Metric AUPRC (in % _ Utility (x100) Linear MLP Head Linear MLP Linear MLP Seq2-Seq-AE 7. 0 + 0. 3 Seq2-Seq-AE-forecast 6. 6 = 0. 3 CL 7. 9 = 0. 4 SACL (Cheng et al,, 2020) 6. 5 =0. 3 CLOCS (Kiyasseh et al,, 2020) 7. 1 +0. 5 NCL(nw ) (Ours) 8. 2 = 0. 4 7. 8 + 0. 4 7. 3 +0. 3 95 + 0. 4 7. 6 =0. 3 7. 3 +0. 4 93 + 0. 5 77. 1 +0. 5 78. 1+0. 6 75. 8 + 0. 9 76. 9 +0. 5 78. 2 =0. 3 80. 2 + 0. 4 73. 0 = 1. 2 75. 3 +0. 8 77. 2 +05 78. 8 = 0. 4 78. 8 = 0. 3 80. 7 = 0. 3 26. 8 = 1. 0 23. 5 + 1. 5 26. 2 + 0. 8 20. 5: 2. 5 23. 0 =1. 1 27. 2 = 1. 0 27. 2 + 1. 0 23. 8 = 1. 2 29. 7 + 1. 0 24. 2+1. 1 25. 8 + 0. 9 30. 2 = 1. 0 End-to-End SCL (Khosla et al,, 2020) NCL(ny ) (Ours) 7. 6 = 0. 2 6. 7 =0. 6 10. 0 = 0. 5 8. 1 +0. 4 6. 0 + 0. 5 10. 1 + 0. 3 78. 9 + 0. 3 73. 1 +1. 7 80. 3 + 0. 4 78. 8 + 0. 4 70. 0 + 1. 9 80. 8 = 0. 2 27. 9 +0. 8 20. 2 +2. 7 32. 6 + 1. 0 27. 5 1. 0 20. 6 = 1. 7 31. 9 + 0. 9 Experiments on MIMIC-IIl and Physionet 2019 datasets. DINFK Gunnar Ratsch 15. 3. 2022 40\n\nmedical_ data_ science _ Ezurich Summary & Take Home messages Representation learning is a recently developed, powerful tool to learn integrative computational summaries of observed data. Health state data is one interesting application where we assume that the patients physiological health state can be accurately represented in vector form: Autoencoders and forecaster models learn vector representations of past data that are predictive of the past and future, respectively: Generative models are an important tool for finding additional representations and to generate realistic data Contrastive learning can improve learning representations of time series DINFK Gunnar Ratsch 41 |
| 10 | OCR_ML4HLecture05-NLP.pptx_.txt | adc6e224-1ea | OCR | \n\nmedical_ data_ science Ezurich Lecture Machine Learning for Health Care" (261-5120-00L) Basics of Natural Language Processing Gunnar Ratsch, Rita Kuznetsova Biomedical Informatics group, Institute for Machine Learning; Department of Computer Science DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022\n\nmedical_ data_ science Ezurich Outline for today Introductionlmotivation Basic preprocessing steps Basic text features LDA algorithm Embeddings: from BoW to word embedding models POS tagging Language Modelling DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022\n\nmedical_ data_ science Ezurich Towards Comprehensive Patient Models for Precision Medicine Distill Data Embed Data Connect Data Clinical Trial Design Precision Medicine Health Records Pathology Images Genomic Data X=w( p(phenotype x, t) Drugs Predict Treatment Mobile Health = argmax p(survival X, DINFK Gunnar Ratsch 15. 3. 2022 3\n\nmedical_ data_ science Ezurich Usefulness of the clinical texts Classification of the clinical notes Binary: mortality Multiclass: phenotyping Sentiment analysis of the clinical reports is diagnosis confirmed, suspected or excluded Topic modelling diseases, treatment etc Medical Named Entity Recognition diseases, drugs etc Text generation medical report generation (for example based on images) Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 Latent Representation DINFK\n\nmedical_ data_ science Ezurich Problems with clinical texts Ungrammatical, has misspellings and concatenations. Contains short telegraphic phrases, acronyms, abbreviations, which are often overloaded It can contain many things that can be typed or pasted, such as long sets of lab values or vital signs Institution-specific template-use is common Pervasive fear; misunderstanding, and confusion around security, privacy, de-identification, and anonymization => significant barriers to get access Some sections might be long and detailed, other sections might be empty or only contain some captions DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 5\n\nmedical_ data_ science Ezurich Basic Text Processing Most NLP tasks needs to do text preprocessing and normalization: a Segmenting tokenizing words b_ Normalization C. Stop-words removal d. Punctuation removal e Lemmatization Istemming Disclaimer: Natural Language Processing is a huge topic on its own and we can only cover the absolute basics. Check out lectures on Natural Language processing and understanding offered by colleagues in the department: DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022\n\nmedical_ data_ science Ezurich Basic preprocessing steps Tokenization "This is a cat:' 99 "This" "is" "a" "cat" 66 9) Issues: "whatre, Im, isn't" "What are, |am; is not" Ways to do: spacy, nltk, custom tokenizer Normalization The indexed text and the query terms must have the same form: e. g., the US, USA & U. S. A. are the same; could be done with RegExp lowercasing Stop-words removal from nltk. corpus import stopwords ['out' on 'off' 'over' under' 'again' 1 further' then' 'once here there'when'where '.. ] Punctuation removal from string import punctuation DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022\n\nmedical_ data_ science Ezurich Stemming vs Lemmatization Stemming reduce terms to their stems in information retrieval: Is crude chopping of affixes automate(s), automatic, automation automat Consult; consultant; consulting, consultative, consultants consult Lemmatization have to find the correct dictionary headword form; use of a vocabulary and morphological analysis of words: Lemma: same stem, part of speech a lemma is the dictionary form of a set of words (headword): cat and cats = same lemma (cat) run; runs, ran, running = same lemma (run) Reduce inflections or variant forms to base form Car; cars, car's, cars car Am, are, is 3 be DINFK NLP Stanford Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022\n\nmedical_ data_ science Ezurich Bag of Words (BoW) How to map the term representation into vector space? One way to do this is to represent documents as the bag (multiset) of its words, disregarding grammar and even word order; but keeping multiplicity "No, my brain did not hurt 9} "Perhaps it was more exasperating this way than if it had."66 would have preferred it to hurt me. We assign every word w from vocabulary Wa one-hot vector: Vw [0, 0, 1, 0,., 0] e RII W 0 The document is represented as d = {w_1, w_2, "J w_n}, then we could assign the vector for the document d: Vd = Vw wed vocab size W O1oo, 000) 0 1 "dog 0 DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022\n\nmedical_ data_ science Ezurich Bag of Words (BoW) & Term Frequency How to map the term representation into vector space? One way to do this is to represent documents as the bag (multiset) of its words, disregarding grammar and even word order; but keeping multiplicity "No, my brain did not hurt 9} "Perhaps it was more exasperating this way than if it had."would have preferred it to hurt me. Term Frequency (TF): raw count of term in document: number of times that term t occurs in document d. {"no" 1, "my": 1, "brain": 1, "did": 1, "not": 1, "hurt": 1} {"Perhaps": 1, "it": 2, 66 was": 1, "more": 1, "exasperating": 1, "this": 1, "way": 1, "if: 1, 66 "had': 1} {": 1, 66 would": 1, "have' 1, "preferred": 1, 66 "it": 1, "to": 1, "hurt": 1, "me": 1} W) Easy to use bag of word representation for vanilla ML models (like SVMs) as inputs (example) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 10\n\nmedical_ data_ science Ezurich TF-IDF (term frequency-inverse document frequency) TF suffers from a critical problem: all terms are considered equally important: In fact, certain terms have little or no discriminating power in determining relevance. How to scale down the term weights? (Simple idea: scale down the term weights of terms with high frequency, defined to be the total Inumber of occurrences of a term in the document collection: Document Frequency (IDF): the number of documents in the collection that contain a term t TF-IDF: assigns to term t a weight in document d that is highest when t occurs many times within a small number of documents (thus lending high discriminating power to those documents) lower when the term occurs fewer times in a document; or occurs in many documents (thus offering a less pronounced relevance signal); lowest when the term occurs in virtually all documents DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 11\n\nmedical_ data_ science Ezurich TF-IDF (term frequency-inverse document frequency) nt TF (t, d) k nk nt the number of times that term t occurs in document d. Denominator the total number of terms in document d. IDI IDF(t, D) = log Kdi EUR D lt e di } Usage: similarity computation feature vector for the classifier (as a baseline) IDl total number of documents in the collection Kd; e D l t e d } number of documents from the collection, where the term t appears, "t = 0. TF-IDF = TF(t, d)* IDF (t, D) TF-IDF values could be obtained with sklearn feature extraction. text. TfidfVectorizer DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 12\n\nmedical_ data_ science Ezurich Latent Dirichlet Allocation (LDA) Topics What is the Topic Modelling? geae 8. 82 genetic 0. 01 Word: element in vocabulary set; W Document: collection of words, d (BoW assumption) life 0. 02 evolve 0. 01 Corpus: collection of documents, D organism 0. 01 Documents Topic proportions and assignments Seeking Life's Bare (Genetic) Necessities COl SPix HARBOR; Vnl FORK nalI "Fectak Hou {mcnWthe 72 Lnme"the dunl Saks t lun "Wuak Urt #itctnt ArFr 'ch'Fnentol Cl Wi U cuulelre: Wetd. Ju" IItIun ( #urh tT= t cnr canluku Wa mtd 4 ~llte Ieuh ~ua I Thetulie he c#l Uk nce MTT Ute T" Arcu: Muhun: nen htum7e et > Icuu Fuky' JeN R Mult RuuueeI k a IntruE tkut Iueuntit R theMnln/ Cmfirin: Go J Fkns m Juthe wat Alhw_h ohe nmker: dent 1 J UEuch Mtl. I~ Frehcts Topic: collection of words (elements from vocabulary set) Document is represented by latent mixture of topics: p(wld) p(wlz)p(zld) where z topic. Rearon 8. 82 nerve 0. 01 Gonomo appinp and SoqvancCcld Sprng Hor Ncx Yok Mey 8 lo 12 Stripping down @omeuter ana y vieis @nes = Iatr ofina Minimum Nocain and &n0 EURnt Denonio; data 0. 02 nunber 0. 02 computer 0. 01 C'IENLE Ml : Mi Each topic is a distribution over words Each document is a mixture of corpus-wide topics Each word is drawn from one of those topics Material from Tutorial by David Blei, 2012 For given corpus, we learn: 1_ Topic: from full vocabulary important subsets 2_ Topic proportion: for each document what is about? Vocabulary ~ basis vector extraction (topics) represent d in topic space (topic proportion) (topic proportion could be used as a feature vector for downstream tasks) Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 13 DINFK\n\nmedical_ data_ science Ezurich Latent Dirichlet Allocation LDA is a generative probabilistic model for collections of discrete data such as text corpora. Proportions Per-word n EUR RV, a EUR RK are the model parameters parameter topic assignment Per-document Observed Assume there are K topics across all the documents: topic proportions word Material from Tutorial by David Blei, 2012 Topic parameter Topics For k in (1, K): choose per-corpus topic distribution Bk e RV ~ Dir(n) For d in (1, D): choose per-document topic proportion 0d EUR RK ~ Dir(a) 04 Bk K n Za;n Wa, n N D For each word w_n: choose topic Zd, n EUR Lx Multinomial(0a) choose word Wd, n eZv Multinomial(Wd, nlzd, m Bk) p(B, 0, 2, wla, n) = K N P(B;In) [ [r(oala) Mp(zd, n| Oa)p(Wd, nl B1. K, _ Zd, n) d=1 a is the parameter of the Dirichlet prior on the per-document topic distributions B is the parameter of the Dirichlet prior on the per-topic word distribution is the topic distribution for document m mn is the topic for the n-th word in document m W is the specific word mn DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 14\n\nmedical_ data_ science _ Ezurich From basic text features to word embeddings Goal: to map each word to the vector: BoW and One-hot: Easy to build; Big dimensionality; Do not reflect the relationship of words in the text: Distributional hypothesis: words that are occur in the same contexts have similar meanings: DINFK Harris (1954). Distributional structure Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 15\n\nmedical_ data_ science _ Ezurich NeurIPS 2013 Distributed Representations of Words and Phrases and their Compositionality word2vec Tomas Mikolov Google Inc Mountain View mikolov@google com Ilya Sutskever Google Inc. Mountain View ilyasu@google com Kai Chen Google Inc _ Mountain View kai@google com arXiv 2013 Greg Corrado Jeffrey Dean Google Inc. Mountain View jeff@google com Efficient Estimation of Word Representations in Vector Space Tomas Mikolov Google Inc.. Mountain View, CA tmikolov@google com Kai Chen Google Inc. Mountain View, CA kaichen@google com Greg Corrado Jeffrey Dean Google Inc, Mountain View, CA gcorrado@google com Google Inc, Mountain View, CA jeff@google com DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 16 https: / /code. google. com/p/ordZvecl\n\nmedical_ data_ science _ 1 Ezurich word2vec The basic idea is to predict a missing word from a context: What is context? Example: "the quick red fox jumps over the lazy brown dog' 1_ Continuous bag of words (CBOW) the 'context' is the sum (or mean) of vectors appearing in the sentence. Based on this context we predict the central' word. dog fox brown quick lazy the jumps red the over 2 Skip-gram the 'context' is each word from the surrounding central word. Based on the central word we predict each word from this surrounding: jumps the jumps quick jumps DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 17 dog\n\nmedical_ data_ science _ Ezurich CBOW VS Skip-gram Continuous bag of words (CBOW) Skip-gram INPUT PROJECTION OUTPUT INPUT PROJECTION OUTPUT w(t-2) w(t-2) w(t-1) w(t-1) SUM w(t) w(t) w(t+1) w(t+1) N log P(w;b i-1 wi-k' wi+k) i=1 N k log P(Wi+ilw;) i=1 j=-k j=O w(t+2) w(t+2) Objective: maximise log-probability DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 18\n\nmedical_ data_ science _ Elzurich Skip-gram model Back to the optimisationl Maximise this: N k log P(Wi+jlwi_ i=1 j=-k j=O This prime is importantl T W0 UwI exp p(wolwi) W T w=1 exp vW Uw[ For each word, J learn two representations: 1. as the context jumps {the quick red fox over the lazy brown dog} Distributed Representations of Words and Phrases and their Compositionality Mikolov; Sutskever; Chen, Corrado, Dean, NIPS 2013 Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 19 2. as the word itself DINFK\n\nmedical_ data_ science _ Ezurich Paradigmatic relations This distinction is importantl "the quick red fox jumps over the lazy brown dog" "the quick red fox leaps over the lazy brown dog 'leaps' and 'jumps' are similar because they fit the context: We don't expect to see them occurring together in a sentence, because they have a paradigmatic relationship. (This means they're somehow interchangeable:) The Distributional Hypothesis Magnus Sahlgren, PhD dissertation (2006) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 20\n\nmedical_ data_ science _ Ezurich Country and Capital Vectors Projected by PCA China Beijing Russias Japans Moscow Ankara 3 Tokyo Turkey 2 1. 5 0. 5 Polandk Germany France Warsaw Berlin Paris0. 5 Italy Greece Athens Rome Spains Madrid Lisbon1. 5 Portugal 22 2150. 5 0. 5 1. 5 2 figure 2 from Mikolov et al. II Distributed Representations of Words and Phrases and their Compositionality NIPS 2013 Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 21 DINFK\n\nmedical_ data_ science _ Ezurich So does it work? Tested vectors using an analogical reasoning task A : B as X : Y'? (e. g: 'puppy is to dog as kitten is to cat) This means asking: vec(A) 5 vec(B) 2 vec(X) vec(Y) Or 2 vec(A) 5 vec(B) + vec(Y) vec(X) Where, as mentioned before, the 'similarity' here is cosine similarity. DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 22\n\nmedical_ data_ science _ Ezurich So does it work? Created a test set with ~9k semantic questions and ~11k syntactic Examples: calm calmly safe safely Athens Greece Japan big bigger small smaller old oldest best Poland zloty Hungary forint move moving fly flying Austin Texas Honolulu Hawaii Ireland Irish Egypt Egyptian dancing danced saying said girl uncle aunt man men cat cats (these are actually most of the categories) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 23 Tokyo good boy\n\nmedical_ data_ science _ Ezurich So does it work? Table 4: Comparison of publicly available word vectors on the Semantic-Syntactic Word Relationship test set; and word vectors from our models: Full vocabularies are used Model Vector Training Accuracy [%] Dimensionality words Semantic Syntactic Total Collobert-Weston NNLM 50 660M 9. 3 12. 3 11. 0 Turian NNLM 50 37M 1. 4 2. 6 2. 1 Turian NNLM 200 37M 1. 4 2. 2 1. 8 Mnih NNLM 50 37M 1. 8 9. 1 5. 8 Mnih NNLM 100 37M 3. 3 13. 2 8. 8 Mikolov RNNLM 80 320M 4. 9 18. 4 12. 7 Mikolov RNNLM 640 320M 8. 6 36. 5 24. 6 Huang NNLM 50 990M 13. 3 11. 6 12. 3 neural net Our NNLM 20 6B 12. 9 26. 4 20. 3 Our NNLM 50 6B 27. 9 55. 8 43. 2 language model Our NNLM 100 6B 34. 2 64. 5 50. 8 CBOW 300 783M 15. 5 53. 1 36. 1 Skip-gram 300 783M 50. 0 55. 9 533 yes! this takes 2 weeks DINFK on 180 cores! 2. 5 days on 125 cores Efficient Estimation of Word Representations in Vector Space' Mikolov, Chen; Corrado, Dean, arXiv 2013\n\nmedical_ data_ science _ Ezurich Analogical reasoning The oft-cited example of successful analogical reasoning is: vec(king) vec(man) vec(woman) = vec(queen) Intuitively; this means vec(king) vec(man) vec("sense of royalty") Which is a pretty fascinating idea: What if. _ vec(gleevec) vec(leukemia)~ vec("treatment")? Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 25 DINFK\n\nmedical_ data_ science _ Ezurich Analogical reasoning Then we could do.. vec(melanoma) vec("treatment") = vec("?? 2") This would certainly be useful for a medical Jeopardy. _ It's also the kind of information we want our embeddings to encode 9 for enabling medical language processing: So we ran wordZvec on some MSKCC text. DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 26\n\nmedical_ data_ science _ Elzurich An example vec(melanoma) [vec(gleevec) vec (leukemia)] =? 22 Top 10 closest results to melanoma: Top 10 closest results to gleevec: Top 10 closest results to Zeukemia: 3 me Lanoma 340 ~2. 22044604925e-16 gLeevec crel 2148 1. 11022302463e-16 Leukemia 1129 ~2. 22044604925e-16 dfsp 12007 0. 156969851964 dasatinib 4299 0. 0436408836789 itp A3744 0. 216802690111 neurotropic 17723 0. 18892856986 imatinib 2596 0. 0444753136031 myelodysplasia 8348 0. 220480414542 neurotrophic 22261 0. 193823458626 nilotinib 5211 0. 0565038545667 cll 1270 0. 229098946201 SCC 4457 0. 199310695303 anagrelide 6961 0. 0806562061213 aml 2236 0. 232815740704 amelanotic 10967 0. 205920966083 hydroxyurea 3921 0. 0824481117079 cmmol 8659 0. 236774032333 choroidal 9357 0. 208689368781 gleevac 16087 0. 0843472720016 mds 2858 0. 23788044386 fibrosarcoma 8679 0. 223705333644 ruxolitinib 11279 0. 0845686271568 coexisting 16242 0. 241202871701 eccrine Jl13344 0. 22791107347 "ieeeie nexavar 7350 0. 0862700069714 Leukemia/sll 35544 0. 245014293364 fibrohistiocytoma 11045 0. 239171736259| hydrea 6445 0. 100871239337 Igl Bxa10616 0. 246571161984 cancer/ 27891 0. 243011714361 afinitor 10465 0. 10846339025 hypogammaglobulinemia 6544 0. 249632640917 Top 10 closest results to UNM+ G-L and we get:.. CMLIALL ponatinib 14236 0. 42982079986 diascopy 23369 0. 435802215172 #I#th 20379 0. 44739503288 eruption 3763 0. 447999880188 gleevac 16087 0. 448643975007 nexavar 7350 0. 452329959825 hive 18576 0. 455971559975 pustule 11842 0. 455989743044 gleevec Ae2148'0. 458117608185 dabrafenib 10448 0. 459866794241 desatinib 32409 0. 46007721604 typo :) sorafenib (kidneylliver cancer drug) (BRAF-mutated, metastatic) MELANOMAI DINFK CMLIALL\n\nmedical_ data_ science _ Ezurich Other embedding techniques 1. Glove [1] aggregated global word-word co-occurrence statistics from a corpus. 2. FastText [2] ~ training process with character n-grams, helps with OOV (Out of vocabulary) problem: 1 GloVe: Global Vectors for Word Representation; Pennington, J. Socher; R,, & Manning; C. D."EMNLP 2014. 2. Enriching_Word_Vectors_with_Subword Information; Bojanowski, P, Grave, E., Joulin, A"& Mikolov, T. TACL, 2017. DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 28\n\nmedical_ data_ science Ezurich Part of Speech (PoS) Tagging Given a sentence W. Wn and a tagset of lexical categories, find the most likely tag t-tn for each word in the sentence Penn_Treebank PQSTags] SecretariatINNP isIVBZ expected/VBN toITO race/VB tomorrow/NN Problem: Many of the words may have unambiguous tags But enough words are either ambiguous or unknown non-trivial task Brown corpus is a general corpus in corpus linguistics (500 samples of English texts, ~IM words). Most words in English have only one Brown Corpus tag: unambiguous (1 tag) 35, 340 words Many of the most common words are ambiguous (over 40% tokens are ambiguous) Obvious strategies may be suggested based on intuition: to/TO race/VB thelDT racelNN Methods: simple baseline with unigrams hardcoded rules vs supervised unsupervised ML approaches. NNP Proper noun, singular; VBZ Verb, 3rd person singular present; VBN Verb, past article; VB verb, base form, NN Noun; singular or mass: DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 29\n\nmedical_ data_ science Ezurich Simplest strategy Choose the most likely tag for each ambiguous word, independent of previous words assign each token the POS category it occurred as most often in the training set This strategy gives 90% accuracy in controlled tests Which POS is more likely in a corpus (1, 273, 000 tokens)? race: NN: 400 VB: 600 P(NNlrace) = P(race, NN) / P(race) by the definition of conditional probability P(race) = 1000/1, 273, 000 =. 0008 P(race, NN) = 400/1, 273, 000 =. 0003 P(race, VB) = 600/1, 273, 000 =. 0004 SO we obtain: P(NNIrace) = P(race, NN)P(race) =. 0003/. 0008 =. 375 P(VBlrace) = P(race, VBYP(race) =. 0004/. 0008 =. 5 DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 30\n\nmedical_ data_ science Ezurich HMMs: A Probabilistic Approach We want to find the "best sequence' 9) of PoS tags T=T_ T for & sentence 1' W=W_ 1"W n' where T; is the PoS tag for word W In other words we want to find a PoS tags T that maximizes P(TIW) Using Bayes' Rule, we can say P(WIT) * P(T) P(TIW) P(W) We want to find the value of T which maximizes the right hand side. note: denominator can be discarded (same for every T) Find T which maximizes P(WIT) * P(T) DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 31\n\nmedical_ data_ science Elzurich Independence Assumptions Assume that current event is based only on previous n-1 events n 1 P(T1, _, Tn) IIP(TIT;_1) i=1 assumes that the event of a PoS tag occurring is independent of the event of any other PoS tag occurring, except for the immediately previous PoS tag from a linguistic standpoint; this seems an unreasonable assumption, due to long-distance dependencies 2_ P(W1.. WnITi. Tn) II P(WAT;) i1 assumes that the event of a word appearing in a category is independent of the event of any surrounding word or tag, except for the tag at this position: Linguists know both these assumptions are incorrect nevertheless, statistical approaches based on these assumptions work pretty well for part-of-speech tagging: Hidden Markov Models (HMMs) is widely used in both PoS-tagging and speech recognition, among other problems DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 32\n\nmedical_ data_ science Ezurich PoS Tagging Based on HMM Problem: Find T which maximizes P(W | T) * P(T) Here W=W, 1W n and T=T Tn Using the HMM; we get: Transition probabilities (prob. of transitioning from one stateltag to another): n P(T1, Tn) IIP(T;IT;_1) i=1 Emission probabilities (prob. of emitting a word at a given state): P(W1. WnITi.. Tn) II P(wt;) i] We want to find the value of T_ 1"~T which maximizes: n n II P(T;) * P(TilTi-1) {=1 DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 33\n\nmedical_ data_ science Ezurich POS probabilities: transitions P(T1, _ '' 1 Tn) ~ IIP(TIT;_1) 0. 8 i=1 1. 0 PRP MD 04 NN 0. 6 0. 3 "He will race J} Possible tag series for T = T1'T2'T3 T =PRP MD NN T = PRP NN NN T = PRP MD VB T =PRP NN VB POS bigram probabilities from training corpus can be used for P(T) 0. 2 NN 0_ B PRP: personal pronoun MD: modal (auxiliary verb) NN: noun VB Verb, base form Which is the most likely path? DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 34\n\nmedical_ data_ science Ezurich Lexical generation probabilities n From the training corpus, we need to find the T; which maximizes: [[ P(WAT;) * P(TilT;_1) C E i] 0. 4 0. 8 MD NN A B 1Q PRP 0. 6 0. 3 0. 2 NN VB 0_ C Emission Probabilities 0. 4 willMD 0. 8 racelNN 0. 4 MD NN VB PRP 0. 8 A B 0. 6 he S | $ 10 helPRP 1. 0 will 0. 8 0. 2 0. 3 0. 22 0. 7 race 0 0. 4 0. 6 willINN 0. 2 racelVB 0. 6 Note 1: sum over column should sum up to 1, the MLE of the emission probability is how many times this word W appears as this tag t, divided by how many times we saw the tag t in training data: DINFK Note 2: the whole table extends further down than is depicted, as there are other words in the dictionary whose emission probability we're not interested in for this sentence. Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 35\n\nmedical_ data_ science Ezurich Using Dynamic Programming To find the most likely sequence of categories for a sequence of words, we don't need to enumerate all possible sequences of categories. Due to the Markov assumption, if you keep track of the most likely sequence found so far for each possible ending category; you can ignore all the other less likely sequences: multiple edges coming into a state, but only keep the value of the most likely path, i. e. use of DP The algorithm to do this is called the Viterbi algorithm: DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 36\n\nmedical_ data_ science Ezurich Viterbi Algorithm recap (informal) 1_ Assume we are at state h in the HMM: a States H, _H m all come into h 2. Obtain a the best probability of each previous state H_Hm b the transition probabilities: P(hIH;); P(hHm) C the emission probability for word w at h: P(wlh) 3_ Multiply the probabilities for each new path: Best(l) = Maxh <h [Best(H)* P(hlH)]* P(wlh) 4_ One of these states (H_ 1"H m) will give the highest probability Only keep the highest probability when using h for the next state E 0 willlM D racelN N 0. 4 0_ 8 A B 1Q helPR P S | 0 03 0 0_ 2 Find the most likely sequencel DINFK willIN N 82 0. 7 racelV B 0. 6 F\n\nmedical_ data_ science Ezurich Introduction to the Transformer model Transformer is the encoder-decoder architecture with self-attention: The key component of the Encoder is Self-Attention mechanism. Layer normalization Input data: X e Rnxd Encoder overview: Q-queries, K keys, V values: linear projections of the input data X, d dimension: # QKT Latent Representation softmax )V Residual 1 Va connections Add & Normalize Feed Forward Feed Forward Projection layer / Add & Normalize Self-Attention Attention weights Positional embeddings POscoDMG Self attention Positional Embeddings is used to learn the order in the sequence. Residual connections mitigates vanishing gradients. Layer normalization helps with numerical stability: POSITIONAL ENCODING EMBEDDINGS X1 Pictures from The Ilustrated transformer_Jay Alammar DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 38 Vaswani, Ashish, et al. "Attention is allyou need"Advances in neural information processing systems 30 (2017).\n\nmedical_ data_ science Ezurich Language Modeling Deep Learning Language Model (LM) computes the probability of p(w_1, w_n) for any W_1, _ w_n = V (vocabulary) How we do LM in modern world? state-of-the art Transformer-based models BERT Input [CLS] my is cute [SEP] he likes play ##ing [SEP] Pretrained WordPiece embeddings Token Embeddings E [CLS] E my E "dog E cute E 'play [SEP] ~likes [SEP] A marker indicating Sentence A or Sentence B Segment Embeddings EA EB To learn about the position in sentence Position Embeddings Eo E Ez E4 E5 E6 E7 Eg E1o 1 [CLS] token is added in the beginning a special classification token, the final hidden state corresponding to this token used as the aggregate sequence representation for classification tasks. 2_ [SEP] token is inserted at the end of each sentence. One sentence also can feed as input. 3_ Both sentences A and B are encoded with the help of Byte-pair encoding (BPE): BPE is a simple form of data compression in which the most common pair of consecutive bytes of data is replaced with a byte that does not occur within that data. ZabdZabac DINFK Devlin et al. 'Bert: Pre-training of_deep bidirectional transformers for language understanding arXiv preprint arXiv:1810. 04805 (2018). Z-aa dog 8 &ne E#ring 4 4 4 4 4 8 8 8 8 S 8\n\nmedical_ data_ science Elzurich Training details 1. Masked-LM: replace n% words in the input by special [MASK] token predict these tokens variants altering Structural DNA [MASK] can modify gene function by [MASK] transcript sequences 2. Next sentence prediction: binarized task; from the [CLS] token need to predict whether one sentence follows another in the text. Class 1: Cancer is a group of diseases involving abnormal cell growth It has the potential to invade or spread to other parts of the body: Class 0: Cancer is a group of diseases involving abnormal cell growth A bear is sunbathing: The basis of the BERT architecture = encoder of the Transformer model Way of using BERT: have pretrained model finetune for the required task on the specific corpora DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 40\n\nmedical_ data_ science Ezurich BERT in biomedical domain Data: MMC I clinical notes, PubMed abstracts, papers from Semantic Scholar etc Tasks: Named Entity Recognition, Q&A, Sentence Similarity etc Hugging Face Search models, datasets, users__ Models Datasets Spaces Docs Solutions Pricing Tasks Models 296 bio Fill-Mask Question Answering dmis-lab/biobert-base-cased-V1. 1 Updated Oct 14, 2020 1. 01M Summarization Table Question Answering Text Classification Text Generation Text2Text Generation 88 Token Classification emilyalsentzer/Bio_ClinicalBERT Fill-Mask Updated 16 days ag0 533k 14 Translation Zero-Shot Classification 218 Sentence Similarity +14 microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext Fill-Mask Updated Sep 22, 2021 121k 19 Libraries PyTorch TensorFlow JAX 4 24 microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract Fill-Mask Updated Sep" 22, 2021 85. 2k Datasets wikipedia common_voice squad dmis-lab/biobert-V1. 1 Feature Extraction Updated May19, 2021 bookcorpus c4 glue conll2003 73k 0 4 dcep europarl jrc-acquis 828 bionlp/bluebert_pubmed_mimic_uncased_L-12_H-768_A-12 Updated Sep 24, 2021 62,. 6k Gunnar Ratsch & Rita Kuznetsova Languages DINFK 22. 3. 2022 41\n\nmedical_ data_ science _ Ezurich Summary & Take Home messages Clinical reports contain relevant information about patient's health state, but are written in challenging, specialized language Standard basic text processing is often applied: tokenization, stemming, lemmatization, Latent Dirichlet Allocation (LDA) is a probabilistic model for text content using topics Word embedding are a recently developed powerful tool in NLP Can be learned unsupervisedly and represent semantic in a vector space Part of Speech Tagging is an import task in NLP that is often solved with HMMs Recent NLP techniques use deep learning and have shown great promise DINFK Gunnar Ratsch & Rita Kuznetsova 22. 3. 2022 42 |
| 11 | OCR_PAPER_dall-e-2-annotated_.txt | 3f42d484-d96 | OCR_academic_paper | \n\nHierarchical Text-Conditional Image Generation with CLIP Latents Aditya Ramesh OpenAI aramesh@openai _ com Prafulla Dhariwal OpenAI prafullaCopenai com Alex NicholOpenAI alexCopenai. com 8 2 2 3 7 1 Casey Chu OpenAI caseyCopenai com Mark Chen OpenAI markCopenai com Abstract Contrastive models like CLIP have been shown to learn robust representations of images that capture both semantics and style. To leverage these representations for image generation; we propose a two-stage model: prior that generates CLIP image embedding given a text caption, and decoder that generates an image conditioned on the image embedding: We show that explicitly generating image representations improves image diversity with minimal loss in photorealism and caption similarity Our decoders conditioned on image representations can also produce variations of an image that preserve both its semantics and style, while varying the non-essential details absent from the image representation Moreover_ the joint embedding space of CLIP enables languageguided image manipulations in zero-shot fashion. We use diffusion models for the decoder and experiment with both autoregressive and diffusion models for the prior; finding that the latter are computationally more efficient and produce higher-quality samples. L Introduction Recent progress in computer vision has been driven by scaling models o large datasets of captioned images collected from the internet [10, [44, 60, [39, B1/[16]. Within this framework, CLIP [39] has emerged as successful representation learner for images CLIP embeddings have number of desirable properties: they are robust to image distribution shift; have impressive zero-shot capabilities, and have been fine-tuned to achieve state-of-the-art results o a wide variety of vision and language tasks [45] Concurrently, diffusion models [46][48, /25/ have emerged as promising generative modeling framework, pushing the state-of-the-art on image and video generation tasks [11 26, 24]. To achieve best results, diffusion models leverage a guidance technique 11 241 which improves sample fidelity (for images, photorealism) at the cost of sample diversity: In this work; we combine these two approaches for the problem of text-conditional image generation We first train a diffusion decoder to invert the CLIP image encoder. Our inverter is non-deterministic and can produce multiple images corresponding to given image embedding: The presence of an encoder and its approximate inverse (the decoder allows for capabilities beyond text-to-image translation. As in GAN inversion [62, [55] encoding and decoding an input image produces semantically similar output images (FigureB} We can also interpolate between input images by inverting interpolations of their image embeddings (Figure[H: However; one notable advantage of using the CLIP latent space is the ability to semantically modify images by moving in the direction of any encoded text vector (FigureE); whereas discovering these directions in GAN latent space involves Equal contribution\n\nVibrant portrait painting of Salvador Dali with a robotic half face shiba inu wearing beret and black turtleneck a close up of a handpalm with leaves growing from it espresso machine that makes coffee from human souls, artstation panda mad scientist mixing sparkling chemicals artstation 1 2 corgi '$ head depicted as an explosion of a nebula propaganda poster depicting cat dressed french emperor napoleon holding a piece of cheese dolphin in an astronaut suit on saturn; artstation teddy bear on a skateboard in times square Figure 1: Selected 1024 X 1024 samples from production version of our model.\n\nCLIP objective img encoder "a corgi playing a flame text throwing encoder trumpet" decoder prior Figure 2: A high-level overview of unCLIP. Above the dotted line, we depict the CLIP training process, through which we learn a joint representation space for text and images. Below the dotted line, we depict our text-to-image generation process: CLIP text embedding is first fed to an autoregressive Or diffusion prior to produce an image embedding, and then this embedding is used to condition diffusion decoder which produces a final image. Note that the CLIP model is frozen during training of the prior and decoder luck and diligent manual examination: Furthermore, encoding and decoding images also provides US with a tool for observing which features of the image are recognized or disregarded by CLIP To obtain a full generative model of images, we combine the CLIP image embedding decoder with a prior model, which generates possible CLIP image embeddings from given text caption. We compare our text-to-image system with other systems such as DALL-E [40] and GLIDE [35], finding that our samples are comparable in quality to GLIDE; but with greater diversity in Our generations We also develop methods for training diffusion priors in latent space, and show that they achieve comparable performance to autoregressive priors, while being more compute-efficient We refer to ur full text-conditional image generation stack as unCLIP, since it generates images by inverting the CLIP image encoder: 2 Method Our 'training dataset consists of pairs (c, y) of images EUR and their corresponding captions y. Given an image EUR, let Zi and Zt be its CLIP image and text embeddings, respectively: We design our generative stack to produce images from captions using two components: A prior P(zily) that produces CLIP image embeddings z; conditioned on captions yA decoder P(clzi, y) that produces images conditioned on CLIP image embeddings zi (and optionally text captions y) The decoder allows US to invert images given their CLIP image embeddings, while the prior allows us t0 learn generative model of the image embeddings themselves. Stacking these two components yields a generative model P(zly) of images z given captions y: P(cly) = P(w, zily) = P(wlzi, y)P(zily). The first equality holds because Zi is a deterministic function of EUR. The second equality holds because of the chain rule _ Thus, we can sample from the true conditional distribution P(zly) by first sampling Zi using the 3\n\nprior, and then sampling EUR using the decoder: In the following sections, we describe our decoder and prior stacks. For training details and hyperparameters, refer to Appendix[ 2. 1 Decoder We use diffusion models [25, [48 to produce images conditioned on CLIP image embeddings (and optionally text captions). Specifically, we modify the architecture described in Nichol et al 2021) by projecting and adding CLIP embeddings to the existing timestep embedding, and by projecting CLIP embeddings into four extra tokens of context that are concatenated t0 the sequence of outputs from the GLIDE text encoder: We retained the text conditioning pathway present in the original GLIDE model, hypothesizing that it could allow the diffusion model to learn aspects of natural language that CLIP fails to capture (e. g variable binding), but find that it offers little help in this regard (Section[}. While we can sample from the conditional distribution of the decoder directly, past work using diffusion models shows using guidance on the conditioning information [11J[24JB5] improves sample quality a lot: We enable classifier-free guidance [24 by randomly setting the CLIP embeddings to zero (Or learned embedding) 10% of the time, and randomly dropping the text caption 50% of the time during training: To generate high resolution images, we train tWo diffusion upsampler models [34, 443]: one to upsample images from 64 X 64t0 256 x 256 resolution, and another to further upsample those to 1024 1024 resolution_ To improve the robustness of our upsamplers, we slightly corrupt the conditioning images during training: For the first upsampling stage, we use gaussian blur [43/, and for the second we use a more diverse BSR degradation [42 [59j. To reduce training compute and improve numerical stability, we follow Rombach et al] [42 and train on random crops of 'images that are one-fourth the target size. We use oly spatial convolutions in the model (i. e,, no attention layers) and at inference time directly apply the model at the target resolution, observing that it readily generalizes to the higher resolution We found no benefit from conditioning the upsamplers on the caption, and use unconditional ADMNets with no guidance 2. 2 Prior While a decoder can invert CLIP image embeddings zi to produce images EUR, we need a prior model that produces Zi from captions y to enable image generations from text captions. We explore two different model classes for the prior model: Autoregressive (AR) prior: the CLIP image embedding zi is converted into sequence of discrete codes and predicted autoregressively conditioned on the caption y. Diffusion prior: The continuous vector Zi is directly modelled using a Gaussian diffusion model conditioned on the caption y: In addition to the caption, we can condition the prior on the CLIP text embedding 2t since it is deterministic function of the caption. To improve sample quality we also enable sampling using classifier-free guidance for both the AR and diffusion prior; by randomly dropping this text conditioning information 1O9 of the time during training: To train and sample from the AR prior more efficiently, we first reduce the dimensionality of the CLIP image embeddings zi by applying Principal Component Analysis (PCA) [377 In particular; we find that the rank of the CLIP representation space is drastically reduced when training CLIP with SAM [15] while slightly improving evaluation metrics We are able to preserve nearly all of the informatiorZby retaining only 319 principal components out of the original 1, 024. After applying PCA, we order the principal components by decreasing eigenvalue magnitude, quantize each of the 319 dimensions into 1, 024 discrete buckets, and 2Le., less than 1% average mean-squared error in reconstructing the image representations\n\n7 Figure 3: Variations of an input image by encoding with CLIP and then decoding with a diffusion model_ The variations preserve both semantic information like presence of a clock in the painting and the overlapping strokes in the logo, as well as stylistic elements like the surrealism in the painting and the color gradients in the logo, while varying the non-essential details predict the resulting sequence using a Transformer [53[ model with a causal attention mask. This results in threefold reduction in the number of tokens predicted during inference, and improves training stability We condition the AR prior on the text caption and the CLIP text embedding by encoding them as a prefix to the sequence Additionally, we prepend a token indicating the (quantized) dot product between the text embedding and image embedding, Zi Zt. This allows uS t0 condition the model on higher dot product, since higher text-image dot products correspond to captions which better describe the image. In practice, we find it beneficial to sample the dot product from the top half of the distributiong] For the diffusion prior; we train decoder-only Transformer with a causal attention mask on sequence consisting 0f, in order: the encoded text, the CLIP text embedding; an embedding for the diffusion timestep, the noised CLIP image embedding, and final embedding whose output from the Transformer is used to predict the unnoised CLIP image embedding We choose not to condition the diffusion prior on Zi Zt like in the AR prior; instead_ we improve quality during sampling time by generating two samples of zi and selecting the one with a higher dot product with Zt_ Instead of using the EUR-prediction formulation from Ho et al. ] [251, we find it better t0 train our model to predict the unnoised Zi directly, and use a mean-squared error loss on this prediction: Lprior E, ~[1, 7]+O~q [Ifo(-{ t, y) 2ll2] 'We swept over percentiles S0%, 709, 8S9, 959 and found S0% to be optimal in all experiments. 5\n\n0 Figure 4: Variations between two images by interpolating their CLIP image embedding and then decoding with a diffusion model. We fix the decoder seed across each row. The intermediate variations naturally blend the content and style from both input images. 3 Image Manipulations Our approach allows us to encode any given image EUR into a bipartite latent representation (Zi; TT) that is sufficient for the decoder to produce an accurate reconstruction. The latent Zi describes the aspects of the image that are recognized by CLIP while the latent TT encodes all of the residual information necessary for the decoder t0 reconstruct x. The former is obtained by simply encoding the image with the CLIP image encoder: The latter is obtained by applying DDIM inversion (Appendix F in [11]) to x using the decoder; while conditioning 0n Zi. We describe three different kinds of manipulations that are enabled by this bipartite representation_ 3. 1 Variations Given an image EUR, we can produce related images that share the same essential content but vary in other apects, such as shape and orientation (FigureB]. To do this, we apply the decoder to the bipartite representation Zi, TT) using DDIM with n 0 for sampling: With n = 0, the decoder becomes deterministic and will reconstruct the given image EUR. Larger values of introduce stochasticity into successive sampling steps, resulting in variations that are perceptually "centered" around the original image EUR. As n increases, these variations tell us what information was captured in the CLIP image embedding (and thus is preserved across samples), and what was lost and thus changes across the samples).\n\nphoto of a cat an anime drawing of a super saiyan cat; artstation photo of a victorian house photo of a modern house photo of an adult lion photo of lion cub photo of a landscape in winter a photo of a landscape in fall Figure 5: Text diffs applied to images by interpolating between their CLIP image embeddings and a normalised difference of the CLIP text embeddings produced from the two descriptions. We also perform DDIM inversion to perfectly reconstruct the input image in the first column, and fix the decoder DDIM noise across each rOW. 3. 2 Interpolations Itis also possible to blend tWo images 11 and T2 for variations (Figure[4Hh; traversing all of the concepts in CLIP' $ embedding space that occur between them To do this, we rotate between their CLIP embeddings Zi1 and #i2 using spherical interpolation, yielding intermediate CLIP representations Zio slerp(zi1Zi2 0) as 0 is varied from 0 to 1. There are two options for producing the intermediate DDIM latents along the trajectory: The first option involves interpolating between their DDIM inverted latents TT1 and TT2 (by setting TTe slerp(TT; TTz 0)), which yields a single trajectory whose endpoints reconstruct T1 and T2 The second option involves fixing the DDIM latent to a randomly-sampled value for all interpolates in the trajectory: This results in an infinite number of trajectories between T1 and. 2, though the endpoints of these trajectories will generally no longer coincide with the original images. We use this approach in Figurel 3. 3 Text Diffs key advantage of using CLIP compared to other models for image representations is that it embeds images and text to the same latent space, thus allowing uS to apply language-guided image manipulations (i. e. _ text diffs), which we show in Figure[5] To modify the image to reflect a new text description y, we first obtain its CLIP text embedding 2t, as well as the CLIP text embedding #t0 of caption describing the current imagq We then compute a text diff vector 2d norm( Zt 2t0 from these by taking their difference and Instead of a description of the current image, we also experimented with using dummy caption like photo for the baseline, Or removing it altogether: These also worked well_\n\nPizz) Feize Ap A Ppck Oc @ Pabl Piot Piz Prpo Ppite PPEale PMAz Granny Smith: 100% iPod: 0% Pizza: 0% Granny Smith: 0. 02% iPod: 99. 98% Pizza: 09 Granny Smith: 94. 33% iPod: 0% Pizza: 5. 669 Figure 6: Variations of images featuring typographic attacks [20] paired with the CLIP model'$ predicted probabilities across three labels. Surprisingly, the decoder still recovers Granny Smith apples even when the predicted probability for this label is near O%. We also find that our CLIP model is slightly less susceptible to the "pizza' attack than the models investigated in [20]. normalizing: Now; we can rotate between the image CLIP embedding Zi and the text diff vector 2d using spherical interpolation, yielding intermediate CLIP representations 2o slerp(zi, 2d, 0 ), where 0 is increased linearly from 0 to a maximum value that is typically in [0. 25, 0. 50]. We produce the final outputs by decoding the interpolates 20, fixing the base DDIM noise t0 TT throughout the entire trajectory: Probing the CLIP Latent Space Our decoder model provides unique opportunity to explore CLIP latent space by allowing US to directly visualize what the CLIP image encoder is seeing: As an example use case, we can revisit cases where CLIP makes incorrect predictions, such as typographic attacks [201. In these adversarial images, a piece of text is overlayed on top of an object; which causes CLIP to predict the object described by the text rather than the object depicted in the image. This piece of text essentially hides the original object in terms of output probabilities In Figure[] we show an example of this attack from [20], wherein an apple can be misclassified as an iPod_ Surprisingly, we find that our decoder still generates pictures of apples with high probability even though the predicted probability of "Granny Smith' is near zero. Even more notable, the model never produces pictures of iPods, despite the very high relative predicted probability of this caption_ iPod Piza PIzA LLC Fpxae 1zlri\n\nFigure 7: Visualization Of reconstructions of CLIP latents from progressively more PCA dimensions (20, 30, 40, 80, 120, 160, 200, 320 dimensions ), with the original source image on the far right: The lower dimensions preserve coarse-= grained semantic information, whereas the higher dimensions encode finer-grained details about the exact form of the objects in the scene: PCA reconstructions offer another tool for probing the structure of the CLIP latent space. In Figure[] we take the CLIP image embeddings of a handful of source images and reconstruct them with progressively more PCA dimensions, and then visualize the reconstructed image embeddings using our decoder with DDIM on fixed seed_ This allows us t0 see what semantic information the different dimensions encode. We observe that the early PCA dimensions preserve coarse-grained semantic information such as what types of objects are in the scene, whereas the later PCA dimensions encode finergrained detail such as the shapes and exact form of the objects. For example, in the first scene, the earlier dimensions seem to encode that there is food and perhaps a container present; whereas the later dimensions encode tomatoes and bottle specifically Figure[] also serves as visualization of what the AR prior is modeling, since the AR prior is trained to explicitly predict these principal components in this order: 5 Text-to-Image Generation 5. 1 Importance of the Prior Although we train a prior to generate CLIP image embeddings from captions, the prior is not strictly necessary for caption-to-image generation: For instance, our decoder can condition on both CLIP image embeddings and captions, but the CLIP image embedding is dropped 59 of the time during training in order to enable classifier-free guidance. Therefore, at sampling time, we can condition on only the caption, although this underperforms model trained fully in this way (this model is GLIDE, and we do a thorough comparison with GLIDE in Sections[ 2]and[3}. Another possibility is to feed the decoder the CLIP text embedding as if it were an image embedding, as previously observed [61J[54]. The first two rows of Figure[8]depicts samples obtained in these two ways; the third row depicts samples obtained with prior: Conditioning the decoder on just the caption is clearly worst; but conditioning On text embeddings zero-shot does produce reasonable results_ Building on this observation, another approach would be to train the decoder to condition on CLIP text embeddings [9] instead of CLIP image embeddings (although we would lose the capabilities mentioned in Section[) To quantify the effectiveness of these alternate approaches, we train two models: small decoder conditioned on CLIP text embeddings, and a small unCLIP stack (diffusion prior and decoder). We then compare samples from the text-embedding decoder; samples from the UnCLIP stack, and samples obtained from feeding text\n\nJ 1 F 1 1 ~A group of baseball an oil painting of a hedgehog using a ~A motorcycle parked in a 'This wire metal rack players is crowded at corgi wearing a calculator" parking space next to holds several pairs of the mound paty hat" another motorcycle. shoes and sandals' Figure &: Samples using different conditioning signals for the same decoder: In the first IOW, We pass the text caption to the decoder; and pass a zero vector for the CLIP embedding: In the second row, we pass both the text caption and the CLIP text embedding of the caption. In the third IOW, we pass the text and CLIP image embedding generated by an autoregressive prior for the given caption_ Note that this decoder is only trained to do the text-to-image generation task (without the CLIP image representation) S% of the time. embeddings to the unCLIP decoder zero-shot; sweeping across guidance scales for all models_ We find that these approaches respectively score FIDs of 9. 16, 7. 99, and 16. 55 on a test set, suggesting the unCLIP approach is best: We also run human evaluations comparing the first two settings, sweeping over sampling hyperparameters for each using our human evaluation proxy model (Appendix[A) We find that humans prefer the full unCLIP stack 57. 0% = 3. 19 of the time for photorealism and 53. 19 E 3. 1% of the time for caption similarity: Given the importance of the prior; it is worth evaluating different approaches for training it. We compare both the AR and diffusion priors throughout our experiments_ In all cases Sections[. 2155. 4 and[5), we find that the diffusion prior outperforms the AR prior for comparable model size and reduced training compute. 5. 2 Human Evaluations We observe in FigurefJthat unCLIP is capable of synthesizing complex, realistic images. While we can compare sample quality to past models using FID, it is not always aligned with human judgment To better gauge the generation capabilities of our system; we conduct systematic human evaluations comparing UnCLIP to GLIDE for photorealism, caption similarity, and sample diversity We follow the protocol of Ramesh et al ] Nichol et al. 40, B5_ for the first two evaluations: for photorealism, users are presented with pairs of images and must choose which looks more photorealistic; for caption 10\n\n3 unCLIP GLIDE Figure 9: Samples when increasing guidance scale for both unCLIP and GLIDE, using the prompt; A green vase filled with red roses sitting on top of table."For UnCLIP, we fix the latent vectors sampled from the prior; and only vary the guidance scale of the decoder: For both models, we fix the diffusion noise seed for each column. Samples from unCLIP improve in quality (more realistic lighting and shadows) but do not change in content aS we increase guidance scale, preser= ving semantic diversity even at high decoder guidance scales unCLIP Prior Photorealism Caption Similarity Diversity 47. 19 = 3. 1% 41. 19 E 3. 0% 62. 6% = 3. 0% 48. 99 = 3. 1% 45. 39 = 3. 0% 70. 59 1 2. 8% AR Diffusion Table 1: Human evaluations comparing UnCLIP to GLIDE We compare to both the AR and diffusion prior for UnCLIP: Reported figures are 95% confidence intervals of the probability that the UnCLIP model specified by the row beats GLIDE: Sampling hyperparameters for all models were wept to optimize an automated proxy for human photorealism evaluations. similarity, users are additionally prompted with a caption, and must choose which image better matches the caption In both evaluations, there is a third "Not sure" option For diversity, we propose a new evaluation protocol in which humans are presented with two 4 X 4 grids of samples and must choose which is more diverse (with a third option; *Not sure") For this evaluation; we produce sample grids using 1, 000 captions from the MS-COCO validation set, and always compare sample grids for the same caption. Before running human comparisons, we swept over sampling hyperparameters for each model using a CLIP linear probe trained to be a proxy for human photorealism evaluations (Appendix[}: These hyperparameters are fixed across all three types of evaluation: We present our results in Table/l In general, the diffusion prior performs better than the AR prior in pairwise comparisons against GLIDE. We find that humans still slightly prefer GLIDE to unCLIP in terms of photorealism, but the gap is very small. Even with similar photorealism, unCLIP is strongly preferred over GLIDE in terms of diversity highlighting one of its benefits.\n\n3 80% 3 70% 1 60% unCLIP is better 50% 2 GLIDE is better 3 40% 30% in terms of photorealism [ + in terms of caption similarity 20% in terms of diversity 1. 0 1. 5 2. 0 2. 5 3. 0 GLIDE guidance scale Figure 10: When comparing UnCLIP (with our best sampling settings) to various settings of guidance scale for GLIDE; unCLIP was preferred by human evaluators on at least one axis among photorealism, caption similarity, and diversity for each comparison: At the higher guidance scales used to generate photorealistic images, UnCLIP yields greater diversity for comparable photorealism and caption similarity 18 216 8 14 12 GLIDE unCLIP (AR) unCLIP (Diffusion) 10 1. 0 1. 5 2. 0 2. 5 3. 0 Guidance Scale 3. 5 4. 0 Figure ]l: FID versus guidance scale for UnCLIP and GLIDE. FOr the UnCLIP priors, we swept over sampling hyperparameters and fixed to the settings with the best minimum FID. 5. 3 Improved Diversity-Fidelity Trade-off with Guidance Compared to GLIDE, we qualitatively observe that unCLIP is able to generate more diverse images while leveraging the guidance technique to improve sample quality To understand why, consider Figure [9]where we increase guidance scale for both GLIDE and unCLIP. For GLIDE, the semantics camera angle, color; size) converge aS we increase guidance scale, whereas for unCLIP the semantic information of the scene is frozen in the CLIP image embedding and therefore does not collapse when guiding the decoder: In Section/5. 2 we observed that unCLIP achieves similar photorealism as GLIDE while maintaining more diversity, but that its caption matching capabilities were slightly worse. It is natural to ask whether GLIDE'$ guidance scale can be lowered to obtain the same diversity level as unCLIP while maintaining better caption 12\n\nModel FID Zero-shot FID Zero-shot FID (filt) AttnGAN Xu et al. 2017 DM-GAN Lhu et al. 2019 DF-GAN Tao et al_ 2020= DM-GAN + CL Ye et al 2021 XMC-GAN Zhang et al. |2021 LAFITE Zhou et al_ 2021 Make-A-Scene Gatn et al_ 2022 DALL-E Ramesh et al. 120217 LAFITE Zhou et al. 12021 GLIDE Nichol et al 2021 Make-A-Scene Gatni et al_ 2022 unCLIP AR prior) unCLIP Diffusion prior) 35. 49 32. 64 21. 42 20. 79 9. 33 8. 12 7. 55 28 26. 94 12. 24 12. 89 11. 84 11. 08 10. 87 10. 63 10. 39 Table 2: Comparison of FID on MS-COCO 256 X 256. We use guidance scale 1. 25 for the decoder for both the AR and diffusion prior; and achieve the best results using the diffusion prior: matching: In Figure[o we conduct a more careful study of this question by performing human evaluations across several GLIDE guidance scales. We find that GLIDE at guidance scale 2. 0 is very close to the photorealism and caption similarity of unCLIP, while still producing less diverse samples. Finally, in Figure[ we compute MS-COCO zero-shot FID [23p while sweeping over guidance scale for both unCLIP and GLIDE, finding that guidance hurts the FID of unCLIP much less so than for GLIDE. In this evaluation; We fix the guidance scale of the unCLIP prior and only vary the guidance scale of the decoder: This is another indication that guidance hurts the diversity of GLIDE much more than unCLIP; since FID heavily penalizes non-diverse generations. 5. 4 Comparison on MS-COCO In the text-conditional image generation literature, it has become standard practice t0 evaluate FID on the MS-COCO [28_ validation set We present results on this benchmark in Tablel] Like GLIDE and DALL-E, unCLIP is not directly trained on the MS-COCO training set; but can still generalize to the validation set zero-shot. We find that, compared to these other zero-shot models, unCLIP achieves a new state-of-the-art FID of 10. 39 when sampling with the diffusion prior: In Figure[2] we visually compare unCLIP to various recent text-conditional image generation models on several captions from MS-COCO. We find that; like the other methods, unCLIP produces realistic scenes that capture the text prompts. 5. 5 Aesthetic Quality Comparison We additionally perform automated aesthetic quality evaluations comparing UnCLIP to GLIDE: Our goal with this evaluation is t0 assess how well each model produces artistic illustrations and photographs. To this end, we generated 512 'artistic" captions using GPT-3 [4] by prompting it with captions for existing artwork (both real and Al generated) Next; we trained CLIP linear probe to predict human aesthetic judgments using the AVA dataset [33| (AppendixAA For each model and set of sampling hyperparameters, We produce four images for each prompt, and report the mean predicted aesthetic judgment over the full batch of 2048 images In Figure[3] we present results 0n our aesthetic quality evaluation: We find that guidance improves aesthetic quality for both GLIDE and unCLIP. For unCLIP we only guide the decoder (we found that guiding the prior hurt results). We also plot the aesthetic quality against Recalf5] since guidance typically induces trade-off SRecall is computed with respect to the training dataset: 13\n\n1 2 7 8 1 8 1 8 a green train is coming down the tracks a group of skiers are preparing to ski down a mountain_ a small kitchen with a lOw ceiling' a group of elephants walking in muddy water: living area with a television and table" Figure 12: Random image samples on MS-COCO prompts_ 14\n\n4. 85 0. 600 0. 575 L 4. 80 0. 550 4. 75? 0. 525 < 4. 70 GLIDE 0. 500 GLIDE 1 4. 65 unCLIP (AR) 0. 475 unCLIP (AR) unCLIP (diffusion) unCLIP (diffusion) 4. 60 0. 450 1. 0 1. 5 2. 0 2. 5 3. 0 3. 5 4. 0 4. 60 4. 65 4. 70 4. 75 4. 80 4. 85 guidance scale mean AVA prediction Figure 13: Aesthetic quality evaluations comparing GLIDE and unCLIP using 512 auto-generated artistic prompts. We find that both models benefit from guidance, but unCLIP does not sacrifice recall for aesthetic quality: between fidelity and diversity: Interestingly; we find that guiding UnCLIP does not decrease Recall while still improving aesthetic quality according to this metric. Related Work Synthetic image generation is a well studied problem, and most popular techniques for unconditional image generation have also been applied to the text-conditional setting Many previous works have trained GANs [21] on publicly available image captioning datasets to produce text-conditional image samples 56, 63_ 49, [58/57]. Other works have adapted the VQ-VAE approach [52] to text-conditional image generation by training autoregressive transformers on sequences of text tokens followed by image tokens 40, [12[1|. Finally, some works have applied diffusion models to the problem, training either continuous 351 or discrete 221 diffusion models with auxiliary text encoders to handle textual input: Previous works have leveraged hierarchical generative processes to create high-quality synthetic images_ Razavi et al. 41 trains multi-layer discrete autoencoder; allowing them to first sample coarse-grained latent codes and then use this as conditioning information when sampling higher-resolution latent codes. Child [Vahdat and Kautz [5J[50] generate images using VAEs with a hierarchy of latent codes that increase progressively with resolution. Concurrently with our work, Gafni et al. ][17| conditions a generative image model on segmentation masks, allowing for a generative process that first samples a semantic map of an image and then conditions the generated image O this information: The computational benefits of using diffusion to model latent space has been noted by previous works Preechakul et al. 38 propose an autoencoder framework where diffusion models are used to render latent variables as images, and a second diffusion model is used to generate these latents (similar to our diffusion prior) [Vahdat et al. ][51] use a score-based model for the latent space of a VAE, whileRombach et al. 42] use diffusion models on the latents obtained from VQGAN [14] like autoencoder: Since its release, CLIP [39_ has been used extensively to steer generative image models towards text prompts_ Galatolo et al. Patashnik et al. Murdock Gal et al. ] [19, 16, B2/, [18/ guide GANs using gradients from CLIP model. For diffusion models, Dhariwal and Nichol 11| introduced classifier guidance as a way to use gradients from classifier trained on noised images to steer the model towards higher quality generations. Nichol et al:. 35 train a CLIP model on noised images and guide text-conditional diffusion model while Crowson Crowson [71/8 use an unnoised CLIP model to guide unconditional or class-conditional diffusion models _ Ho and Salimans 24] introduced classifier-free guidance and showed that one can perform guidance 15\n\nunCLIP GLIDE Figure 14: Samples from unCLIP and GLIDE for the prompt a red cube on top of a blue cube implictly from the predictions of the model with and without the conditioning information, thus removing the need for a classifier: Nichol et al. [35 | showed classifier-free guidance works more favorably than CLIP guidance for text conditional image generation: Several previous works have trained generative image models that are directly conditioned on CLIP embeddings Zhou et al:] 61] condition GAN models on randomly perturbed CLIP image embeddings, finding that these models can generalize to CLIP text embeddings to produce text-conditional images Crowson [9] trained diffusion models conditioned on CLIP text embeddings, allowing for direct text-conditional image generation: Wang et all] 54] train an autoregressive generative model conditioned o CLIP image embeddings, finding that it generalizes to CLIP text embeddings well enough to allow for text-conditional image synthesis Bordes et al ] [3] train diffusion models conditioned on image representations from contrastive models While the diffusion models themselves cannot generate images unconditionally, the authors experimented with a simple approach for two-stage image generation by employing Kernel Density Estimation to sample image representations By feeding these generated representations to the diffusion model, they can generate imas ges end-to-end in a way similar t0 our proposed technique. However; our work differs from this in two ways: first, we use multimodal contrastive representations rather than image-only representations; second, we employ much more powerful generative models for the first stage of the generation hierarchy, and these generative models are conditioned on text: 7 Limitations and Risks Although conditioning image generation on CLIP embeddings improves diversity, this choice does come with certain limitations In particular; UnCLIP is worse at binding attributes t0 objects than a corresponding GLIDE model_ In Figure[4 we find that unCLIP struggles more than GLIDE with prompt where it must bind two separate objects (cubes) to two separate attributes (colors)_ We hypothesize that this occurs because the CLIP embedding itself does not explicitly bind attributes to objects, and find that reconstructions from the decoder often mix up attributes and objects, as shown in Figure[[5} A similar and likely related issue is that unCLIP 16\n\nFigure 15: Reconstructions from the decoder for difficult binding problems_ We find that the reconstructions mix up objects and attributes In the first two examples, the model mixes up the color of two objects In the rightmost example, the model does not reliably reconstruct the relative size of two objects Deinp Lerpt: Diep PONEELH Deep Figure 16: Samples from unCLIP for the prompt; "A sign that says deep learning:' struggles at producing coherent text; as illustrated in Figure[6} it is possible that the CLIP embedding does not precisely encode spelling information of rendered text This issue is likely made worse because the BPE encoding we use obscures the spelling of the words in caption from the model, so the model needs to have independently seen each token written out in the training images in order to learn to render it: We also note that our stack still has hard time producing details in complex scenes (Figure[7. We hypothesize that this is a limitation of our decoder hierarchy producing an image at a base resolution of 64 X 64 and then upsampling it Training our unCLIP decoder at a higher base resolution should be able to alleviate this, at the cost of additional training and inference compute_ As discussed in the GLIDE paper; image generation models carry risks related to deceptive and otherwise harmful content: unCLIP' $ performance improvements also raise the risk profile over GLIDE. As the technology matures, it leaves fewer traces and indicators that outputs are Al-generated, making it easier to mistake generated images for authentic ones and vice versa. More research is also needed on how the change in architecture changes how the model learns biases in training data. 17 Deep\n\nA high quality photo of a dog playing in a green field next to a lake. b) high quality photo of Times Square_ Figure 17: unCLIP samples show low levels of detail for some complex scenes The risks of these models should be assessed in relation to the particular deployment context; which includes training data, guardrails in place, the deployment space, and who will have access: A_ preliminary analysis of these issues in the context of the DALLE 2 Preview platform (the first deployment of an unCLIP model), can be found in Mishkin et al. [30]. 8 Acknowledgements We 'd like to thank Jong Wook Kim, Hyeonwoo Noh, Alec Radford, Pranav Shyam, and Ilya Sutskever for helpful discussions and contributions t0 our work: We'd also like to thank Yunxin Jiao for creating several figures used in the paper: We are grateful to the Acceleration and Supercomputing teams at OpenAL for their work o software and hardware infrastructure this project used: 18\n\nReferences [1] Armen Aghajanyan, Bernie Huang; Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, and Luke Zettlemoyer: CM3: A Causal Masked Multimodal Model of the Internet. arXiv:2201. 07520, 2022 [2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang: Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models_ CoRR, abs/2201. 06503, 2022. URL https / [arxiv_ org abs_ 2201. 06503 [3] Florian Bordes, Randall Balestriero, and Pascal Vincent High Fidelity Visualization of What Your Self-Supervised Representation Knows About: arXiv:2112. 09164 2021. [4] Tom B. Brown, Benjamin Mann, Nick Ryder; Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel HerbertVoss, Gretchen Krueger; Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler; Jeffrey Wu; Clemens Winter; Christopher Hesse, Mark Chen, Eric Sigler; Mateusz Litwin, Scott Gray; Benjamin Chess, Jack Clark, Christopher Berner; Sam McCandlish, Alec Radford, Ilya Sutskever; and Dario Amodei. Language Models are Few-Shot Learners arXiv:2005. 14165 2020. [5] Rewon Child. Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images arXiv:2011. 10650, 2021. [6] Katherine Crowson_ AVA Linear Probe. https [twitter com/RiversHaveW_ Jings /status 14723461867281735687s-20&t=T HRr3GwSHRGj QaMDtRe3a 2021. [7] Katherine Crowson: CLIP guided diffusion HQ 256x256. https I/colab. research. google _ com drive/12a_Wrfi2_gwwluN3VvMTwVMz9TfqctNj 2021. [8] Katherine Crowson. CLIP Guided Diffusion 512x512, Secondary Model Method. https I [twitter com/RiversHavew Jings, status 1462859669454536711/ 2021. [9] Katherine Crowson V-diffusion_ https: github. com_ crowsonkb/v-diffusion-pytorch, 2021. [10] Karan Desai and Justin Johnson: VirTex: Learning Visual Representations from Textual Annotations arXiv:2006. 06666, 2020. [11] Prafulla Dhariwal and Alex Nichol. arXiv:2105. 05233, 2021. Diffusion Models Beat GANs on Image Synthesis. [12] Ming Ding, Zhuoyi Yang; Wenyi Hong; Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang: Cog View: Mastering Text-to-Image Generation via Transformers_ arXiv:2105. 13290, 2021. [13] Alexey Dosovitskiy, Lucas Beyer; Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner; Mostafa Dehghani, Matthias Minderer; Georg Heigold, Sylvain Gelly, Jakob Uszkoreit; and Neil Houlsby: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale: arXiv:2010. 11929, 2020. [14] Patrick Esser; Robin Rombach, and Bjorn Ommer: Taming Transformers for High-Resolution Image Synthesis arXiv: 2012. 09841 2020. [15] Pierre Foret, Ariel Kleiner; Hossein Mobahi, and Behnam Neyshabur: Sharpness-Aware Minimization for Efficiently Improving Generalization arXiv:2010. 01412 2020. 19\n\n[16] Andreas Fiirst, Elisabeth Rumetshofer; Viet Thuong Tran, Hubert Ramsauer; Fei Tang; Johannes Lehner; D P Kreil, Michael K Kopp; Giinter Klambauer; Angela Bitto-Nemling, and Sepp Hochreiter: CLOOB: Modern Hopfield Networks with InfoLOOB Outperform CLIP, 2022. URL https / /openreview net / forum? id-qw674LIPfQE [17] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman Make-AScene: Scene-Based Text-to-Image Generation with Human Priors arXiv:2203. 13131 2022. [18] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or: StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators_ arXiv: 2108. 00946, 2021. [19] Federico A. Galatolo, Mario G. C. A. Cimino, and Gigliola Vaglini. Generating images from caption and vice versa via CLIP-Guided Generative Latent Space Search: arXiv:2102. 01645, 2021. [20] Gabriel Goh; Nick Cammarata Chelsea Voss t, Shan Carter; Michael Petrov, Ludwig Schubert, Alec Radford, and Chris Olah: Multimodal Neurons in Artificial Neural Networks_ Distill, 2021. doi: 10. 23915/distill. 00030. https:/ Ildistill-pub/202l/multimodal-neurons [21] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu; David Warde-Farley, Sherjil Ozair; Aaron Courville, and Yoshua Bengio. Generative Adversarial Networks arXiv: 1406. 2661 2014. [22] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang; Dongdong Chen, Lu Yuan, and Baining Guo. Vector Quantized Diffusion Model for Text-to-Image Synthesis. arXiv:2 111. 14822 2021. [23] Martin Heusel, Hubert Ramsauer; Thomas Unterthiner; Bernhard Nessler; and Sepp Hochreiter: GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. Advances in Neural Information Processing Systems 30 NIPS 2017), 2017. [24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In NeurIPS 202 1 Workshop on Deep Generative Models and Downstream Applications, 2021. URL https '/openreview net_ forum? id-qw8AKxfYbI [25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. arXiv:2006. 11239, 2020. [26] Jonathan Ho, Chitwan Saharia, William Chan, David J. Fleet; Mohammad Norouzi, and Tim Salimans_ Cascaded Diffusion Models for High Fidelity Image Generation. arXiv:2106. 15282, 2021. [27] Diederik P Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. arXiv:1412. 6980 2014 [28] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev; Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C Lawrence Zitnick, and Piotr Dollar: Microsoft COCO: Common Objects in Context. arXiv:1405. 0312 2014. [29] Ilya Loshchilov and Frank Hutter: Decoupled Weight Decay Regularization: arXiv: 1711. 05101 2017. [30] Pamela Mishkin, Lama Ahmad, Miles Brundage, Gretchen Krueger; and Girish Sastry. DALLE 2 Preview Risks and Limitations 2022_ URL https github. com/openai/dalle-2-preview/ blob/main/system-card. md [31] Norman Mu; Alexander Kirillov, David Wagner; and Saining Xie. SLIP: Self-supervision meets Language-Image Pre-training. arXiv:2 112. 12750, 2021. [32] Ryan Murdock The Big Sleep. https: / /twitter com_ advadnoun status_ 1351038053033406468, 2021. 20\n\n[33] Naila Murray, Luca Marchesotti, and Florent Perronnin: AVA: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408-2415, 2012. doi: 10. 1109/CVPR. 2012. 6247954_ [34] Alex Nichol and Prafulla Dhariwal Improved Denoising Diffusion Probabilistic Models: arXiv:2102. 09672, 2021. [35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever; and Mark Chen_ GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. arXiv:2112. 10741 2021. [36] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or; and Dani Lischinski. StyleCLIP: TextDriven Manipulation of StyleGAN Imagery. arXiv:2103. 17249, 2021. [37] Karl Pearson. LIII. On lines and planes of closest fit to systems of points in space, November 1901. URL https I Idoi org 10. 1080/14786440109462720 [38] Konpat Preechakul, Nattanat Chatthee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion Autoencoders: Toward Meaningful and Decodable Representation. arXiv:2111. 15640, 2021_ [39] Alec Radford Jong Wook Kim, Chris Hallacy, Aditya Ramesh; Gabriel Goh; Sandhini Agarwal, Girish Sastry; Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever: Learning Transferable Visual Models From Natural Language Supervision: arXiv:2103. 00020 2021. [40] Aditya Ramesh, Mikhail Pavlov; Gabriel Goh, Scott Gray; Chelsea Voss, Alec Radford, Mark Chen; and Ilya Sutskever Zero-Shot Text-to-Image Generation. arXiv:2102. 12092, 2021. [41] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating Diverse High-Fidelity Images with VQ-VAE-2. arXiv:1906. 00446, 2019. [42] Robin Rombach; Andreas Blattmann, Dominik Lorenz, Patrick Esser; and Bjorn Ommer: HighResolution Image Synthesis with Latent Diffusion Models. arXiv:2112. 10752, 2021. [43] Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, and Mohammad Norouzi _ Image Super-Resolution via Iterative Refinement arXiv:arXiv:2104. 07636, 2021. [44] Mert Bulent Sariyildiz, Julien Perez. and Diane Larlus. Learning Visual Representations with Caption Annotations_ arXiv:2008. 01392, 2020. [45] Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach; Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer How Much Can CLIP Benefit Vision-and-Language Tasks? arXiv:2107. 06383, 2021. [46] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep Unsupervised Learning using Nonequilibrium Thermodynamics arXiv: 1503. 03585, 2015. [47] Jiaming Song; Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models arXiv: 2010. 02502, 2020. [48] Yang Song and Stefano Ermon. Improved Techniques for Training Score-Based Generative Models arXiv:2006. 09011 2020. [49] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing; Fei Wu; and Bingkun Bao. DF-GAN: Deep Fusion Generative Adversarial Networks for Text-to-Image Synthesis arXiv:2008. 05865, 2020. [50] Arash Vahdat and Jan Kautz. NVAE: A Deep Hierarchical Variational Autoencoder: arXiv:2007. 03898, 2020. 21\n\n[51] Arash Vahdat, Karsten Kreis, and Jan Kautz_ Score-based Generative Modeling in Latent Space. In Neural Information Processing Systems (NeurIPS), 2021. [52] Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu: Neural Discrete Representation Learning: arXiv: 1711. 00937, 2017. [53] Ashish Vaswani, Noam Shazeer; Niki Parmar; Jakob Uszkoreit; Llion Jones, Aidan N. Gomez, Lukasz Kaiser; and Illia Polosukhin: Attention Is All You Need_ arXiv:1706. 03762, 2017. [54] Zihao Wang; Wei Liu, Qian He, Xinglong Wu, and Zili Yi. CLIP-GEN: Language-Free Training of a Text-to-Image Generator with CLIP. arXiv:2203. 00386, 2022. [55] Weihao Xia, Yulun Zhang: Yujiu Yang; Jing-Hao Xue, Bolei Zhou, and Ming-Hsuan Yang: GAN Inversion: A Survey arXiv:2101. 05278, 2021. [56] Tao Xu, Pengchuan Zhang; Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He_ AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks arXiv: 1711. 10485, 2017. [57] Hui Ye, Xiulong Yang, Martin Takac, Rajshekhar Sunderraman, and Shihao Ji. Improving Text-to-Image Synthesis Using Contrastive Learning: arXiv:2107. 024232021. [58] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang: Cross-Modal Contrastive Learning for Text-to-Image Generation: arXiv:2101. 04702 2021. [59] Kai Zhang; Jingyun Liang; Luc Van Gool, and Radu Timofte. Designing a Practical Degradation Model for Deep Blind Image Super-Resolution. 2021 IEEEICVF International Conference on Computer Vision (ICCV), Oct 2021. doi: 10. 1109/iccv48922. 2021. 00475. URL http: / /dx doi org 10. 1109/ ICCV48922. 2021. 00475 [60] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P Langlotz. Contrastive Learning of Medical Visual Representations from Paired Images and Text arXiv:2010. 00747, 2020. [61] Yufan Zhou; Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer; Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. LAFITE: Towards Language-Free Training for Text-to-Image Generation. arXiv:2111. 13792, 2021. [62] Jun-Yan Zhu, Philipp Krahenbihl, Eli Shechtman, and Alexei A. Efros Generative Visual Manipulation on the Natural Image Manifold. arXiv:1609. 03552, 2016. [63] Minfeng Zhu; Pingbo Pan; Wei Chen; and Yi Yang DM-GAN: Dynamic Memory Generative Adversarial Networks for Text-to-Image Synthesis arXiv: 1904. 01310, 2019. 22\n\nLinear Probes for Evaluations For our evaluations, we leverage two new linear probes on top of a CLIP ViT-L/I4 [13 model To automate aesthetic quality evaluations, we follow the procedure used by Crowson [6], training a linear regression model on images and mean ratings from the AVA dataset [33 To reduce the cost of hyperparameter sweeps before conducting human evaluations, we train logistic regression model to predict win probabilities between pairs of images To train this model, we used 15, 000 pairwise image comparisons gathered from all of Our previous human evaluations_ For each comparison i, we computed CLIP image embeddings Ti and yi for the two images in the pair: We then trained a linear model f (x) such that 1/(1 + exp (f(Ti) f(yi))) approximates the probability that a human prefers the image for Yi: This can be reduced to a logistic regression problem with inputs equal t0 Yi Ti _ B Error Bars for Human Evaluation When computing error bars for human evaluations, we use the normal approximation interval with p = 0. 95. We expect the normal approximation to be accurate for such a large sample size of n 1000. Training Details The unCLIP models used for the experiments in this paper were trained with the hyperparameters described below, unless otherwise noted_ We additionally trained production version of unCLIP using similarly sized models but with modified architectures and trained for longer; we include changes to accommodate product and safety requirements (e. g. inpainting, preventing unwanted memorization), and train On a larger dataset that is filtered for aesthetic quality and safety: We report model and training hyperparameters for the paper models in Table/3 All models were trained using Adam [27 | with corrected weight decay 29| and momentum 81 0. 9_ Our CLIP model uses ViT-H/16 [13| image encoder that consumes 256 X 256 resolution images, and has width 1280 with 32 Transformer [53| blocks_ The text encoder also follows the architecture described in Radford et al. 39 it is Transformer [53 with a causal attention mask; with width 1024 and 24 Trans former blocks. Both models are trained with learning rate 3 10 and SAM [15] with p = 0. 1, where the perturbations are applied independently by the replicas, each of which uses batch size 64_ The remaining hyperparameters are the same as those reported in Radford et al, 39]_ When training the encoder; we sample from the CLIP [39| and DALL-E 40] datasets (approximately 650M images in total) with equal probability: When training the decoder; upsamplers, and prior; we use only the DALL-E dataset [40] (approximately 250M images). Incorporating the noisier CLIP dataset while training the generative stack negatively impacted sample quality in our initial evaluations_ Our decoder architecture is the 3. 5 billion parameter GLIDE model with the same architecture and diffusion hyperparameters as in Nichol et al. ] 351_ We train with learned sigma and sample with 250 strided sampling steps as in Nichol and Dhariwal [34|We use the ADMNet architecture |11 for the upsamplers_ In the first upsampling stage, we use a cosine noising schedule, 320 channels and depth of 3 resblocks per resolution inside the ADMNet We also apply gaussian blur (kernel size 3, sigma 0. 6) as described in Saharia et al. 431In the second upsampling stage, we use a linear noising schedule, 192 channels, a depth of 2 resblocks per resolution, and train with the BSR degradation from Rombach et al. [42]. Neither upsampler uses attention. To reduce inference time, we use DDIM 47 and manually tune the number of steps, with 27 steps for 256 x 256 model, and 15 steps for the 1024 X 1024 model. 23\n\nFor the AR prior; we use Transformer text encoder with width 2048 and 24 blocks and a decoder with causal attention mask width 1664, and 24 blocks. For the diffusion prior; we use Transformer with width 2048 and 24 blocks, and sample with Analytic DPM [2] with 64 strided sampling steps. To reuse hyperparameters tuned for diffusion noise schedules on images from Dhariwal and Nichol [H1 we scale the CLIP embedding inputs by 17. 2 to match the empirical variance of RGB pixel values of ImageNet images scaled to [~1, 1]: AR prior Diffusion prior 64 64 256 256 1024 Diffusion steps 1000 1000 10oO 100o Noise schedule cosine cosine cosine linear Sampling steps 64 250 27 15 Sampling variance method analytic [2| learned 134 DDIM [47| DDIM [473 Crop fraction 0. 25 0. 25 Model size IB 3. SB 700M 300M Channels 512 320 192 Depth Channels multiple 1, 2, 3, 4 1, 2, 3, 4 1, 1, 2, 2, 4, 4 Heads channels 64 Attention resolution 32, 16, 8 Text encoder context 256 256 256 Text encoder width 2048 2048 2048 Text encoder depth 24 24 24 Text encoder heads 32 32 32 Latent decoder context 384 Latent decoder width 1664 Latent decoder depth 24 Latent decoder heads 26 Dropout 0. 1 0. 1 Weight decay 4. 0e-2 6. 0e-2 Batch size 4096 4096 2048 1024 512 Iterations IM 6OOK 8OOK IM IM Learning rate 1. 6e-4 1. le-4 1. 2e-4 1. 2e-4 1. 0e-4 Adam 82 0. 91 0. 96 0. 999 0. 999 0. 999 Adam 1. Oe-10 1. Oe-6. Oe-8 1. Oe-8 1. Oe-8 EMA decay 0. 999 0. 9999 0. 9999 0. 9999 0. 9999 Table 3: Hyperparameters for the models 24\n\nD Random samples In Figures[18[gandpOwe show random samples from our production model for some of the prompts from Figure p} Figure 18: Random samples from unCLIP for prompt "Vibrant portrait painting of Salvador Dali with robotic half face" 25\n\nFigure 19: Random samples from unCLIP for prompt *A close up of a handpalm with leaves growing from it' 26\n\nIdd tandy Figure 20: Random samples from unCLIP for prompt "A teddybear on a skateboard in Times Square. 27 |
| 12 | OCR_PAPER_Hong et al. - 2022 - CogVideo Large-scale Pretraining for Text-to-Video Generation via Transformers-annotated_.txt | 66f03e4f-bd9 | OCR_academic_paper | \n\nCog Video: Large-scale Pretraining for Text-to-Video Generation via Transformers Wenyi Hongi Ming Ding' Wendi Zheng Xinghan Liut Jie Tangtt tTsinghua University IBAAI {hongwy18@mails dm18@mails jietang@mail}. tsinghua. edu. cn 8 2 8 3 ] 1 Abstract Large-scale pretrained transformers have created milestones in text (GPT-3) and text-to-image (DALL-E and CogView) generation: Its application to video generation is still facing many challenges: The potential huge computation cost makes the training from scratch unaffordable; The scarcity and weak relevance of text-video datasets hinder the model understanding complex movement semantics. In this work, we present 9B-parameter transformer Cog Video, trained by inheriting a pretrained text-to-image model, Cog View2. We also propose multi-frame-rate hierarchical training strategy to better align text and video clips. As (probably) the first open-source large-scale pretrained text-to-video model, Cog' Video outperforms all publicly available models at a large margin in machine and human evaluations A man is skiing: (running an tnet beach in the late afternoon_ A couple are having dinner: Nietrooisa A lion is Idrinking water: | A girl is dancing: Anime Figure I: Samples generated by Cog' Video. The actual text inputs are in Chinese. Each sample is 4-second clip of 32 frames, and here we sample 9 frames uniformly for display purposes More samples, models and codes will be available athttps : / /github. com THUDM/CogVideo, Equal contribution_ Preprint_ Under review.\n\nL Introduction Autoregressive transformers, eg DALL-E 183 and Cog View [5], have revolutionized text-to-image generation recently. It is natural to investigate the potential of autoregressive transformers on textto-video generation: Previous works followed this basic framework [35, /9, e. g: VideoGPT [361, verifying its superiority over GAN-based methods [4J[26], but are still far from satisfactory: One common challenge is that the generated video frames tend to gradually deviate from the text prompt, making the generated characters hard to perform the desired actions. Vanilla autoregressive models might be good at synthesizing videos with regular (e. g. straightly moving cars) Or random patterns (e. g: speaking by randomly moving lips), but fail on text prompt such as ta lion is drinking water"= The main difference between the two cases is that; in the former case the first frame already provides sufficient information for the subsequent changes, while in the latter the model has to precisely understand the action "drink in order to correctly generate the desired action the lion lifts the glass to its lip, drinks and then puts down the glass. Why do the autoregressive transformers well understand the text-image relations, but struggle to understand the text-action relations in videos? We hypothesize that the datasets and the way to utilize them are the main reasons_ First; it is possible t0 collect billions of high-quality text-image pairs from Internet [18], but the text-video data are more scarce. The largest annotated text-video dataset, VATEX [31/, has only 41, 250 videos. The retrieval-based text-video pairs, e. g: HowtolOOM [16], are weakly relevant and most of them only describe the scene without the temporal information: Second, the duration of videos varies lot. Previous models split the video into many clips of a fixed number of frames for training, which destroys the alignment between the text and its temporal counterparts in the video. If a ~drinking video is split into four individual clips of "holding glass' ~lifting' ~drinking" and "putting down with the same text "drinking" the model will be confused t0 learn the accurate meaning of drinking: Present Work. Here we present a large-scale pretrained text-to-video generative model, Cog Video, which is of 9. 4 billion parameters and trained on 5. 4 million text-video pairs_ We build Cog Video based on a pretrained text-to-image model, Cog View2 [6/, in order to inherit the knowledge learned from the text-image pretraining: To ensure the alignment between text and its temporal counterparts in the video, we propose the multi-frame-rate hierarchical training: The flexibility of the textual condition makes it possible to simply prepend a piece of text describing the frame rate to the original text prompt for modeling different frame rates To keep the text-video alignment; we choose a proper frame rate description to include the complete action in each training sample. The frame rate token also controls the intensity of the changes throughout continuous frames in generation. Specifically, we train a sequential generation model and frame interpolation model The former model generates key frames according to the text, and the latter recursively fill the middle frames by varying the frame rates t0 make the video coherent: As shown in Figure[] Cog' Video can generate high-resolution (480x480) videos Human evaluation demonstrates that Cog Video outperforms all publicly available models at a large margin. Our main contributions can be concluded as follows: We present Cog' Video, which is the largest and the first open-source pretrained transformer for text-to-video generation in the general domain: Cog'Video elegantly and efficiently finetunes a pretrained text-to-image generative model for text-to-image generation, avoiding the expensive full pretraining from scratch_ We propose the multi-frame-rate hierarchical training to better align text-clip pairs, which significantly improves the generation accuracy; in particular for movements of complex semantics. This training strategy endows Cog Video with the capacity of controlling the intensity of changes during the generation. 2 Related Work 2. 1 Video Generation Video generation is a long-standing research topic. Most previous works focus on the next-frame prediction task forecasting the future frames based on the first video frame. Early works, e. g\n\nCDNA [8p and PredRNN [32], leverage deterministic methods to directly predict the next frame via CNNs or RNNs. However; these deterministic models are unable to capture the stochastic temporal patterns and synthesize coherent complex scenes. Generative models, especially Generative Adversarial Networks [10] (GANs), begin to dominate the area as they can perform unconditional o class-conditional video synthesis without the first frames. VGAN [30] is the first one t0 use GAN for video generation: It decomposes video to a static background and moving foreground, and then generates them with 2D and 3D convolutional networks respectively. TGANI19 proposes to separately generate the temporal latent variables and spatial information, and MoCoGAN [26] similarly decomposes the latent space into context and motion subspaces DIGAN [37| applies implicit neural representations for video encoding: Recently; text-to-video generation emerges as a promis Ising direction: The framework of VQVAE [28 and autoregressive transformers 29, [1] quickly becomes the mainstream method 34185, 9]. Ho et al. 11 proposes video diffusion model along with gradient method recently for text-to-video generation. The previous methods are basically trained on specific dataset; e. g: UCF-101 [22], making the trained model domain-specific. Moreover; most of these models are not publicly available 2. 2 Autoregressive Transformer Recent years have witnessed the autoregressive transformer emerging as a powerful generative model. The autoregressive models become the most prevalent framework for text generation [23] With its prominent capacity of fitting; transformer [29] gradually becomes the standard neural structure for text generation One milestone is GPT-3 [1J. In computer vision, van den Oord et al_ [28 first proposes to train VQVAE to compress the image into sequence of tokens from learned dictionary, which can be efficiently handled by autoregressive models. VQ-GAN [7 | learns a more semantic-aware dictionary for unconditional image generation: In the text-to-image generation; pretrained autoregressive transformers such as DALL-E [18] and CogView [5] have shown superiority in open-domain image generation_ Besides the pure GPT-style generation, Cog View2 [6] proposes a new language model CogLM for infilling in the image generation. Recent autoregressive transformers [17] B36, [34} B5[ have also shown their superiority in video generation. Among them, GODIVA [347 and NUWA [351 focus on the open-domain text-to-video generation. However; they simply generate frames or frame blocks one by one in chronological order; and may suffer from poor text-video alignment (Cf: s[ 3 Method In this section, we first introduce multi-frame-rate hierarchical training to better align text and video semantics in $. 4 and then illustrate an efficient method dual-channel attention t0 inherit the knowledge in pretrained text-image models for video generation in $B. 2] To overcome the large memory and time overhead caused by the large model and long sequence, we refer to Swin Attention [14] and extend it to autoregressive video generation in $B3] 3. 1 Multi-frame-rate Hierarchical Training Here we present the multi-frame-rate hierarchical training and generation. We follow the framework of VQVAE [28] and first tokenize each frame into image tokens_ Each training sample consists of 5 frames of tokens, but our training method differs in the construction of training sequences and generation process. Training: The key design is to add a frame-rate token to the text and sample frames at this frame rate to compose a fixed-length training sequence. The motivations are two folds: Directly separating the long video into clips at fixed frame rate often leads to semantic mismatching: We still use the full text but the truncated clip might oly contain incomplete action (2) The adjacent frames are usually very similar: A giant change over the previous frame will probably incur a large loss This will lead the models less inclined to explore the long-range correlation because simply copying the previous frame acts like a shortcut:\n\nInput Text: Input Frames Image Tokenizer A lion is drinking water: RWFETIBjK 2 Discretize 20*20-400 image tokens per frame 5 frames Text tokenization Frame Rate Flatten Text Frame-1 Frame-2 Frame-3 Frame-Frame-5 Transformer (Stage 1: Sequential Generation) Interpolate frames Sequence IB] FrameSequence 2 Frame-3 Frame-4 Frame Rate Text Frame-2 Frame-5 Transformer (Stage 2: Recursive Interpolation) Figure 2: Multi-frame-rate hierarchical generation framework in Cog'Video. Input sequence includes frame rate, text, frame tokens. [B] (Begin-of-image) is separator token, inherited from Cog View2. In stage 1, Ts frames are generated sequentially on condition of frame rate and text Then in stage 2, generated frames are re-input as bidirectional attention regions to recursively interpolate frames_ Frame rate can be adjusted during both stages. Bidirectional attention regions are highlighted in blue and unidirectional regions are highlighted in green Therefore; in each training sample, we want the text and the frames to match as possible. We predefined series of frame rates, and select the lowest frame rate for each text-video pair; as long as we can sample at least 5 frames at this frame rate in the video Although the above method increases the alignment of text and video, the generation at a low frame rate could be incoherent. We train another frame interpolation model t0 insert transition frames t0 the generated samples of the sequential generation model Thanks to the generality of CogLM [6], the two models can share the same structure and training process only with different attention masks_ Generation. The multi-frame-rate hierarchical generation is recursive process, illustrated in Figurep] Specifically, the generation pipeline consists of a sequential generation stage and a recursive interpolation stage: Sequentially generate Ts key frames based on low frame rate and text: The input sequence is [{Frame Rate}{Text} [B] {Framel} {Frame Ts}]. In practice, we always set Ts 5 and the minimum sampling frame rate to 1 fps. (2) Recursively interpolate frames based on the text, frame rate and known frames. In each round of interpolation, we split generated frames into multiple % ]-frame blocks overlapping at the beginning and the end, and interpolate a frame between the successive frames in each block: The input sequence is [{Frame Rate} {Text} [B] {Frame1} {Frame Ts}] where Frame 2i(i = 1, 2, 2 J) are to be autoregressively generated. By recursively halfing {Frame Rate}, we can conduct finer and finer interpolation to generate videos of many frames. The effect of CogLM. Tasks such as frame interpolation rely heavily on bidirectional information_ However; most previous works use GPT [34, B6] [5], which is unidirectional. To be aware of the bidirectional context; we adopt Cross-Modal General Language Model (CogLM) proposed in [6] which unites bidirectional context-aware mask prediction and autoregressive generation by dividing tokens into unidirectional and bidirectional attention regions While bidirectional regions can attend to all bidirectional regions, unidirectional regions can attend to all bidirectional regions and previous unidirectional regions As shown inp] (1) all frames in stage and the Znd, 4th frames in stage\n\n2 are in the unidirectional region; (2) {Frame Rate}, {Text} and all other frames belong to the bidirectional region. In this way, bidirectional attention context is fully exploited in text and given frames without interfering with auto-regressive frame prediction. 3. 2 Dual-channel Attention Large-scale pretraining usually demands a large dataset. For the Xout open-domain text-to-video generation;, ideally we need the dataset Addition to cover sufficient text-video pairs to infer both spatial and temporal correlation between video and text: However; to collect Layer Norm high-quality text-video pairs is often difficult; expensive and timeDual-channel Attention consuming: Addition 10 A natural idea is to make use of the image data to facilitate the Attention-base Attention-plus learning of spatial semantics. Video Diffusion Model 11 and Spatial Channel) (Temporal Channel) NUWA [35| try to add text-image pairs into text-video training; which achieves better results on multiple metrics_ However; as Layer Norm for training a video-only generation model, adding image data Xin will significantly increase training costs, especially in large-scale pretraining scenarios Figure 3: The dual-channel attention block: We initialize In this paper; we propose to leverage pretrained image generation the Attention-plus the same as models instead of image data Pretrained text-to-image models, Attention-base so that the model e. g: Cog'View2 [6], already have good command of the textbehaves exactly the same as image relations. The coverage of the dataset to train these models CogView2 when it is initialized. is also larger than that of videos_ The proposed technique is dual-channel attention, where we only add a new spatial-temporal attention channel to the pretrained Cog View2 [6] at each transformer layer: All the parameters in the Cog View2 are frozen in the training, and only the parameters in the newly added attention layer (See the attentionplus in FigureB] are trainable_ We denote the original attention block in CogView2 as attention-base Here we also emphasize that directly finetuning CogView2 for text-to-video generation cannot well inherit the knowledge, because the temporal attention follows a different attention pattern and quickly ruins the pretrained weights during the initial phase of training with large gradients. Specifically, the dual-channel attention block with Sandwich-LN [5] can be computed as x =0 attention-base(LayerNorm (Tin _ + (1 a). attention-plus(LayerNorm(Tin) ), (2) Tout Tin + LayerNorm(x). The mixture factor & is a vector EUR (0, 1)&, where d is the hidden size of the input feature TinTo restrict the range of a within (0, 1), we reparameterize it as & sigmoid(a) EUR (0, 1)8 where a EUR Rd is a learnable parameter: The attention-plus block has the same shape of parameters as the normal multi-head attention block, attention-base, but differs in the procedure of computation as follows. In our training, we tried two kinds of attention, 3D local attention and 3D Swin [14] attention for attention-plus block: Here we depict the 3D local attention, and the latter is a natural replacement introduced in sectionB3 In 3D local attention, the receptive field (RF) for the token at (t, EUR, y) (where (t, EUR, y) corresponds to the coordination along time, height and width), is a 3D block with extent lt, lz, ly EUR Nt: RF(t. f, y) = {(k, i, j) Ix i] < 1, ly jl < ly;lt _ kl < lt; (k, i, j) $ Mask(t, z, y)}; (3) where Maskat. f '9) represents an attention mask for token (t,., y). In the sequential generation model (Stage 1), the Mask ensures the auto-regressive order; In the interpolation model (Stage 2), the Mask is designed as in as CogLM [6] to make the known frames visible to all the frames_ It is worth noting that two channels are fused and share the same FFN in each layer; because FFN is a module ofheavy parameters containing much vision knowledge. Due to the similarity between images and videos, bringing its knowledge to the temporal channel will facilitate video modeling: Finally, sharing FFN can reduce parameters, thus speeding up training and reducing memory overhead: 5\n\n3. 3 Shifted Window Attention in Auto-regressive Generation To further alleviate the large time and memory overhead in the temporal channel during training and inference, we refer t0 Swin Attention [14]. The original Swin attention is only applied to non-autoregressive scenarios, We extend it t0 the autoregressive and temporal scenario by applying an auto-regressive attention mask in the shifted windows An interesting finding is that, the Swin attention provide a chance for parallel generation in faraway regions of different frames, which further accelerates the auto-regressive genera tion: The dependence of the generation of a specific token relies on Auto-regressive mask: token can only attend to previous frames Or tOt-i+1 t-i+2 kens before itself in the current frame_ Shifted window. Only tokens within Figure 4: In 3D autoregressive swin attention (winthe distance of window size in both dow size 2 X 2 as an example), the token in the red width and height dimensions can be box can only attend to (either directly or indirectly) directly attended to. the yellow O green tokens The gray tokens in the i-th frame and the token in the red box can be As shown in Figurel] we can start generating generated in parallel. parts of the tokens in the following frames before finishing the generation of all the previous frames they can work in parallel. Suppose X, Y is the height and width of each frame, and Ar, Ay are the height and width of shifted window. For two tokens at (t1, 11, 91) and (t2, *2, Y2), +1 t2, the latter cannot attend to the former either directly O indirectly if (81 T2)Y + (y1 Y2) > (t2 t1 +1)(AcY + Ay); which means that the i-th token in the t-th frame can be generated with the (i A, Y _ Ay)-th token in the (t + 1)-th frame in parallel. In this way; we can generate [z44w, tokens in parallel at most_ thus greatly enhance parallelism and accelerate inference compared to auto-regressive with standard attention which can only generate one token at a time_ Training Based on the methods above, the training details of Cog'Video are listed as follows: Model: The backbone of Cog' Video in both stages is Transformer with dual-channel attention The Transformer has 48 layers, with hidden size of 3, 072 in each attention channel, 48 attention heads and 9. 4 billion parameters in total. Among them, 6 billion parameters are fixed to CogView? s parameters, which include Position-wise Feed-Forward Networks (FFN), the spatial channel of dualchannel attention, first frame '$ positional embeddings and all image and text vocabulary embeddings. The specific implementation of Transformer structure is almost identical to CogView [5| such as using Sandwich LayerNorm and PB-Relax to stabilize training. Shifted CogLM attention window is adopted in recursive interpolation model with window size 10 x 10. Dataset: We pretrain our model on dataset of 5. 4 million captioned videos with a spatial resolution of 160 x 160 (can be upsampled to 480 X 480 by Cog View2). For the sequential generation model (Stage 1), we adjust the frame rate in each sample to accommodate the whole video, while the minimum frame rate is set t0 [ fps. For the recursive interpolation model (Stage 2), we split videos into clips of different lengths to accommodate prediction on multiple frame rates including 2, 4, 8 fps. Pretraining: The sequence lengths in both stages are 2, 065, consisting of 64 text tokens, 5 (frames) 400 (per frame) image tokens, and seperator token. Both text and images are tokenized with icetkH]The parameters are updated by Adam with max learning rate = 2 x 10 81 0. 9, 82 0. 95, weight decay = 1 10-2. See Appendix for pretraining details. ~https: '/github _ com/ THUDM_ icetk\n\nTable I: (Left) Video generation performance on UCF-1O1. Class labels are used as the text inputs means that the model is only trained on the training split of UCF-1O1. (Right) Video generation performance on Kinetics-600_ The metrics are based on the 16-frame generated videos priming on first 5 frames, following settings of Rakhimov et al 17|**means that the ground truth used in FVD testing is the reconstruction result of the tokenizer: Method IS FVD 24. 69 27. 38 28. 87 1209 32. 36 838 29. 71 655 32. 70 577 79. 28 332 VideoGPTI36 DVD-GANI4 TGANv2[20p MoCoGAN-HDI24 DIGAN[37 DIGAN[37 TATS-basel9l Cog"Video (Ours Cog' Video (Ours_ Method Latent Video Tranformerl 17 Video Transformer[33 DVD-GAN-FPI4_ TriVD-GAN-FP[15] Cog Video (Ours, Cog Video (Ours FVD(L) 224. 73 170 69. 15 25. 74 109. 23 59. 55 50. 46 626 545 5 Experiments 5. 1 Machine Evaluation Machine evaluation is conducted on two popular benchmarks for video generation; ie, UCFIOI [22] and Kinetics-600 [3]. Following Rakhimov et al. [17|, Yu et al. [37/, we use Frechet Video Distance (FVD) [27E and Inception score (IS [21] as metrics in the evaluation_ FVD is calculated based on I3D modell2| trained on Kinetics-400, and IS is based on C3D model [25_ which was first trained on the Sports-IM dataset 12 and then finetuned on the UCFIOI dataset_ Our evaluation code is the same as the official TGAN-v2 implementationz] UCF-IOI is a human action dataset consisting of 13, 320 videos annotated with 101 action classes. Due to the gaps of image style and frame rate between Cog' Video 's training set and UCF-101, we use class labels as the input text and finetune Cog'Video on the whole dataset for 10, 000 iterations with a batch size of 192. During inference, we generate samples of various classes according to the class distribution: FVD and IS are evaluated over 2, 048 and 10, 000 samples respectively, following Yu et al. [37]. Results are shown in Table[](Left). Kinetics-600 contains 600 classes of human action videos, with roughly 350, 000 train and 50, C 000 test videos in total. We use the action category as input text, and finetune Cog Video on the training set for 12, 000 iterations with batch size of 640. Following the setup of Weissenborn et al. [33], Rakhimov et al. [17], we center-crop and downsample each frame to 64X64 to measure the FVD of the model_ Results are shown in Table[@](Right). 5. 2 Human Evaluation To further evaluate Cog'Video, we invite 90 anonymous evaluators to rate for Cog' Video and other open source baselines including GAN-based model TGANv2 [20] and GPT-based model VideoGPT 36]. 30 classes in UCFIOL are randomly picked as text conditions, and several aspects are rated (See Appendix for details). For VideoGPT; we use the official unconditional pretrained model to generate samples. For TGANv2, we use the official source code to train an unconditional generation model under the same setting as that in Saito et al. [20[ To assign unconditionally generated samples into corresponding categories, we choose TSM [13 as the action recognition model for postclassification. We only keep the samples whose likelihood to a certain class is at least 80%. Results in Figure[lshow that Cog Video significantly outperforms baselines on multiple important aspects including frame texture, motion realism and semantic relevance, and achieves the top score by the overall quality It can be seen that 49. 53% evaluators choose Cog Video as the best method, and only 15. 42% and 5. 6% favor VideoGPT and TGANv2, respectively: https: Igithub com_ [pfnet research/tgan2 https : '/github _ com/ wilsonlyan/VideoGPT\n\nTGAN videoGPT Cog video IStage Cogideo Ground Truth TGe VidecGpt Cuyvicec St10" Convicen Ground Truch Lonvide? Staje videoGPt 015, 42 5. 61 TGAN 4957 CogVideo Ovrnl Scnre Frame exturz Kotion Kejiism bemantic evonce Human preference The percentage of being chosen as the best:. Overall scores (1-10) for each method: Scores (1-5) on three important aspects_ Figure 5: Human evaluation results. ~Cog Video IStage" refers to the method in ablation study, which only generates videos sequentially with the CogVideo's Stage 1 to the desired number of frames: Table 2: Ablation study on 5, 000-sample subset of Kinetics-600 s testset: FVD is evaluated on generated H-frame samples priming on 5 frames and the recovered ground-truth by the image tokenizer: The setting column indicates the difference between each method and CogVideo. Models of each setting are trained on Kinetics-600 trainset for 11, 000 iterations with a batch size of 160. Method Setting None FVD 108. 27 Cog Video 1-stage Generation( Noverlap 1-stage Generation( Noverlap 2) Initialized with Cog View2 Randomly Initialized hierarchical hierarchical 137. 13 120. 82 124. 92 166. 13 Pretrain Pretrain Cog' View2 5. 3 Ablation Study To verify the effectiveness of hierarchical multi-frame-rate generation and incorporating Cog View2, we conduct ablation studies on Kinetics-600 and UCF-1O1 datasets. We will first briefly introduce the compared methods and analyze the quantitative results in $[5. 3 Tand qualitative results in $653. 2 Hierarchical multi-frame-rate generation. In comparison with CogVideo; we finetune I-stage video generation model on Kinetics-600 from the sequential generation model in CogVideo, which generates long videos by sliding windows In each window, We generate the rest frames based on Noverlap previous known frames. Larger Noverlap means more previous frames can be utilized during the inference, but will increase time overhead: Dual-channel attention with Cog View2's weights. To highlight the effectiveness of our finetuning strategy, we additionally finetune (1) randomly initialized model, (2) model incorporating Cog' View2'$ weights but leaving the temporal channel unfixed (equivalent t0 CogVideo without pretraining O videos) on Kinetics-600 for comparison 5. 3. 1 Quantitative Evaluation Cogvideolfinetune) CogView Initialized Randomly Initialized All aforementioned models have been trained for 11, 000 5. 75 iterations with 5, 50 batch size of 160. Quantitative results are 5. 25 shown in Tablep We can see that the hierarchical method 8 is clearly superior to the 1-stage generation with differ1 4. 75 4. 50 ent Noverlap; and the model initialized with Cog'View2 s 4. 25 weights has lower FVD than the randomly initialized one: 3. 75 Figure plots the training loss curve of (1) finetuning 3. 50 Cog' Video; (2) training model from random initialization; 3, 25 (3) training model initialized with Cog View2 and partially thousand iteration fixed_ Figure 6: Training loss in ablation study: We can see that Cog View2 endows the model with 8\n\nRandomly Initialized Input Text: Lunge Given frames: Finetuned CogVideo, hierarchical generation Finetuned CogVideo, 1-Stage Noverlap (b) Initialized with CogView2 Finetuned CogVideo, 1-Stage Noverlap Figure 7: Video samples in ablation study, which are generated priming on the class label and first 5 frames in Kinetics-600. All samples are downsampled by extracting one in every three frames for display purposes. (a) Use finetuned Cog Video to hierarchically generate samples. (b) Train a model on Kinetics-600 which is initialized as and partially fixed to Cog' View2, and hierarchically generate samples (c) Train a model on Kinetics-600 which is randomly initialized, and hierarchically generate samples. (d)(e) Use finetuned Cog Video to generate frames in stage with different Noverlap_ good initialization point from which the loss can decrease faster: Moreover; fixing part of the parameters reduces the time and memory cost: 5. 3. 2 Qualitative Evaluation Qualitative comparison is shown in Figure[ While the model trained from random initialization tends to produce irrational deformation, the model incorporating CogView2 is able to generate realistic objects, and the hierarchical generation performs better on content consistency and motion realism We also conduct human evaluation between I-stage and hierarchical video generation model under the same setting as in $6. 2 As shown in Figure/5] the hierarchical model, i. e. Cog'Video, outperforms the 1-stage model on semantic relevance, motion realism as well as texture quality: This is probably because the 1-stage model cannot estimate a proper intensity of change from the previous frames in the window, as shown in Figure[Kd)(e) 6 Conclusion We present Cog' Video, to the best of our knowledge, the largest and the first open-source pretrained transformer for text-to-video generation in general domain. CogVideo is also the first attempt to efficiently leverage the pretrained text-to-image generative model to the text-to-video generation model without hurting its image generation capacity: With the proposed multi-frame-rate hierarchical training framework; CogVideo is endowed with better understanding of text-video relations and abilities to control the intensity of changes during generation. We extend swin attention to CogLM; which achieves acceleration in both training and inference. There are still some limitations in Cog'Video, e. g restriction on the length of the input sequence still exists due to the large scale of the model and limitation of GPU memory; and we leave them for future work: Broader Impact: This paper aims to advance the open-domain text-to-video generation, which will ease the effort of short video and digital art creation The efficient training method transfers knowledge from text-to-image models to text-to-video models, which helps avoid training from scratch, and thus reduces energy consumption and carbon emission_ A negative impact is the risk of misinformation_ To alleviate it; we can train an additional classifier t0 discriminate the fakes. We believe the benefits outweigh the downsides.\n\nAcknowledgments and Disclosure of Funding We would like t0 thank Zhao Xue, Shuai Zhao, Sha Yuan for their help in data collection, Weidong Guo, Fengyu Rao, Zhaoyang Zeng; Mingkang Tang for their useful discussion, Hanxiao Qu for maintaining the machines and the computational resources supported by BAAL References T B. Brown B Mann_ N. Ryder; M Subbiah; J. Kaplan, P Dhariwal, A. Neelakantan, P Shyam; G. Sastry, A_ Askell, et al. Language models are few-shot learners. arXiv preprint arXiv: 2005. 14165, 2020. [2] J. Carreira and A Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset_ In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. [3] J. Carreira; E. Noland, A_ Banki-Horvath, C. Hillier; and A. Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808. 01340, 2018. [4] A. Clark, J. Donahue, and K. Simonyan Adversarial video generation on complex datasets. arXiv preprint arXiv: 1907. 06571, 2019. [5] M. Ding, Z Yang, W: Hong; W. Zheng; C. Zhou; D. Yin, J. Lin, X Zou; Z. Shao, H Yang; et al_ Cogview: Mastering text-to-image generation via transformers_ Advances in Neural Information Processing Systems, 34, 2021. [6] M Ding, W: Zheng; W: Hong; and J. Tang: Cogview2: Faster and better text-to-image generation via hierarchical transformers_ arXiv preprint arXiv:2204. 14217, 2022. [7] P Esser; R Rombach; and B Ommer Taming transformers for high-resolution image synthesis arXiv preprint arXiv:2012. 09841, 2020. [8] C Finn, I Goodfellow, and S. Levine. Unsupervised learning for physical interaction through video prediction. Advances in neural information processing systems; 29, 2016. [9] S. Ge_ T Hayes, H: Yang; X. Yin, G. Pang; D. Jacobs, J. -B_ Huang; and D_ Parikh_ Long video generation with time-agnostic vqgan and time-sensitive transformer: arXiv preprint arXiv:2204. 03638, 2022_ [10] I. J. Goodfellow; J. Pouget-Abadie, M. Mirza, B. Xu; D. Warde-Farley, S. Ozair, A. Courville_ and Y Bengio. Generative adversarial networks arXiv preprint arXiv:1406. 2661, 2014. J. Ho, T Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet: Video diffusion models_ arXiv preprint arXiv:2204. 03458, 2022. [12] A. Karpathy, G. Toderici, S. Shetty, T: Leung, R Sukthankar; and L Fei-Fei: Large-scale video classification with convolutional neural networks_ In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725-1732, 2014. [13] J. Lin, C. Gan, and S. Han. Tsm: Temporal shift module for efficient video understanding_ In Proceedings of the IEEEICVF International Conference on Computer Vision, pages 7083-7093, 2019. [14] Z Liu; Y Lin, Y Cao, H. Hu, Y. Wei, Z Zhang, S. Lin, and B_ Guo. Swin transformer: Hierarchical vision transformer using shifted windows_ In Proceedings of the IEEEICVF International Conference on Computer Vision, pages 10012-10022, 2021. [15] P Luc, A Clark, S_ Dieleman D. d_ L: Casas, Y Doron, A Cassirer; and K Simonyan_ Transformation-based adversarial video prediction on large-scale data_ arXiv preprint arXiv:2003. 04035, 2020. [16] A. Miech; D Zhukov; J-B. Alayrac, M. Tapaswi, I. Laptev, and J. Sivic. HowtolOOm: Learning text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 10\n\n[17] R Rakhimov, D. Volkhonskiy, A. Artemov; D Zorin, and E. Burnaev. Latent video transformer: arXiv preprint arXiv:2006. 10704, 2020. [18] A Ramesh; M. Pavlov, G. Goh; S. Gray, C. Voss, A. Radford, M. Chen, and I. Sutskever: Zero-shot text-to-image generation: arXiv preprint arXiv:2102. 12092, 2021_ [19] M. Saito, E. Matsumoto, and S_ Saito_ Temporal generative adversarial nets with singular value clipping: In Proceedings of the IEEE international conference on computer vision, pages 2830-2839, 2017. [20] M. Saito, S. Saito, M. Koyama, and S. Kobayashi. Train sparsely, generate densely: Memoryefficient unsupervised training f high-resolution temporal gan_ International Journal of Computer Vision, 128(10). 2586-2606, 2020. [21] T Salimans I. Goodfellow, W Zaremba, V. Cheung, A. Radford, and X. Chen_ Improved techniques for training gans. In Proceedings of the 3Oth International Conference on Neural Information Processing Systems, pages 2234-2242, 2016. [22] K Soomro, A. R Zamir; and M. Shah: UcflOl: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv: 1212. 0402, 2012. [23] I Sutskever; J. Martens, and G. Hinton. Generating text with recurrent neural networks In ICML' 11, page 1017-10024, 2011. [24] Y Tian, J. Ren, M. Chai, K. Olszewski, X Peng; D. N. Metaxas, and S. Tulyakov good image generator is what you need forhigh-resolution video synthesis. arXiv preprint arXiv:2104. 15069, 2021. [25] D: Tran, L. Bourdev, R Fergus, L. Torresani, and M. Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. [26] S. Tulyakov; M-Y Liu, X Yang; and J. Kautz Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. [27] T Unterthiner; S_ van Steenkiste, K Kurach_ R. Marinier; M Michalski, and $. Gelly: Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv: 1812. 01717, 2018. [28] A_ van den Oord, 0. Vinyals, and K Kavukcuoglu_ Neural discrete representation learning: In Proceedings of the 31st International Conference on Neural Information Processing Systems; pages 6309-6318, 2017. [29] As Vaswani N. Shazeer; N. Parmar; J Uszkoreit; L. Jones, A_ N. Gomez, L Kaiser; and [. Polosukhin. Attention is all you need_ arXiv preprint arXiv: 1706. 03762, 2017. [30] C. Vondrick, H. Pirsiavash; and A_ Torralba. Generating videos with scene dynamics Advances in neural information processing syStems, 29, 2016. [31] X Wang, J. Wu; J. Chen, L Li; Y-F Wang; and W. Y Wang: Vatex: large-scale, highquality multilingual dataset for video-and-language research: In Proceedings of the IEEEICVF International Conference on Computer Vision, pages 4581-4591, 2019. [32] Y. Wang; M Long, J. Wang, Z. Gao, and P S. Yu_ Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms Advances in neural information processing systems; 30, 2017. [33] D. Weissenborn; 0. Tackstrom; and J. Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv: 1906. 02634, 2019. [34] C. Wu, L. Huang, Q. Zhang; B. Li, L. Ji, F Yang; G. Sapiro, and N. Duan: Godiva: Generating open-domain videos from natural descriptions arXiv preprint arXiv:2104. 14806, 2021.\n\n[35] C. Wu; J. Liang; L. Ji, F Yang; Y Fang, D_ Jiang, and N. Duan_ Nl uwa: Visual synthesis pre-training for neural visual world creation. arXiv preprint arXiv:2111. 12417, 2021. [36] W: Yan, Y Zhang; P Abbeel, and A_ Srinivas. Videogpt: Video generation using Vq-vae and transformers arXiv preprint arXiv:2104. 10157, 2021. [37] S. Yu; J. Tack, S. Mo, H. Kim; J. Kim; J. -W. Ha, and J. Shin. Generating videos with dynamicsaware implicit generative adversarial networks arXiv preprint arXiv:2202. 10571, 2022. Attention Analysis To explore the attention mechanism of dual-channel attention; we visualize (1) the attention distribution in the temporal channel and (2) the mixture factor & controlling the ratio between the spatial and temporal channel in equation[ Figure[8]visualizes the distribution among frames and texts in sequential generation (Stage 1) with heat maps, where only 24 of 48 attention heads in 6 layers are shown for display purposes The attention patterns can be broadly classified into the following categories: Most of the attention is on the text: E. g: the attention heads in violet Most of the attention is on a certain frame. Eg: the attention heads in pink focus mainly on the previous frame; the attention heads in blue focus mainly on the first frame besides the text; the attention heads in yellow focus mostly on the frame itself: Attention is spread over several frames Eg: the attention heads in green Some attention heads exhibit a single pattern; while others may exhibit a mixture of them. Attention heads in the same layer tend to show similar patterns: In lower layers (e. g: layer 4, 12) the heads tend to allocate attention according to position, while in higher layers more attention is allocated to text (e. g. layer 44) O spread over multiple frames. One possible explanation is that there are more high-level features in higher layers such as video semantics, by which more frames and texts can interact with each other to make high-level feature analysis. It is worth noting that many heads in temporal channel do not allocate much attention to the frame itself, especially in higher layers, while attending to itself is important for inference This shows that the Cog Video performs a certain degree of decoupling in the analysis of temporal and spatial features While the spatial channel is in charge of feature analysis within the frame, the temporal channel can allocate more resources t0 explore relationships among different frames We further illustrate this perspective with FigureC which shows that features calculated by Cog View2 in the spatial channel are heavily relied on_ B Training Details Cog' Video consists of two models corresponding to two stages, i. e. sequential generation and recursive interpolation_ Both models have 7. 7 billion parameters while 6 billion of them are fixed to Cog View2, thus Cog Video has 9. 4 billion different parameters in total_ CogVideo is trained on a dataset of 5. 4 million captioned videos with spatial resolution of 160x 160 (can be upsampled to 480x480 by Cog View2) Each model is pretrained separately. The model in stage 1 is first pretrained for 76, 000 iterations on video clips with minimum frame rate of 0. 25 fps, then trained for 15, 000 iterations with minimum frame rate of 1 fps The model in stage 2 is pretrained for 78, 500 iterations with the frame rate of 2_ 4, and & fps. Both models are trained in FPI6 with a batch size of416, and optimized by Adam with max learning rate = 2 x 10-4, 81 0. 9_ 82 0. 95, weight decay = 1 x 10-2_ 12\n\nLayer 4 Layer 12 Layer 20 Layer 28 Layer 36 Layer 44 Figure &: The attention distribution among frames and texts in sequential generation (Stage 1) Only 24of 48 attention heads in 6 layers are selected for display purposes. Each attention head is visualized with a heat map of size 5X6, where lighter color represents larger value The 5x5 block on the left indicates the sum of attention scores (after softmax between each pair of frames, and the rightmost column indicates the sum of the attention score of each frame t0 text That is to say, the grid in row i column j (j < 5) represents EreF yeF; attn_, y, and the grid in row column 6 represents ZreF, yeT attnc y, where Fi, T denotes the set of tokens in the i-th frame and text respectively; and attnz; y denotes the attention score of token x t0 y: Details about Human Evaluation In this section, we introduce more details about the human evaluation for measuring generation quality: The conduction of our human evaluation generally follows previous works including Ramesh et al. [18], Ding et al. [5] We randomly extract 30 classes from UCFIOL for video generation, using corresponding video samples in the dataset as ground truth items in the evaluation. Based on captions of selected classes, we generate video samples from models including TGANv2, VideoGPT; and our model_ Cog' Video. To further illustrate the effectiveness of hierarchical multi-frame-rate generation, we also include I-stage version of CogVideo model fine-tuned on Kinetics-600 which is described in $63] For TGANv2, we use the official source code to train an unconditional generation model under the same setting as that in Saito et al. [20]. For VideoGPT, we use the official unconditional pretrained model to generate samples To assign unconditionally generated samples into corresponding categories; we choose TSM[13/ as the action recognition model for a post-classification: We only keep the samples whose likelihood to a certain class is at least 80%. A randomly selected subset of samples is displayed in Figure [0] For each sample of the video mentioned above, we ask evaluators to give scores between and 5 ( 5 indicates the best while indicates the worst) from three aspects including frame texture, motion realism, and semantic relevance Then the evaluators are required to give a general score of quality for each sample between and 10, where higher score indicates better quality. After video samples 13\n\n8 0. 4 12 14 32 Layer Figure 9: The scale factor & controlling the ratio between the spatial and temporal channel in equation[Jin dual-channel attention. Only a in half of the layers are shown for display reasons As & is a vector of dimension 3072, We show the mean and variance among all of its dimensions in this figure_ Skiing, {83 Biking; 95A(j4 CogVideo CogVideo Stage) 0, 'Y VideoGPT TGANv2 Groundtruth Figure 1O: A subset of human evaluation samples. The captions are randomly selected from UCF-101. The original samples are clips of 16 frames, which are downsampled to 4 frames uniformly for display purposes: from each caption are all evaluated, the evaluators are asked to select the best one from them We show snapshots of the evaluation website in Figure[ Throughout the process of human evaluation, we invited nearly 100 anonymous evaluators, while 90 of them completed the whole evaluation and were counted in the final results. None of the questions in the evaluation have any time limit We offer each evaluator 75 RMB as a reward for the evaluation_ Results of the human evaluation, including the average score and standard deviation for each group, have already been introduced in Figure[lin the main body: As ground truth samples take an absolute predominance in the best selection question; we have removed the part of ground truth samples in the selection pie plot for clearer model comparison. 14\n\nCoayilee TAo CogVideo Task 0 Arudhaurick UFtnilsta EZEpEmi AAAIHm, T^Sin@0 6 @rrpRuaezix Be-%eil) jitmpistmtejiare s41 ~Wnt4a, TBM, Tratir5-Wnirbtatiq) Mnmzrab (4A246 Q@R) #0ade9ziybeias(g (1-@ipekraitStEz? 74* 5-6FDZ+0 #6E#haSI#8M2) Kaazrjad (WetTKAMiM) #Mariemt swoaatwo (1-Jfijizmu4, 1868X, 05542*T+* 10+07123180*4R*RSMw Ditmtlar erimRMiRZ 10-GARS) Figure 1: Snapshots of the evaluation website. 15 |
| 13 | OCR_PAPER_Kandpal, Nieto, Jin - 2022 - Music Enhancement via Image Translation and Vocoding-annotated_.txt | 110b05be-f8d | OCR_academic_paper | \n\nMUSIC ENHANCEMENT VIA IMAGE TRANSLATION AND VOCODING Nikhil Kandpal* Oriol Nieto, Zeyu Jin University of North Carolina at Chapel Hill Computer Science Department Chapel Hill, NC, USA Adobe Research San Francisco, CA, USA ABSTRACT to develop a solution that works for polyphonic signal enhancement and reflects the unique qualities of music perception. Our approach performs enhancement on the recording' s melspectrogram representation. This is achieved by treating the melspectrogram as an image and training an image-to-image translation model similar t0 Pix2Pix [3] to transform low-quality melspectrogram into that of a high-quality signal. We hypothesize that it is easier to enhance polyphonic signals in the mel-spectrogram domain as polyphonic sources are additive and have a very small temporal span compared to waveforms Finally; to map generated high-quality mel-spectrograms to perceptually realistic waveforms we train vocoding model based on DiffWave [4|. Training this model on only the high quality samples of music performance makes it robust to the artifacts that reside in the synthetic mel-spectrogram_ We evaluate Our approach by performing listening test with 211 participants, and we show that this approach achieves a much better perceptual enhancement than several state-of-the-art techniques_ We also compare the subjective listening test scores with widely used audio quality metrics and suggest that, similar to speech enhancement, these metrics correlate poorly with human perception [1[5 : With this work; we hope to motivate both future research in music enhancement as well as music quality perceptual metrics akin to those in the speech literature [6], [1. To promote further research, audio samples generated in our experiments and source code are provided at Our project website In this paper; we refer to Pix2Pix models operating on melspectrograms as MelzMel models and vocoding applied to the music domain as musecoding. We summarize our contributions as follows: A music enhancement model leveraging recent work on conditional image synthesis and vocoding: generative process for simulating realistic low-quality music recordings from professional-quality recordings An analysis of the reliability of common audio enhancement evaluation metrics in the music domain_ 8 Consumergrade music recordings such as those captured by mo bile devices typically contain distortions in the form of background noise, reverb_ and microphone-induced EQ. This paper presents 2 deep learning approach to enhance low-quality music recordings by combining (i) an image-to-image translation model for manipulating audio in its mel-spectrogram representation and (ii) a music vocod8 ing model for mapping synthetically generated mel-spectrograms to perceptually realistic waveforms We find that this approach to music enhancement outperforms baselines which use classical methods for mel-spectrogram inversion and an end-to-end approach directly mapping noisy waveforms t0 clean waveforms. Additionally, 8 in evaluating the proposed method with listening test; we analyze the reliability of common audio enhancement evaluation metrics when used in the music domain: Index Terms _ Music Enhancement; Image-to-Image Translation, Diffusion Probabilistic Models, Vocoding 7 1. INTRODUCTION With the rise of Internet influencers and music hobbyists, large portion of music content is created with cheap and accessible recording devices in non-treated environments _ While being audible, these recordings often have degraded quality stemming from background noise, unpleasant reverb, and resonance caused by the microphone and the environment_ This prompts US t0 investigate quality enhance1 ment for music signals, transforming low-quality amateur recordings into professional ones. The main difficulty of such an endeavor is that $o many aspects of the low-quality recording setup are unknown Parameters of the recording device, such as frequency response characteristics, vary drastically across different hardware. Additionally, acoustic properties such as the size, shape, and reflectivity of the recording environment vary between different recording setups_ Finally, background noise is hard to capture and generalize, especially non-stationary 2. RELATED WORK noise. A solution that faithfully transforms low-quality recording into what it would sound like recorded professionally must implicTo our knowledge there is little prior work studying music quality itly or explicitly infer all of these aspects from the signal alone In enhancement. The work most similar to our contributions focuses on speech enhancement; end-to-end methods such as HiFi-GAN [1 and speech enhancement; conditional speech synthesis, O music source Demucs |2] achieve this by extracting the speech source from a mixseparation_ ture of sources However; music signals are often polyphonic, i. eEarly approaches to speech enhancement have used classithere can be an arbitrary number of sources to be extracted at once_ cal signal processing techniques such as Wiener filtering [8] and Moreover; the perception of music quality typically differs from that non-negative matrix factorization [9|. More recently, deep learningof speech: For example, human listeners may find reverb pleasant based methods have achieved state-of-the-art on speech enhancein music, while it is usually undesired in speech. Therefore, we aim ment_ These methods either manipulate the audio in its magnitude Work done during an internship with Adobe Research Ihttps: nkandpa2. github 10 music enhancement\n\nHigh-Quality Spectrogram MelzMel GAN Musecoder DDPM Low-Quality Spectrogram Conv Encoder Conv Decoder Denoising Steps High-Quality Waveform Noise Diffusion Steps Fig: 1. Model architecture of our Mel2Mel + Diffwave model. First, low-quality mel-spectrogram is enhanced by conditional GAN. The resulting synthetic mel-spectrogram is then "musecoded" into a waveform by a Denoising Diffusion Probabilistic Model (DDPM) spectrogram representation (followed by spectrogram inversion method to recreate the corresponding waveform) |10} H[2| or map directly from the low-quality waveform to a cleaned waveform 113] 2 : Methods that operate on the time~frequency domain generally produce audible artifacts due to the use of phase reconstruction algorithms like the Griffin-Lim algorithm [14] A recent work addresses this with neural-network based vocoders [15], yet its quality is not 0 par with an end-to-end approach [16p Alternatively; methods that work on the time domain typically require more training steps |1| Conditional speech synthesis techniques produce speech wave _ forms from conditioning information such as magnitude spectrograms, problem commonly known as vocoding Some state-of-theart vocoding methods involve using generative adversarial networks [17[8, or denoising diffusion probabilistic models 4/19] for generating audio. Music source separation focuses on taking a mix of multiple music stems (vocals, drums, etc. ) and separating the mix into its individual sources Some approaches to music source separation operate by masking spectrograms |20] or directly mapping the mix waveform t0 individual source waveforms [21122/_ The music enhancement problem is different than music source separation, since our goal is not only to extract all musical sources from a noisy mixture but also to reduce reverb and adjust EQ such that the listening experience is improved dataset we assume access to high-quality recordings and define generative process for simulating low-quality ones. First, we simulate the reverb and varied microphone placements of a nonprofessional recording environment by convolving the high-quality music signal with a room impulse response. Next; we apply additive background noise scaled to achieve randomly sampled SNR between 5 and 30 dB. Finally, we simulate a low-quality microphone frequency response by applying 4-band equalization with randomly sampled gains between15 and 15 dB and frequency bands from 0-200, 200-1000, 1000-4000, and 4000-8000 Hz 3. 3. Mel-Spectrogram Enhancement with Mel2Mel Our first step in music enhancement is modeling the distribution of high-quality mel-spectrograms conditioned 0n their low-quality counterparts To estimate this distribution, we use existing work on image-to-image translation with conditional adversarial networks in an approach similar to [12|. In this framework a generator and a discriminator are trained using an aligned dataset of low and high-quality recording pairs. The generator maps from low to high-quality mel-spectrograms with the objective of maximizing the discriminator's loss and minimizing the C1 distance between the generated mel-spectrogram and the ground truth high-quality mel-spectrogram. The discriminator is trained to classify whether a given mel-spectrogram is generated or comes from the true data distribution. It performs this classification on patch-wise basis, predicting a class for each patch in the input melspectrogram_ For this reason, the discriminator acts aS a learned loss function for the generator which enforces realistic local features and the /1 loss enforces global consistency with the ground truth melspectrogram_ 3. 0 METHODS 3. 1. Modeling Approach In this paper; we investigate the approach of enhancing music in its mel-spectrogram domain, as it is easier to represent complex harmonic structures and polyphonic sound sources_ We then transform the resulting mel-spectrograms to waveforms through Diffwavebased vocoder (a process that in this context could be more aptly named "musecoding"). Decoupling waveform generation from melspectrogram enhancement allows us to train a musecoder that is not only robust to noise and other artifacts, but can also be used for any generation and enhancement task without the need of retraining: Figure [I depicts a block diagram of our proposed architecture: This approach is motivated by recent advances in vocoding that generate natural-sounding speech from mel-spectrograms [4]. 3. 4. Musecoding Recent work has shown that deep learning models can generate perceptually realistic waveforms from speech mel-spectrograms _ In our experiments, we evaluate the Diffwave |4] vocoder applied to music a process that we call "musecoding" Diffwave is a denoising diffusion probabilistic model (DDPM) This class of models defines forward diffusion process which iteratively adds gaussian noise to audio waveforms from the training dataset A model is then trained to estimate the reverse transition distributions of each noising step conditioned on the mel-spectrogram of the clean audio. Sampling from this model requires sampling noise from a standard gaussian and iteratively denoising using the reverse transition probability distributions from the model_ For further discussion of DDPMs see [4] and [23|. 3. 2 Data Simulation The modeling techniques we consider in this paper require aligned pairs of highand low-quality music recordings. To construct such\n\nModel Clean MOS 1 4. 39 = 0. 05 4. 06 = 0. 06 3. 01 + 0. 09 2. 85 =0. 09 4. 3. Baselines We evaluate our approach against two separate baselines First, we pair Mel2Mel for mel-spectrogram enhancement with inverse mel-scaling and the Griffin-Lim algorithm for musecoding. Both inverse mel-scaling and Griffin-Lim require solving optimization problems [291, So we run both solvers for 100 iterations, which yields per-sample runtime comparable to that of the Diffwave musecoder: Our second baseline is an end-to-end approach for music enhancement: Namely, we use the Demucs model architecture [21 and train it using the /1 reconstruction loss on our dataset of lowand high-quality recording pairs This matches the original training objective used for this architecture on the task of music source separation. We train this model for 360 epochs with batch size 64 and learning rate 0. 0003. We find that after this number of epochs the validation loss plateaus MelzMel + Diffwave Mel2Mel + Griffin-Lim No Enhancement Table 1. Mean Opinion Scores in a human listening test. As a musecoding baseline, we also consider mel-spectrogram inversion with inverse mel-scaling and the Griffin-Lim algorithm [14]. 4. EXPERIMENT SETUP 4. 1. Dataset We train and evaluate models on the Medley-solos-DB dataset [24]_ containing 21, 572 three-second, single-instrument samples recorded in professional studios. We exclude the distorted electric guitar samples to avoid fitting our models to production effects. We use 5841 samples for training, 3494 for validation and the rest for testing: We start by downsampling Our data to 16 kHz following the setup of prior vocoding work [4[71. This sample rate has shown to be favored by most speech enhancement work [D[2 and can be pOtentially super-resolved to 48 kHz with bandwidth extension techniques [5]. Using the procedure described in Section[. 2] we generate a dataset of 'highand low-quality recording pairs. For simulation of low-quality recordings, we source room impulse responses from the DNS Challenge dataset [25_ and realistic background noise from the ACE Challenge dataset |26/. As a final step, we apply a low-cut filter to remove nearly inaudible low frequencies below 35 Hz and normalize the waveforms to have a maximum absolute value of 0. 95_ We find that this treatment helps improve our models' training stability: When evaluating, we apply the same treatment (low-cut filter at 35 Hz and normalization) before applying OUr enhancement models. 4. 4. Evaluation Metrics To evaluate the results of different enhancement models we conducted Mean Opinion Score (MOS) test with human listeners on Amazon Mechanical Turk (AMT) Additionally; we evaluate enhancement methods by computing the frequency-weighted segmental SNR (fwSSNR) 1301, multi-resolution spectrogram loss (MRS) [34| C1 spectrogram distance, and Frechet Audio Distance (FAD) [32 between enhanced and clean reference signals_ In Sec tion|5. 3 we analyze the effectiveness of these objective metrics at approximating human listener ratings in the music domain. 5. 0 RESULTS 5. 1. Mean Opinion Score Test To evaluate our proposed MelzMel Diffwave music enhancement model, we conducted an MOS test with human listeners on AMT. We used 200 audio samples from our test set, added 8 different types of simulated degradation and passed these low-quality waveforms through our method, MelzMel + Griffin-Lim; and Demucs_ The lowquality, enhanced, and ground truth high-quality samples were then presented to human listeners who were asked to give a quality score from to 5. We used the ground truth high-quality recording as high anchor and the same recording with 0 dB white noise as low anchor Each Human Intelligence Task (HIT) started with screening test in which human listeners were required to identify which one of 5 audio samples sound the same aS a reference sample. 4 out of the 5 samples are passed a small amount of effects including low pass filters, high pass filters, comb filters, and added noise. Passing the screening test was required to continue The rest of the HIT consisted of 34 tests in which were validation tests to check if listeners were paying attention_ If they failed the validation test, the entire HIT was invalidated. In the end we collected 9, 095 answers from 214 listeners. The results shown in Table 1 suggests that MelzMel with a Diffwave musecoder achieves the highest MOS with a score near that of clean audio from the dataset_ 4. 2. Model Architectures and Hyperparameters In all experiments, we compute mel-spectrograms with 128 mel bins, an FFT size of 1024, and 256 sample hop length: When training models that generate O are conditioned 0n mel-spectrograms, we use log-= scale amplitudes to reduce the range of values and to avoid positive restrictions on our models domain Or range. The Mel2Mel generator described in SectionB3] consists of 2 downsampling blocks; each containing a 2D convolutional kernel of size 3 and stride 2, instance normalization |27 | and ReLU activation functions. This is followed by 3 ResNet blocks 28] with kernel size 3 and instance normalization_ Finally, the representation is upsampled back to the original dimensionality of the input with two upsampling blocks, each containing a transposed convolutional kernel of size 3 and stride 2, instance normalization, and ReLU activation functions. The Mel2Mel discriminator is a fully convolutional model made Up of three blocks, each containing a convolutional kernel of size 4 and stride 2, instance normalization, and LeakyReLU activation function. The last layer does not have any normalization O activation function Both the generator and discriminator are trained with batch size of 64 and learning rate of 0. 0002 for 200 epochs The Diffwave model described in SectionBAuses the architecture and training objective described in /4]. We train this model for 3000 epochs using a batch size of & and a learning rate of 0. 0002 5. 2. Perturbation Ablation Study To gain insight into which perturbations are handled most effectively by each enhancement model, we perform an ablation study isolating each perturbation introduced in the low-quality signal generative process Table[contains mean opinion scores for each enhancement\n\nModel Clean Random EQ 4. 35 = 0. 06 4. 15 =0. 07 2. 98 = 0. 1 3. 39 + 0. 10 3. 99 + 0. 08 SNR 5 SNR 10 4. 27 =0. 06 4. 24 = 0. 06 3. 53 + 0. 09 3. 07 = 0. 1 2. 71 +0. 1 SNR 15 4. 46 = 0. 06 3. 96 = 0. 09 3. 18 = 0. 11 2. 85 = 0. 11 3. 04 = 0. 12 DRR 0 DRR 3 DRR 6 4. 24 =0. 07 4. 01 = 0. 08 3. 10 = 0. 08 2. 55 = 0. 10 2. 48 + 0. 11 4. 28 = 0. 04 4. 19 = 0. 06 3. 77 = 0. 06 3. 84 = 0. 06 2. 82 + 0. 07 2. 77 _ 0. 09 3. 13 = 0. 07 3. 21 + 0. 07 4. 01 = 0. 06 3. 91 = 0. 07 4. 42 =0. 06 3. 96 = 0. 08 2. 99 + 0. 10 3. 30 + 0. 10 4. 21 +0. 07 Mel2Mel + Diffwave Mel2Mel + Griffin-Lim Demucs No Enhancement Table 2. Mean Opinion Scores in a human listening test. Each column contains the ratings for single perturbation type: EQ, additive background noise at different signal-to-noise ratios (SNR), and reverb at different direct-to-reverberant ratios (DRR)_ Enhancement Metric Rank Correlation with MOS fwSSNR 0. 5 ~MRS 056 ~LI 0. 4 ~FAD 053 Model fwSSNR 9. 04 7. 61 6. 58 8. 23 6. 96 MRS 1. 40 1. 57 1. 65 1. 80 1. 89 Ll + 1350 1, 57 1. 69 1. 83 2. 16 FAD 4. 73 4. 54 3. 98 5. 54 5. 90 Independent Training Joint Fine-tuning Joint Training Sequential Training No Enhancement Table 3. Spearman rank correlation between MOS test ratings and audio enhancement metrics. Table 4. Performance of MelzMel + Diffwave enhancement models using different training schemes model applied to signals with randomly sampled EQ, additive noise with signal-to-noise ratios (SNR) of 5, 10, and 15 dB, and reverb with direct-to-reverberant ratios (DRR) of 0, 3, and 6 dB. This ablation shows that the Mel2Mel Diffwave model excels at removing noise even at SNR values as low as 5 dB and at undoing 4-band equalization simulating a non-flat microphone frequency response. Interestingly, none of the models tested perform dereverberation very well, and in fact degrade signals that contain no noise and only simulated reverb: This may be due to train-test mismatch; since all samples enhanced during training time contained some level of additive noise. This ablation also lends insight into the types of perturbations that affect human listeners"perception of music_ From the difference between the scores given to clean samples and non-enhanced samples, it is clear that additive noise impacts the listener'$ perception significantly while reverb is mostly ignored. 5. 4. Alternate Training Schemes In Section |3. 1] we motivated approaching music enhancement by training two decoupled models that separately handle melspectrogram enhancement and musecoding: Here, we investigate training schemes for these models other than independently training them on their respective tasks In addition t0 independent training, we (1) finetune the Mel2Mel generator and Diffwave musecoder jointly using the Diffwave objective, (2) train the models sequentially by first training the musecoder and then training the Mel2Mel generator with musecoder parameters frozen; and (3) train the MelzMel generator and musecoder jointly as single model using the Diffwave objective. Table shows the performance of the resulting models. In Section[3]we discussed the reliability of using these metrics for evaluating algorithms, and find that FAD is the most perceptually aligned metric when it comes to denoising Given this observation, our results suggest that joint training may yield better denoising performance than independent training: Joint training has the added benefit that only a single model is trained using a non-adversarial objective. However; this comes with the downside that the trained model cannot be split into enhancement and musecoding sub-models: Future work could focus on further exploring such training schemes. 5. 3. Perceptual Alignment of Objective Metrics The results of the MOS test also provide a mechanism t0 evaluate how well objective metrics for audio quality align with human perception in the music domain We measure fwSSNR, MRS, FAD, and 61 spectrogram distance on the same samples submitted for MOS evaluation_ We then take the mean score across all samples with given perturbation type (i. e. SNR 5, DRR 0, etc. ) and perform Spearman rank correlation with the mean scores measured in the human MOS test_ In Tablel3] we show the rank correlation for each objective metric_ We find that none of the four metrics evaluated correlate very strongly with human opinion scores, the highest achieving a rank correlation of 0. 56_ We also identify particular failure modes of these metrics AlL four metrics fail to identify robotic artifacts induced by the GriffinLim algorithm and actually rate the Mel2Mel + Griffin-Lim model as the best of all models we tested. Additionally, fwSSNR MRS, and 61 spectrogram distance all fail to identify additive noise effectively, and rate non-enhanced samples at SNR values of 10 and 15 dB as being better than any enhancement model output: FAD does not have this failure mode. 6. CONCLUSION We propose a music enhancement model that decomposes the task into mel-spectrogram enhancement and waveform synthesis from mel-spectrograms_ This model was trained using high-quality samples from a public dataset paired with low-quality samples generated by simulating artifacts that typically appear in amateur recordings_ A human MOS test shows that this model outperforms state-of-theart baselines Additionally, we found that current objective metrics for audio enhancement do not accurately reflect human perception of music. We hope this work encourages researchers to further advance the rather unexplored and yet timely topic of automatic music enhancement; either by designing more performant models 01 by proposing metrics that better align with human music perception.\n\n7 _ REFERENCES [17] Kundan Kumar; Rithesh Kumar; Thibault de Boissiere, Lucas Gestin, Wei Zhen Teoh, Jose Sotelo, Alexandre de Brebisson Yoshua Bengio, and Aaron Courville_ "Melgan: Generative adversarial networks for conditional waveform synthesis,"2019. [18] Jaeseong You_ Dalhyun Kim; Gyuhyeon Nam, Geumbyeol Hwang, and Gyeongsu Chae, "Gan vocoder: Multi-resolution discriminator is all you need;" 2021. [19] Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss Mohammad Norouzi, and William Chan, Wavegrad: Estimating gradients for waveform generation; 2020. [20] Romain Hennequin, Anis Khlif, Felix Voituret; and Manuel Moussallam_ "Spleeter: fast and efficient music source separation tool with pre-trained models Journal of Open Source Software, vol. 5, pp. 2154, 06 2020. [21] Alexandre Defossez, Nicolas Usunier; Leon Bottou, and Francis Bach; "Music source separation in the waveform domain; 2021. [22] Yi Luo and Nima Mesgarani_ "Conv-tasnet: Surpassing ideal time-frequency magnitude masking for speech separation;' IEEEIACM TASLP vol. 27, no. &, pp. 1256-1266, Aug 2019. [23] Jonathan Ho, Ajay Jain, and Pieter Abbeel, "Denoising diffusion probabilistic models;' 2020. [24] Vincent Lostanlen and Carmine-Emanuele Cella, Deep convolutional networks on the pitch spiral for musical instrument recognition;' 2017. [25] Chandan K Reddy, Harishchandra Dubey; Kazuhito Koishida, Arun Nair; Vishak Gopal_ Ross Cutler Sebastian Braun, Hannes Gamper; Robert Aichner; and Sriram Srinivasan ~Interspeech 2021 deep noise suppression challenge;' 2021 _ [26] J. Eaton, N. D. Gaubitch, A_ H. Moore, and P A. Naylor; "The ace challenge corpus description and performance evaluation,'in 2015 IEEE WASPAA, 2015, pp. 1-5_ [27] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky, Instance normalization: The missing ingredient for fast stylization, ; 2017. [28] Kaiming He, Xiangyu Zhang, Shaoqing Ren; and Jian Sun_ Deep residual learning for image recognition;' in 2016 IEEE CVPR, 2016, pp. 770-778_ [29] Yao-Yuan Yang; Moto Hira, Zhaoheng Ni, Anjali Chourdia, Artyom Astafurov, Caroline Chen, Ching-Feng Yeh; Christian Puhrsch_ David Pollack; Dmitriy Genzel, Donny Greenberg, Edward Z Yang; Jason Lian, Jay Mahadeokar; Jeff Hwang; Ji Chen, Peter Goldsborough, Prabhat Roy; Sean Narenthiran, Shinji Watanabe, Soumith Chintala, Vincent Quenneville-Belair; and Yangyang Shi, Torchaudio: Building blocks for audio and speech processing; arXiv preprint arXiv:2110. 15018, 2021. [30] YHu and Philipos C. Loizou; valuation of objective quality measures for speech enhancement_ IEEE TASLP vol: 16, pp_ 229-238,. 2008. [31] Ryuichi Yamamoto, Eunwoo Song; and Jae-Min Kim; Parallel wavegan: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram;" 2020_ [32] Kevin Kilgour; Mauricio Zuluaga; Dominik Roblek; and Matthew Sharifi, Frechet audio distance: A metric for evaluating music enhancement algorithms,"2019. 1] Jiaqi Su, Zeyu Jin, and Adam Finkelstein, Hifi-gan: Highfidelity denoising and dereverberation based o speech deep features in adversarial networks 2020. [2] Alexandre Defossez, Gabriel Synnaeve, and Yossi Adi, "Real time speech enhancement in the waveform domain, Interspeech2020, 2020. [3] Phillip Isola, Jun-Yan Zhu; Tinghui Zhou, and Alexei A Efros, ~Image-to-image translation with conditional adversarial networks_ 2018_ [4] Zhifeng Kong; Wei Ping; Jiaji Huang, Kexin Zhao, and Bryan Catanzaro, Diffwave: versatile diffusion model for audio synthesis,'2021_ [5] Jiaqi Su, Yunyun Wang. Adam Finkelstein; and Zeyu Jin Bandwidth extension is all you need, in ICASSP 2021-2021. IEEE, 2021, pp. 696-700. [6] Pranay Manocha, Zeyu Jin, Richard Zhang; and Adam Finkelstein, "Cdpam: Contrastive learning for perceptual audio similarity;" in ICASSP 2021-2021. IEEE, 2021, pp. 196-200. [7] Chandan KA Reddy, Vishak Gopal, and Ross Cutler; Dnsmos: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors in ICASSP 2021-2021. IEEE, 2021, pp. 6493-6497. [8] P Scalart and J. V. Filho, "Speech enhancement based on priori signal to noise estimation;" in 1996 IEEE ICASSP Proceedings, 1996, vol. 2, pp. 629-632 vol: 2 [9] Hideaki Kagami, Hirokazu Kameoka; and Masahiro Yukawa, "Joint separation and dereverberation of reverberant mixtures with determined multichannel non-negative matrix factorization,'in 2018 IEEE ICASSP, 2018, pp. 31-35_ [10] Kun Han, Yuxuan Wang, DeLiang Wang; William $. Woods, Ivo Merks, and Tao Zhang; Learning spectral mapping for speech dereverberation and denoising;' IEEEIACM TASLP, vol. 23, no. 6, pp. 982-992, 2015. [11] Donald S_ Williamson and DeLiang Wang; "Speech dereverberation and denoising using complex ratio masks;" in 2017 IEEE ICASSP, 2017, pp. 5590-5594. [12] Daniel Michelsanti and Zheng-Hua Tan; "Conditional generative adversarial networks for speech enhancement and noiserobust speaker verification,"Interspeech 2017, Aug 2017. [13] Santiago Pascual, Joan Serra, and Antonio Bonafonte, ~Towards generalized speech enhancement with generative adversarial networks,'2019. [14] D. Griffin and Jae Lim; ~Signal estimation from modified short-time fourier transform IEEE Transactions on Acoustics, Speech, and Signal Processing, vol_ 32, no. 2, Pp. 236 243, 1984 [15] Adam Polyak; Lior Wolf;, Yossi Adi, Ori Kabeli, and Yaniv Taigman, High fidelity speech regeneration with application to speech enhancement;' in ICASSP 2021-2021. IEEE, 2021, pp. 7143-7147 [16] Jiaqi Su, Zeyu Jin, and Adam Finkelstein; "Hifi-gan-2: studioquality speech enhancement via generative adversarial networks conditioned on acoustic features;' in 2015 IEEE WASPAA, 2021_ |
| 14 | script_findingnemo.txt | 04a90337-527 | Script | ------------------------------------------------------------\n\nFINDING NEMO Transcript v1.0\n\nCopyright 2003 Walt Disney Pictures, Pixar Animation Studios\n\n------------------------------------------------------------\n\nTranscribed by BaD_BURN\n\nemail : markgonzalez154@hotmail.com\n\n ------------------------------------------------------------------\n\n| Okay, this is the work-in-progress FINDING NEMO film transcript. |\n\n| Why is it 'work-in-progress' you might ask? Well for one, this |\n\n| isn't a 100% accurate transcript: some words might be missing, |\n\n| may not be right. Second, some lines may or may not have been |\n\n| spoken by the right character. There are instances in the film |\n\n| where a line is spoken but the character isn't on screen, which |\n\n| makes things complicated. But I'd say this transcript is about |\n\n| 98-99% accurate. Dialogue for each scene is seperated by a line |\n\n| of equal signs (=). |\n\n| |\n\n| This transcript is open for corrections, additions if you have |\n\n| any. What you CAN'T do, however, is to edit it and take credit |\n\n| for it. Although I do not own the movie or it's screenplay, this |\n\n| transcript was made with no intention of copyright infringement |\n\n| and the like. Enjoy. And remember: 'Fish are friends, not food'. |\n\n ------------------------------------------------------------------\n\n======================================================================================\n\nMARLIN\n\nWow.\n\nCORAL\n\nMmm.\n\nMARLIN\n\nWow.\n\nCORAL\n\nMmm-hmm.\n\nMARLIN\n\nWow.\n\nCORAL\n\nYes, Marlin. No, I see it. It's beautiful.\n\nMARLIN\n\nSo, Coral, when you said you wanted an ocean view, you didn't think that we we're gonna\n\nget the whole ocean, did you? Huh? [sighs] Oh yeah. A fish can breath out here. Did your\n\nman deliver or did he deliver?\n\n 1\n\nCORAL\n\nMy man delivered.\n\nMARLIN\n\nAnd it wasn't so easy.\n\nCORAL\n\nBecause a lot of other clownfish had their eyes on this place.\n\nMARLIN\n\nYou better believe they did--every single one of them.\n\nCORAL\n\nMm-hmm. You did good. And the neighborhood is awesome.\n\nMARLIN\n\nSo, you do like it, don't you?\n\nCORAL\n\nNo, no. I do, I do. I really do like it. But Marlin, I know that the drop off is desirable\n\nwith the great schools and the amazing view and all, but do we really need so much space?\n\nMARLIN\n\nCoral, honey, these are our kids we're talking about. They deserve the best. Look, look,\n\nlook. They'll wake up, poke their little heads out and they'll see a whale! See, right by\n\ntheir bedroom window.\n\nCORAL\n\nShhh, you're gonna wake the kids.\n\nMARLIN\n\nOh, right. Right.\n\nCORAL\n\nAww, look. They're dreaming. We still have to name them.\n\nMARLIN\n\nYou wanna name all of 'em, right now? All right, we'll name this half Marlin Jr. and then\n\nthis half Coral Jr. Okay, we're done.\n\nCORAL\n\nI like Nemo.\n\nMARLIN\n\nNemo? Well, we'll name one Nemo but I'd like most of them to be Marlin Jr.\n\nCORAL\n\nJust think that in a couple of days, we're gonna be parents!\n\nMARLIN\n\nYeah. What if they don't like me?\n\nCORAL\n\nMarlin.\n\nMARLIN\n\nNo, really.\n\nCORAL\n\nThere's over 400 eggs. Odds are, one of them is bound to like you.\n\nCORAL\n\nWhat?\n\nMARLIN\n\nYou remember how we met?\n\nCORAL\n\nWell, I try not to.\n\nMARLIN\n\nWell, I remember. 'Excuse me, miss, can you check and see if there's a hook in my lip?'\n\nCORAL\n\nMarlin!\n\nMARLIN\n\n 2\n\n'Well, you gotta look a little closer because it's wiggling'.\n\nCORAL\n\nGet away!\n\nMARLIN\n\nHere he is. Cutie's here! Where did everybody go?\n\nMARLIN\n\n[gasps] Coral, get inside the house, Coral. No, Coral, don't. They'll be fine. Just get\n\ninside, you, right now.\n\nMARLIN\n\nNo!\n\nMARLIN\n\nCoral! Coral?\n\nMARLIN\n\nCoral? Oh!\n\nMARLIN\n\nOhh. There, there, there. It's okay, daddy's here. Daddy's got you. I promise, I will\n\nnever let anything happen to you...Nemo.\n\n======================================================================================\n\nNEMO\n\nFirst day of school! First day of school! Wake up, wake up! C'mon, first day of school!\n\nMARLIN\n\nI don't wanna go to school. Five more minutes.\n\nNEMO\n\nNot you, dad. Me!\n\nMARLIN\n\nOkay...huh?\n\nNEMO\n\nGet up, get up! It's time for school! It's time for school! It's time for school!\n\nIt's time for school! Oh boy! Oh boy!\n\nMARLIN\n\nAll right, I'm up.\n\nNEMO\n\nOh boy--whoa!\n\nMARLIN\n\nNemo!\n\nNEMO\n\nFirst day of school!\n\nMARLIN\n\n[gasps] Nemo, don't move! Don't move! You'll never get out of there yourself. I'll do it.\n\nAll right, where's the break? You feel a break?\n\nNEMO\n\nNo.\n\nMARLIN\n\nSometimes you can't tell 'cause fluid is rushing to the area. Now, any rushing fluids?\n\nNEMO\n\nNo.\n\nMARLIN\n\nAre you woozy?\n\nNEMO\n\nNo.\n\nMARLIN\n\nHow many stripes do I have?\n\n 3\n\nNEMO\n\nI'm fine.\n\nMARLIN\n\nAnswer the stripe question!\n\nNEMO\n\nThree.\n\nMARLIN\n\nNo! See, something's wrong with you. I have one, two, three--that's all I have? Oh,\n\nyou're okay. How's the lucky fin?\n\nNEMO\n\nLucky.\n\nMARLIN\n\nLet's see.\n\nMARLIN\n\nAre you sure you wanna go to school this year? 'Cause there's no problem if you don't.\n\nYou can wait 5 or 6 years.\n\nNEMO\n\nCome on, dad. It's time for school.\n\nMARLIN\n\nAh-ah-ah! Forgot to brush.\n\nNEMO\n\nOhh...\n\nMARLIN\n\nDo you want this anemone to sting you?\n\nNEMO\n\nYes.\n\nMARLIN\n\nBrush.\n\nNEMO\n\nOkay, I'm done.\n\nMARLIN\n\nYou missed a spot.\n\nNEMO\n\nWhere?\n\nMARLIN\n\nThere. Ha ha! Right there. And here and here and here!\n\n======================================================================================\n\nMARLIN\n\nAll right, we're excited. First day of school, here we go. We're ready to learn to get\n\nsome knowledge. Now, what's the one thing we have to remember about the ocean?\n\nNEMO\n\nIt's not safe.\n\nMARLIN\n\nThat's my boy. So, first we check to see that the coast is clear. We go out and back in.\n\nAnd then we go out, and back in. And then one more time--out and back in. And sometimes,\n\nif you wanna do it four times--\n\nNEMO\n\nDad..\n\nMARLIN\n\nAll right. Come on, boy.\n\nNEMO\n\nDad, maybe while I'm at school, I'll see a shark!\n\nMARLIN\n\n 4\n\nI highly doubt that.\n\nNEMO\n\nHave you ever met a shark?\n\nMARLIN\n\nNo, and I don't plan to.\n\nNEMO\n\nHow old are sea turtles?\n\nMARLIN\n\nSea turtles? I don't know.\n\nNEMO\n\nSandy Plankton from next door, he said that sea turtles, said that they live to be about\n\na hundred years old!\n\nMARLIN\n\nWell, you know what, if I ever meet a sea turtle, I'll ask him. After I'm done talking\n\nto the shark, okay? Whoa, whoa, whoa! Hold on, hold on, wait to cross. Hold my fin,\n\nhold my fin.\n\nNEMO\n\nDad, you're not gonna freak out like you did at the petting zoo, are you?\n\nMARLIN\n\nHey, that snail was about to charge. Hmm, I wonder where we're supposed to go.\n\nFISH KIDS\n\nBye, mom!\n\nFISH MOM\n\nI'll pick you up after school.\n\nCRAB KID\n\nCome on, you guys. Stop it! Give it back!\n\nMARLIN\n\nCome on, we'll try over there.\n\nMARLIN\n\nExcuse me, is this where we meet his teacher?\n\nBOB\n\nWell, look who's out of the anemone.\n\nMARLIN\n\nYes. Shocking, I know.\n\nBOB\n\nMarty, right?\n\nMARLIN\n\nMarlin.\n\nBOB\n\nBob.\n\nTED\n\nTed.\n\nBILL\n\nBill. Hey, you're a clownfish. You're funny, right? Hey, tell us a joke.\n\nBOB/TED\n\nYeah, yeah. Come on, give us a funny one.\n\nMARLIN\n\nWell, actually, that's a common misconception. Clownfish are no funnier than any\n\nother fish.\n\nBILL\n\nAw, come on, clownie.\n\nTED\n\nYeah, do something funny.\n\n 5\n\nBOB\n\nYeah!\n\nMARLIN\n\nAll right, I know one joke. Um, there's a mollusk, see? And he walks up to a sea, well he\n\ndoesn't walk up, he swims up. Well, actually the mollusk isn't moving. He's in one place\n\nand then the sea cucumber, well they--I mixed up. There was a mollusk and a sea cucumber.\n\nNone of them were walking, so forget that I--\n\nBOB\n\nSheldon! Get out of Mr. Johansenn's yard, now!\n\nKIDS\n\nWhoa!\n\nMR. JOHANSSEN\n\nAll right, you kids! Ooh! Uuh, where'd you go? Where'd you go? Where, where'd you go?\n\nNEMO\n\nDad, dad...can I go play too? Can I?\n\nMARLIN\n\nI would feel better if you go play over on the sponge beds.\n\nMARLIN\n\nThat's where I would play\n\nPEARL\n\nWhat's wrong with his fin?\n\nTAD\n\nHe looks funny!\n\nSHELDON\n\nOw! Hey, what'd I do? What'd I do?\n\nBOB\n\nBe nice. It's his first time at school.\n\nMARLIN\n\nHe was born with it, kids. We call it his lucky fin.\n\nNEMO\n\nDad.\n\nPEARL\n\nSee this tentacle? It's actually shorter than all my other tentacles but you can't really\n\ntell.Especially when I twirl them like this.\n\nSHELDON\n\nI'm H2O-intolerant. [sneezes]\n\nTAD\n\nI'm obnoxious.\n\nMR. RAY\n\n[singing] Oooh, let's name the zones, the zones, the zones. Let's name the zones of the\n\nopen sea.\n\nKIDS\n\nMr. Ray!\n\nSHELDON\n\nCome on, Nemo.\n\nMARLIN\n\nWhoa, you better stay with me.\n\nMR. RAY\n\n[singing]..mesopolagic, bathyal, abyssalpelagic. All the rest are too deep for you and\n\nme to see.\n\nMR. RAY\n\nHuh, I wonder where my class has gone?\n\nKIDS\n\n 6\n\nWe're under here!\n\nMR. RAY\n\nOh, there you are. Climb aboard, explorers. [singing] Oh, knowledge exploring is oh so\n\nlyrical, when you think thoughts that are empirical.\n\nNEMO\n\nDad, you can go now.\n\nMR. RAY\n\nWell, hello. Who is this?\n\nNEMO\n\nI'm Nemo.\n\nMR. RAY\n\nWell, Nemo, all new explorers must answer a science question.\n\nNEMO\n\nOkay.\n\nMR. RAY\n\nYou live in what kind of home?\n\nNEMO\n\nAn anemo-none. A nemenem-menome-nememen-nenemone--\n\nMR. RAY\n\nOkay, okay, don't hurt yourself. Welcome aboard, explorers!\n\nMARLIN\n\nJust so you know, he's got a little fin. I find if he's having trouble swimming, let him\n\ntake a break. Ten, fifteen minutes.\n\nNEMO\n\nDad, it's time for you to go now.\n\nMR. RAY\n\nDon't worry. We're gonna stay together as a group. Okay, class, optical orbits up front.\n\nAnd remember, we keep our supraesophogeal ganglion to ourselves...that means you, Jimmy.\n\nJIMMY\n\nAw, man!\n\nMR. RAY\n\n[singing]\n\nMARLIN\n\nBye, Nemo!\n\nNEMO\n\nBye, dad!\n\nMARLIN\n\nBye, son! Be safe.\n\nBOB\n\nHey, you're doing pretty well for a first timer.\n\nMARLIN\n\nWell, you can't hold onto them forever, can you?\n\nBILL\n\nYeah, I had a tough time when my oldest went out at the drop off.\n\nMARLIN\n\nThey just gotta grow up--the drop off?! They're going to the drop off?! Wh-what are you,\n\ninsane?! Why don't we fry 'em up now and serve them with chips!?\n\nBOB\n\nHey, Marty. Calm down.\n\nMARLIN\n\nDon't tell me to be calm, pony boy!\n\nBOB\n\n'Pony boy'?\n\n 7\n\nBILL\n\nYou know for a clownfish, he really isn't that funny.\n\nTED\n\nPity.\n\n======================================================================================\n\nMR. RAY\n\n[singing] Oh, let's name the species, the species, the species. Let's name the species\n\nthat live in thesea.\n\nNEMO\n\nWhoa.\n\nMR. RAY\n\n[singing] There's porifera, coelenterata, hydrozoa, scyphozoa, anthozoa, ctenophora,\n\nbryozoas, three! Gastropoda, arthropoda, echinoderma, and some fish like you and me. Come\n\non, sing with me. Oh...!\n\nMR. RAY\n\nJust the girls this time. [singing] Oh, seaweed is cool. Seaweed is fun. It makes it's food\n\nwith the rays of the sun...\n\nMR. RAY\n\nOkay, the drop off. All right, kids, feel free to explore but stay close. [gasps]\n\nStromalitic cyanobacteria! Gather. An entire ecosystem contained in one infinitesimal speck.\n\nThere are as many protein pairs contained in this...\n\nTAD\n\nCome on, let's go.\n\nMR. RAY\n\nCome on, sing with me! [singing] There's porifera, coelentera, hydrozoa, scyphozoa, anthozoa,\n\nctenophora, bryozoas, three!\n\nNEMO\n\nHey guys, wait up! Whoa.\n\nTAD\n\nCool.\n\nTAD\n\nSaved your life!\n\nPEARL\n\nAw, you guys made me ink.\n\nNEMO\n\nWhat's that?\n\nTAD\n\nI know what that is. Oh, oh! Sandy Plankton saw one. He called, he said it was called a...a\n\nbutt.\n\nNEMO\n\nWhoa.\n\nPEARL\n\nWow. That's a pretty big butt.\n\nSHELDON\n\nOh, look at me. I'm gonna go touch the butt. [sneezes] Whoa!\n\nSHELDON\n\nOh yeah? Let's see you get closer.\n\nPEARL\n\nOkay. Beat that.\n\nTAD\n\nCome on, Nemo. How far can you go?\n\nNEMO\n\nUh, my dad says it's not safe.\n\n 8\n\nMARLIN\n\nNemo, no!\n\nNEMO\n\nDad?\n\nMARLIN\n\nYou were about to swim into open water!\n\nNEMO\n\nNo, I wasn't go out--but dad!\n\nMARLIN\n\nIt was a good thing I was here. If I hadn't showed up, I don't know--\n\nPEARL\n\nSir, he wasn't gonna go.\n\nTAD\n\nYeah, he was too afraid.\n\nNEMO\n\nNo, I wasn't.\n\nMARLIN\n\nThis does not concern you, kids. And you're lucky I don't tell your parents you were out\n\nthere.\n\nYou know you can't swim well.\n\nNEMO\n\nI can swim fine, dad, okay?\n\nMARLIN\n\nNo, it's not okay. You shouldn't be anywhere near here. Okay, I was right. You'll start school\n\nin a year or two.\n\nNEMO\n\nNo, dad! Just because you're scared of the ocean--\n\nMARLIN\n\nClearly, you're not ready. And you're not coming back until you are. You think you can do\n\nthese\n\nthings but you just can't, Nemo!\n\nNEMO\n\nI hate you.\n\nMR. RAY\n\nThere's--nothing to see. Gather, uh, over there. Excuse me, is there anything I can do? I am a\n\nscientist, sir. Is there any problem?\n\nMARLIN\n\nI'm sorry. I didn't mean to interrupt things. He isn't a good swimmer and it's a little\n\ntoo soon for him to be out here unsupervised.\n\nMR. RAY\n\nWell, I can assure you, he's quite safe with me.\n\nMARLINLook, I'm sure he is. But you have a large class and he can get lost\n\nfrom sight if you're not looking. I'm not saying you're not looking--\n\nFISH KID\n\nOh my gosh! Nemo's swimming out to sea!\n\nMARLIN\n\nNemo! What do you think you're doing? You're gonna get stuck out there and I'll have to get\n\nyou before another fish does! Get back here! I said get back here, now! Stop! You take one\n\nmove, mister. Don't youdare! If you put one fin on that boat..are you listening to me?\n\nDon't touch the bo--Nemo!\n\nTAD\n\n[whispering] He touched the butt.\n\nMARLIN\n\nYou paddle your little tail back here, Nemo. That's right. You are in big trouble, young man.\n\nDo you hear me? Big...big--\n\n 9\n\nNEMO\n\nAaaah! Daddy! Help me!\n\nMARLIN\n\nI'm coming, Nemo!\n\nKIDS\n\nAaaah!\n\nMR. RAY\n\nGet under me, kids!\n\nNEMO\n\nAh! Oh no! Dad! Daddy!\n\nMARLIN\n\nOh! Nemo! Unh! Nemo! Nemo, no! Nemo! Nemo! Nemo! No! No! Aah! Nemo! Nemo!\n\nDIVER\n\nWhoa! Hold on.\n\nMARLIN\n\nOh no. No, no. It's gone, it's gone. No, no, it can't be gone. No, no! Nemo! Nemo! Nemo! No!\n\nNemo! Nemo! No! No, please, no! No, no!\n\nMARLIN\n\nHas anybody seen a boat!? Please! A white boat! They took my son! My son! Help me, please!\n\nDORY\n\nLook out!\n\nMARLIN\n\nWaaaah!\n\nMARLIN\n\nOoh, ooh...\n\nDORY\n\nOhh. Oh, oh. Sorry! I didn't see you. Sir, are you okay?\n\nMARLIN\n\nHe's gone, he's gone..\n\nDORY\n\nThere, there. It's all right.\n\nMARLIN\n\nHe's gone.\n\nDORY\n\nIt'll be okay.\n\nMARLIN\n\nNo, no. They took him away. I have to find the boat.\n\nDORY\n\nHey, I've seen a boat.\n\nMARLIN\n\nYou have?\n\nDORY\n\nIt passed by not too long ago.\n\nMARLIN\n\nA white one?\n\nDORY\n\nHi. I'm Dory.\n\nMARLIN\n\nWhere!? Which way!?\n\nDORY\n\nOh, oh, oh! It-it went, um, this way! And it went this way! Follow me!\n\nMARLIN\n\n 10\n\nThank you! Thank you, thank you so much!\n\nDORY\n\nNo problem.\n\nMARLIN\n\nHey! Wait!\n\nDORY\n\nWill you quit it?\n\nMARLIN\n\nWhat?\n\nDORY\n\nI'm trying to swim here. What, ocean ain't big enough for you?\n\nMARLIN\n\nHuh?\n\nDORY\n\nYou got a problem, buddy? Huh? Huh? Do 'ya? Do 'ya? Do 'ya? You want a piece of me? Yeah,\n\noooh, I'm scared now. Whaat!?\n\nMARLIN\n\nWait a minute..\n\nDORY\n\nStop following me, okay!?\n\nMARLIN\n\nWhat? You're showing me which way the boat went!\n\nDORY\n\nA boat? Hey, I've seen a boat. It passed by not too long ago. It went this way, it went this\n\nway. Follow me!\n\nMARLIN\n\nWait a minute, wait a minute! What is going on? You already told me which way the boat\n\nwas going!\n\nDORY\n\nI did? Oh dear...\n\nMARLIN\n\nIf this is some kind of practical joke, it's not funny! And I know funny..I'm a clownfish!\n\nDORY\n\nNo, it's not. I know it's not. I'm so sorry. See, I suffer from short-term memory loss.\n\nMARLIN\n\nShort-term memory loss..I don't believe this!\n\nDORY\n\nNo, it's true. I forget things almost instantly. It runs in my family..or at least I think\n\nit does. Hmmm..where are they? Can I help you?\n\nMARLIN\n\nSomething's wrong with you, really. You're wasting my time. I have to find my son. [gasps]\n\nBRUCE\n\nHello.\n\nDORY\n\nWell, hi!\n\nBRUCE\n\nName's Bruce. It's all right, I understand. Why trust a shark, right? So, what's a couple of\n\nbites like you doing out so late, eh?\n\nMARLIN\n\nNothing. We're not doing anything. We're not even out.\n\nBRUCE\n\nGreat! Then how'd you morsels like to come to a little get-together I'm havin'?\n\nDORY\n\n 11\n\nYou mean like a party?\n\nBRUCE\n\nYeah, yeah, that's right--a party! What do you say?\n\nDORY\n\nOoh, I love parties! Parties are fun!\n\nMARLIN\n\nParties are fun, and it's tempting but--\n\nBRUCE\n\nOh, come on, I insist.\n\nMARLIN\n\nO-okay..that's all that matters.\n\nDORY\n\nHey, look--balloons! It is a party!\n\nBRUCE\n\nHa ha ha! Mind your distance, though. Those balloons can be a bit dodgy. You wouldn't want\n\none of them to pop.\n\nBRUCE\n\nAnchor! Chum!\n\nANCHOR\n\nThere you are, Bruce, finally!\n\nBRUCE\n\nWe got company.\n\nANCHOR\n\nIt's about time, mate.\n\nCHUM\n\nWe've already gone through all the snacks and I'm still starvin'!\n\nANCHOR\n\nWe almost had a feeding frenzy.\n\nCHUM\n\nCome on, let's get this over with.\n\n======================================================================================\n\nBRUCE\n\nRight, then. The meeting has officially come to order. Let us all say the pledge..\n\nBRUCE/ANCHOR/CHUM\n\n'I am a nice shark, not a mindless eating machine. If I am to change this image, I must\n\nfirst change myself. Fish are friends, not food'.\n\nANCHOR\n\nExcept stinkin' dolphins.\n\nCHUM\n\nDolphins! Yeah, they think they're sooo cute! 'Hey, look at me. I'm a flippin' little dolphin!\n\nLet me flip for 'ya! Ain't I a somethin'!'\n\nBRUCE\n\nRight, then. Today's meeting is step 5, 'BRING A FISH FRIEND'. Now do you all have your\n\nfriends?\n\nANCHOR\n\nGot mine.\n\nDORY\n\nHey there!\n\nBRUCE\n\nHow 'bout you, Chum?\n\nCHUM\n\nOh, um, I seem to have misplaced my uh, friend.\n\n 12\n\nBRUCE\n\nThat's all right, Chum. I had a feeling this would be a difficult step, you can help yourself\n\nto one of my friends.\n\nCHUM\n\nOh, thanks, mate. A little chum for Chum, eh?\n\nBRUCE\n\nI'll start the testimonies. Hello, my name is Bruce.\n\nANCHOR/CHUM\n\nHello, Bruce.\n\nBRUCE\n\nIt has been three weeks since my last fish, on my honor, or may I be chopped up and\n\nmade into soup.\n\nCHUM\n\nYou're an inspiration to all of us.\n\nANCHOR\n\nAmen.\n\nBRUCE\n\nRight, then. Who's next?\n\nDORY\n\nOoh! Pick me! Pick me!\n\nBRUCE\n\nYes, the little Sheila down the front.\n\nDORY\n\nWoo-hoo!\n\nBRUCE\n\nCome on up here.\n\nDORY\n\nHi. I'm Dory.\n\nBRUCE/ANCHOR/CHUM\n\nHello, Dory.\n\nDORY\n\nAnd, uh, well, I don't think I've ever eaten a fish.\n\nCHUM\n\nHey, that's incredible.\n\nBRUCE\n\nGood on 'ya, mate!\n\nDORY\n\nWhew! I'm glad I got that off my chest.\n\nBRUCE\n\nAll right, anyone else? Hello, how 'bout you, mate? What's your problem?\n\nMARLIN\n\nMe? I don't have a problem.\n\nBRUCE\n\nOh. Okay..\n\nBRUCE/ANCHOR/CHUM\n\nDenial.\n\nBRUCE\n\nJust start with your name.\n\nMARLIN\n\nOkay. Uh, hello. My name is Marlin. I'm a clownfish--\n\nCHUM\n\nA clownfish? Really?!\n\n 13\n\nBRUCE\n\nGo on, tell us a joke!\n\nCHUM\n\nOoh! I love jokes!\n\nMARLIN\n\nActually I do know one that's pretty good. There was this mollusk and he walks up to a sea\n\ncucumber. Normally, they don't talk, sea cucumbers, but in a joke, everyone talks. So the\n\nsea mollusk says to the cucumber...\n\nNEMO\n\nDaddy!\n\nMARLIN\n\nNemo!\n\nCHUM\n\nNemo! Ha ha ha! Nemo! I don't get it.\n\nBRUCE\n\nFor a clownfish, he's not that funny.\n\nMARLIN\n\nNo, no, no, no. He's my son. He was taken by these divers.\n\nDORY\n\nOh my, you poor fish.\n\nCHUM\n\nHumans. Think they own everything.\n\nANCHOR\n\nProbably American.\n\nBRUCE\n\nNow there is a father looking for his little boy.\n\nMARLIN\n\nUgh! What do these markings mean?\n\nBRUCE\n\nI never knew my father! [sobs]\n\nCHUM\n\nAw, come here.\n\nANCHOR\n\nGroup hug.\n\nCHUM\n\nWe're all mates here, mate.\n\nMARLIN\n\nI can't read human.\n\nDORY\n\nWell then we gotta find a fish who can read this. Hey, look. Sharks!\n\nMARLIN\n\nNo, no, no, Dory!\n\nDORY\n\nGuys, guys!\n\nMARLIN\n\nNo, Dory!\n\nDORY\n\nThat's mine! Give it to me! Gimme! Oww!\n\nMARLIN\n\nOh, I'm sorry. Are you okay?\n\nDORY\n\nOw, ow, ow.\n\n 14\n\nMARLIN\n\nI'm so sorry.\n\nDORY\n\nYou really clocked me there. Am I bleeding?\n\nMARLIN\n\nOhh...\n\nDORY\n\nOw, ow, ow.\n\nBRUCE\n\nDory, are you oka--oohh. Oohh, that's good.\n\nANCHOR/CHUM\n\nIntervention!\n\nBRUCE\n\nJust a bite!\n\nANCHOR\n\nHold it together, mate!\n\nCHUM\n\nRemember, Bruce, fish are friends, not food!\n\nBRUCE\n\nFOOD!\n\nMARLIN\n\nDory, look out!\n\nBRUCE\n\nI'm havin' fish tonight!\n\nCHUM\n\nRemember the steps, mate!\n\nBRUCE\n\nJust one bite!\n\nBRUCE\n\nG'day!\n\nMARLIN/DORY\n\nAaaaaaaah!\n\nBRUCE\n\nArrrr!\n\nMARLIN\n\nThere's no way out! There's got to be a way to escape!\n\nDORY\n\nWho is it?\n\nMARLIN\n\nDory, help me find a way out!\n\nDORY\n\nSorry, you'll have to come back later. We're trying to escape.\n\nMARLIN\n\nThere's gotta be a way out!\n\nDORY\n\nLook, here's something! 'ESSS-CA-PE'! I wonder what that means. It's funny, it's spelled\n\njust like the word 'escape'.\n\nMARLIN\n\nLet's go!\n\nBRUCE\n\nHere's Brucey!\n\nMARLIN\n\n 15\n\nWait a minute..you can read?!\n\nDORY\n\nI can read? That's right, I can read!\n\nMARLIN\n\nWell, then here. Read this now!\n\nANCHOR\n\nHe really doesn't mean it, y'know! He never even knew his father!\n\nCHUM\n\nDon't fall off the wagon!\n\nMARLIN\n\nOh no, it's blocked!\n\nANCHOR\n\nNo, Bruce. Focus!\n\nCHUM\n\nSorry about--this, mate!\n\nANCHOR\n\nHe's really--a nice guy!\n\nMARLIN\n\nI need to get that mask!\n\nDORY\n\nYou want that mask? Okay.\n\nMARLIN\n\nNo, no, no, no, no, no!\n\nMARLIN\n\nQuick grab the mask!\n\nANCHOR\n\nOh no. Bruce?\n\nBRUCE\n\nWhat? [gasps] Swim away! Swim away!\n\nDORY\n\nAw, is the party over?\n\nPELICAN\n\nNice.\n\n======================================================================================\n\nNEMO\n\nDad? Daddy?\n\nDENTIST\n\nBarbara?\n\nBARBARA\n\nUh-huh?\n\nDENTIST\n\nPrep for his anterior crown, would you, please? And I'm going to need a few cotton rolls.\n\nBARBARA\n\nOkay.\n\nDENTIST\n\nHello, little fella!\n\nNEMO\n\nAah!\n\nDENTIST\n\nHeh heh heh! Beauty, isn't he? I found that guy struggling for life out on the reef and\n\nI saved him. So, has that novocaine kicked in yet?\n\n 16\n\nPATIENT\n\nI think so. We're ready to roll.\n\nBUBBLES\n\nBubbles! [muttering] My bubbles.\n\nPEACH\n\nHe likes bubbles.\n\nNEMO\n\nAah! Ohh! No! Uhh!\n\nJACQUES\n\nBonjour.\n\nNEMO\n\nAah!\n\nBLOAT\n\nHeh heh! Slow down, little fella. There's nothing to worry about.\n\nDEB\n\nOh, he's scared to death.\n\nNEMO\n\nI wanna go home. Do you know where my dad is?\n\nPEACH\n\nHoney, your dad's probably back at the pet store.\n\nNEMO\n\nPet store?\n\nBLOAT\n\nYeah, you know, like I'm from Bob's Fish Mart.\n\nGURGLE\n\nPet Palace.\n\nBUBBLES\n\nFish-O-Rama.\n\nDEB\n\nMail order.\n\nPEACH\n\nEbay.\n\nGURGLE\n\nSo which one is it?\n\nNEMO\n\nI'm from the ocean.\n\nGURGLE\n\nAh, the ocean. The ocean! Aaah! He hasn't been decontaminated yet! Jacques!\n\nJACQUES\n\nOui.\n\nGURGLE\n\nClean him!\n\nJACQUES\n\nOui.\n\nGURGLE\n\nOcean!\n\nJACQUES\n\nOoh, la mer. Bon. Voila. He is clean.\n\nBUBBLES\n\nWow. The big blue. What's it like?\n\nNEMO\n\nBig...and blue?\n\n 17\n\nBUBBLES\n\nI knew it.\n\nDEB\n\nKid, if there's anything you need, just ask your auntie Deb, that's me. Or if I'm not\n\naround, you can always talk to my sister Flo. Hi,how are you? Don't listen to anything\n\nmy sister says, she's nuts! Ha ha ha ha!\n\nPEACH\n\n[muffled] We got a live one!\n\nBLOAT\n\nCan't hear you, Peach.\n\nPEACH\n\nI said we got a live one.\n\nGURGLE\n\nYes!\n\nBLOAT\n\nOh boy, oh boy, oh boy, oh boy!\n\nDEB\n\nWhat do we got?\n\nPEACH\n\nRoot canal, and by the looks of those x-rays it's not gonna be pretty.\n\nPATIENT\n\nOwwwwwwwww!\n\nBLOAT\n\nRubber dam and clamp installed?\n\nPEACH\n\nYep.\n\nGURGLE\n\nWhat did he use to open?\n\nPEACH\n\nGator-Glidden drill. He seems to be favoring that one lately.\n\nDEB\n\nI can't see, Flo.\n\nPATIENT\n\nYou're getting a little too--aaaaah!!!\n\nPEACH\n\nNow he's doing the Schilder technique.\n\nBLOAT\n\nOooh, he's using a Hedstrom file.\n\nGURGLE\n\nThat's not a Hedstrom file. That's a K-Flex.\n\nBLOAT\n\nIt's got a teardrop cross-section. Clearly a Hedstrom.\n\nGURGLE\n\nNo, no. K-Flex.\n\nBLOAT\n\nHedstrom!\n\nGURGLE\n\nK-Flex!\n\nBLOAT\n\nHedstro--! [inflates] There I go. A little help over here.\n\nDEB\n\nI'll go deflate him.\n\n 18\n\nDENTIST\n\nAll right, go ahead and rinse.\n\nGURGLE\n\nUgh! The human mouth is a disgusting place.\n\nPEACH\n\nHey, Nigel.\n\nNIGEL\n\nWhat did I miss? Am I late?\n\nPEACH\n\nRoot canal and it's a doozy.\n\nNIGEL\n\nRoot canal, eh? What did he use to open?\n\nPEACH\n\nGator-Glidden drill.\n\nNIGEL\n\nHe seems to be favoring that one. Hope he doesn't get surplus sealer at the portal terminus...\n\nhello.\n\nNEMO\n\n[gasps]\n\nNIGEL\n\nWho's this?\n\nDEB\n\nNew guy. Ha ha ha!\n\nGURGLE\n\nThe dentist took him off the reef.\n\nNIGEL\n\nAn outie. From my neck of the woods, eh? Sorry if I ever took a snap at you. Fish gotta swim,\n\nbirds gotta eat. [gasps]\n\nDENTIST\n\nHey! No, no, no, no! They're not your fish. They're my fish. Come on, go! Go on, shoo! Oh,\n\nthe picture broke. This here's Darla. She's my niece. She's going to be eight next week.\n\nHey, little fella. Say hello to your new mummy. She'll be here Friday to pick you up. You're\n\nher present. Shh, shh, shh! It's our little secret. Well, Mr. Tucker, while that sets up\n\nI'm going to see a man about a wallaby.\n\nBLOAT\n\nOh, Darla.\n\nNEMO\n\nWhat? What's wrong with her?\n\nGURGLE\n\nShe wouldn't stop shaking the bag.\n\nBUBBLES\n\nPoor Chuckles.\n\nDEB\n\nHe was her present last year.\n\nBLOAT\n\nHitched a ride on the porcelain express.\n\nPEACH\n\nShe's a fish killer.\n\nNEMO\n\nI can't go with that girl! I have to get back to my dad! Aaah! Daddy! Help me!\n\nGURGLE\n\nOh, he's stuck!\n\nGILL\n\n 19\n\nNobody touch him! Nobody touch him.\n\nNEMO\n\nCan you help me?\n\nGILL\n\nNo. You got yourself in there, you can get yourself out.\n\nPEACH\n\nGill..\n\nGILL\n\nI just wanna see him do it, okay? Calm down. Alternate wiggling your fins and your tail.\n\nNEMO\n\nI can't. I have a bad fin.\n\nGILL\n\nNever stopped me.\n\nGILL\n\nJust think about what you need to do.\n\nBLOAT\n\nCome on.\n\nGILL\n\nPerfect.\n\nBUBBLES\n\nYay!\n\nGURGLE\n\nYou did it!\n\nDEB\n\nGood squirming! Ha ha ha!\n\nPEACH\n\nWow. From the ocean. Just like you, Gill.\n\nGILL\n\nYeah.\n\nPEACH\n\nI've seen that look before. What are you thinking about?\n\nGILL\n\nI'm thinking, tonight, we give the kid a proper reception.\n\nBLOAT\n\nSo kid, you got a name or what?\n\nNEMO\n\nNemo. I'm Nemo.\n\n======================================================================================\n\nMARLIN\n\nNemo. Nemo. [mutters]\n\nDORY\n\nAre you gonna eat that? Careful with that hammer...\n\nMARLIN\n\nHuh? No, no! What does it say? Dory!\n\nDORY\n\nSea monkey has my money...\n\nMARLIN\n\nWake up! Get up! Come on! Come on!\n\nDORY\n\nYes, I'm a natural blue...\n\nMARLIN\n\n 20\n\nGet up!\n\nDORY\n\nLook out! Sharks eat fish! Aaaaaah!\n\nMARLIN/DORY\n\nAAAAAAAAAAHHH!!!\n\nDORY\n\nWow. Dusty.\n\nMARLIN\n\n[gasps] The mask! Where's the mask? No! No, not the mask! Get it! Get the mask!\n\nGet the mask! Get it!\n\nDORY\n\n[singing] Hoo doot doo doot doot doo doot. Whoo-hoo! La la la la la la. Just keeps\n\ngoing on, doesn't it? Echo! Echo! Hey, what are you doing?\n\nMARLIN\n\nIt's gone. I've lost the mask.\n\nDORY\n\nDid you drop it?\n\nMARLIN\n\nYou dropped it! That was my only chance of finding my son, now it's gone.\n\nDORY\n\nHey, Mr. Grumpy Gills. When life gets you down, you know what you gotta do?\n\nMARLIN\n\nI don't wanna know what you gotta do when life gets you down.\n\nDORY\n\n[singing] Just keep swimming. Just keep swimming, swimming, swimming. What do we do?\n\nWe swim, swim.\n\nMARLIN\n\nDory, no singing.\n\nDORY\n\n[singing] Ho ho ho ho ho ho! I love to swim! When you want to swim..\n\nMARLIN\n\nSee, I'm going to get stuck now with that song now it's in my head!\n\nDORY\n\nSorry.\n\nMARLIN\n\nDory, do you see anything?\n\nDORY\n\nAaah! Something's got me!\n\nMARLIN\n\nThat was me. I'm sorry.\n\nDORY\n\n[gasps] Who was that?\n\nMARLIN\n\nWho could it be? It's me!\n\nDORY\n\nAre..are you my conscience?\n\nMARLIN\n\nYeah, yeah. I'm your conscience. We haven't spoken for a while. How are you?\n\nDORY\n\nHmm, can't complain.\n\nMARLIN\n\nYeah? Good. Now, Dory. I want you to tell me..do you see anything?\n\n 21\n\nDORY\n\nI see..I see a light.\n\nMARLIN\n\nA light.\n\nDORY\n\nYeah. Over there. Hey, conscience. Am I dead?\n\nMARLIN\n\nNo, I see it too. What is it?\n\nDORY\n\nIt's so pretty.\n\nMARLIN\n\nI'm feeling...happy. Which is a big deal for me.\n\nDORY\n\nI want to touch it. Oh!\n\nMARLIN\n\nHey, come back. Come on back here.\n\nDORY\n\n[singing] I'm gonna get you. I'm gonna get you. I'm gonna swim with you.\n\nMARLIN\n\nI'm gonna get you. I'm gonna be your best friend...good feeling's gone.\n\nMARLIN\n\nI can't see! I don't know where I'm going!\n\nDORY\n\nHaah!\n\nMARLIN\n\nThe mask!\n\nDORY\n\nWhat mask?\n\nDORY\n\nOkay, I can't see a thing.\n\nMARLIN\n\nOh, gee!\n\nDORY\n\nHey, look! A mask!\n\nMARLIN\n\nRead it!\n\nDORY\n\nI'm sorry, but if you could just bring it a little closer, I kind of need the light.\n\nThat's great, keep it right there.\n\nMARLIN\n\nJust read it!\n\nDORY\n\nOkay, okay. Mr. Bossy. Uh, 'P'. Okay, 'P'. 'Shh-eer...Sher--P. Sher--P. Shirley? P.--'. Oh!\n\nThe first line's 'P. Sherman'!\n\nMARLIN\n\nP. Sherman doesn't make any sense!\n\nDORY\n\nOkay, second line. '42'.\n\nMARLIN\n\nDon't eat me! Don't eat me! Aaaah!\n\nDORY\n\nLight, please. 'Walla--Walla--Walla-beee'...\n\n 22\n\nMARLIN\n\nWaah! Waaah! Waaaah!\n\nDORY\n\nThe second line's '42 Wallaby Way'!\n\nMARLIN\n\nThat's great! Speed read! Take a guess! No pressure! No problem! There's a lot of pressure!\n\nPressure! Take a guess now with pressure!\n\nDORY\n\n'Sydney'. It's 'Sydney'!\n\nMARLIN\n\nDuck!\n\nDORY\n\nAaah!\n\nMARLIN\n\nI'm dead, I'm dead, I'm dead, I'm dead, I'm dead, I died, I'm dead.\n\nMARLIN\n\nWhoo-hoo! [singing] We did it, we did it! Oh yeah, yeah, yeah! No eating here tonight, whoo!\n\nBOTH\n\n[singing] Eating here tonight!\n\nMARLIN\n\nDory.\n\nDORY\n\n[singing] No, no, no eating here tonight. You on a diet--\n\nMARLIN\n\nDory! What did the mask say?\n\nDORY\n\n'P. Sherman, 42 Wallaby Way, Sydney'. [gasps] I remember what it said! I usually forget\n\nthings, but I remembered it this time!\n\nMARLIN\n\nWhoa, whoa, wait! Where is that?\n\nDORY\n\nI don't know. But who cares? I remembered!\n\nMARLIN/DORY\n\nAaah!\n\nDORY\n\nP. Sherman, 42 Wallaby Way, Sydney. I remembered it again!\n\n======================================================================================\n\nJACQUES\n\nPsst. Nemo.\n\nNEMO\n\nMmmm...\n\nJACQUES\n\nNemo.\n\nNEMO\n\nHuh?\n\nJACQUES\n\nSuivez-moi. Follow me.\n\nBLOAT/BUBBLES/GURGLE\n\n[chanting] Ha! Ho! Hwa! Hwee! Ha! Ho! Ho! Ho! Ha! Ho! Hwa! Hwee! Ha! Ho! Ho! Ho! Ha! Ho!\n\nHwa! Hwee! Ha! Ho! Ho! Ho! Hahoo! Wahoo! Yahoo! Ho! Ha! Ho! Wahee! Ha! Ho! Ho! Ho! Hoo!\n\nGILL\n\nState your name.\n\n 23\n\nNEMO\n\nNemo.\n\nGILL\n\nBrother Bloat, proceed.\n\nBLOAT\n\nNemo! Newcomer of orange and white, you have been called forth to the summit of Mount\n\nWannahockaloogie to join with us in the fraternal bonds of tankhood.\n\nNEMO\n\nHuh?\n\nPEACH\n\nWe want you in our club, kid.\n\nNEMO\n\nReally?\n\nBLOAT\n\nIf you are able to swim through..THE RING OF FIRE! [whispers to Jacques] Turn on the\n\nRing of Fire! The Ring of Fire, you said you could do it--THE RING OF FIRE!\n\nBUBBLES\n\nBubbles! Bubbles! Let me--oww!\n\nBLOAT/BUBBLES/GURGLE\n\n[chanting]\n\nPEACH\n\nIsn't there another way? He's just a boy!\n\nJACQUES\n\n[wailing]\n\nGILL\n\nFrom this moment on, you will now be known as Sharkbait.\n\nBLOAT/BUBBLES/GURGLE\n\nSharkbait! Ooh ha ha!\n\nGILL\n\nWelcome, brother Sharkbait!\n\nBLOAT/BUBBLES/GURGLE\n\nSharkbait! Ooh ha ha!\n\nGILL\n\nEnough with the Sharkbait.\n\nGURGLE\n\nSharkbait! Ooh..ba-ba-doo.\n\nGILL\n\nOkay, Sharkbait's one of us now, agreed?\n\nBLOAT/BUBBLES/GURGLE\n\nAgreed!\n\nGILL\n\nWe can't send him off to his death. Darla's coming in 5 days, so what are we gonna do?\n\nI'll tell you what we're gonna do: we're gonna get him outta here. We're gonna help\n\nhim escape.\n\nNEMO\n\nEscape? Really?\n\nGILL\n\nWe're all gonna escape!\n\nGURGLE\n\nGill, please, not another one of your escape plans.\n\nDEB\n\nSorry, but they, they just, they never work.\n\n 24\n\nBLOAT\n\nYeah. Why should this be any different?\n\nGILL\n\n'Cause we've got him.\n\nNEMO\n\nMe?\n\nGILL\n\nYou see that filter?\n\nNEMO\n\nYeah?\n\nGILL\n\nYou're the only one who can get in and out of that thing. What we need you to do is take\n\na pebble inside and jam the gears. You do that and this tank's gonna get filthier and\n\nfilthier by the minute. Pretty soon, the dentist'll have to clean the tank himself. And\n\nwhen he does, he'll take us out of the tank, put us in the individual baggies, then we roll\n\nourselves down the counter, out of the window, off the awning, into the bushes, across the\n\nstreet and into the harbor! It's foolproof! Who's with me?\n\nBLOAT\n\nAye!\n\nJACQUES\n\nAye!\n\nDEB\n\nAye!\n\nBUBBLES\n\nAye!\n\nGURGLE\n\nI think your nuts.\n\nGILL/NEMO\n\n[sighs]\n\nGURGLE\n\nNo offense, kid, but, um..you're not the best swimmer.\n\nGILL\n\nHe's fine, he can do this. So Sharkbait, what do you think?\n\nNEMO\n\nLet's do it.\n\n======================================================================================\n\nDORY\n\nI'm going to P. Sherman, 42 Wallaby Way, Sydney. Where are you going? I'm going to P.\n\nSherman, 42 Wallaby Way, Sydney. If you're askin' where I'm goin'. I'll tell you that's\n\nwhere I'm going. It's P. Sherman, 42 Wallaby Way, Sydney. Where? I'm sorry, I didn't hear\n\nyou. P. Sherman, 42 Wallaby Way...\n\nMARLIN\n\nExcuse me. Ex-excuse me, um, hi. Do you know how to get to--hello? W-w-w-wait! Can you\n\ntell me--hey! Hold it! Wait a minute! I'm trying to talk to you. Okay, fellas, come back\n\nhere. Please, one quick question. I need to aaaaand they're gone again. [sighs]\n\nDORY\n\nP. Sherman 42 Wallaby Way, Sydney. Why do I have to tell you over and over again? I'll tell\n\nyou again. I don't get tired of it--\n\nMARLIN\n\nOkay, all right.\n\nDORY\n\nHuh?\n\nMARLIN\n\nHere's the thing.\n\nDORY\n\n 25\n\nUh-huh.\n\nMARLIN\n\nY'know, I just, I-I think it's best if I just, if I just, carry on from here by..by myself.\n\nDORY\n\nOkay.\n\nMARLIN\n\nY'know, alone.\n\nDORY\n\nUh-huh.\n\nMARLIN\n\nWithout, without..well, I mean, not without you. I mean, it's just that I don't want you...\n\nwith me.\n\nDORY\n\nHuh?\n\nMARLIN\n\nI don't wanna hurt your feelings..\n\nDORY\n\nYou want me to leave?\n\nMARLIN\n\nWell, I mean not..yes, yeah. It's just that you know I-I just can't afford anymore delays\n\nand you're one of those fish that cause delays. And sometimes it's a good thing. There's\n\na whole group of fish. They're..'delay fish'.\n\nDORY\n\nYou mean..[whimper]you mean you don't..like me? [sobs]\n\nMARLIN\n\nNo, of course I like you. It's because I like you I don't wanna be with you. It's a\n\ncomplicated emotion. Oh, don't cry. I like you.\n\nMOONFISH LEADER\n\nHey, you! Lady, is this guy botherin' you?\n\nDORY\n\nUm, I don't remember. Were you?\n\nMARLIN\n\nNo, no, no, no, no. We're just, we're..hey, do you guys know how I can get to--\n\nMOONFISH LEADER\n\nLook, pal. We're talkin' to the lady, not you. Hey-hey, you like impressions?\n\nDORY\n\nMm-mmm-mmmm.\n\nMOONFISH LEADER\n\nOkay. Just like in rehearsals, gentlemen. So, what are we? Take a guess.\n\nDORY\n\nOh, oh, I've seen one of those.\n\nMOONFISH LEADER\n\nI'm a fish with a nose like a sword.\n\nDORY\n\nWait, wait, um..\n\nMARLIN\n\nIt's a swordfish.\n\nMOONFISH LEADER\n\nHey, clown boy! Let the lady guess. Where's the butter?\n\nDORY\n\nOh-oh-oh! It's on the tip of my tongue.\n\nMARLIN\n\n[coughs up answer]Lobster.\n\n 26\n\nMOONFISH LEADER\n\nSaw that.\n\nMARLIN\n\nWhat?\n\nMOONFISH LEADER\n\nLots of legs, lives in the ocean.\n\nDORY\n\nClam!\n\nMOONFISH LEADER\n\nClose enough. [singing] Oh, it's a whale of a tale, I'll tell you lad, a whale of a tale.\n\nDORY\n\nOh, they're good.\n\nMARLIN\n\nWill somebody please give me directions?\n\nMOONFISH LEADER\n\n[impersonating Marlin] Will somebody please give me directions?\n\nDORY\n\nHa ha ha ha ha!\n\nMARLIN\n\nI'm serious.\n\nMOONFISH LEADER\n\nBlah-blah-blah! Me-me-blah! Blah-blah-blah-blah-me-me-me!\n\nMARLIN\n\nThank you.\n\nDORY\n\nOh dear. Hey, hey come back! Hey, what's the matter?\n\nMARLIN\n\nWhat's the matter? While they're doing their silly little impressions, I am miles from\n\nhome, with a fish that can't even remember her own name.\n\nDORY\n\nBoy, bet that's frustrating.\n\nMARLIN\n\nYeah. Meanwhile my son is out there.\n\nDORY\n\nYou're son Chico?\n\nMARLIN\n\nNemo.\n\nDORY\n\nRight. Got it.\n\nMARLIN\n\nBut it doesn't matter, 'cause no fish in this entire ocean is gonna help me.\n\nDORY\n\nWell, I'm helping you. Wait right here. Hey, guys.\n\nMOONFISH LEADER\n\nWhat, is he bothering you again?\n\nDORY\n\nNo, no, he's a good guy. Go easy on him, he's lost his son, Fabio. Any of you heard of\n\nP. Sherman, 42 Wallaby Way, Sydney?\n\nMOONFISH LEADER\n\nSydney? Oh sure. Why, Ted here's got relatives in Sydney. Don't you, Ted?\n\nMOONFISH TED\n\nSure do.\n\n 27\n\nDORY\n\nOh, hey! They know Sydney!\n\nMARLIN\n\n[gasps]\n\nDORY\n\nYou wouldn't know how to get there, would you?\n\nMOONFISH LEADER\n\nWhat you wanna do is follow the EAC, that's the East Australian Current. Big current,\n\ncan't miss it, it's in..that direction. And then you gotta follow that for about, I\n\ndon't know, what do you guys think? About three leagues? And that little baby's gonna\n\nput you right past Sydney.\n\nMOONFISH SCHOOL\n\nTA-DAA!\n\nMARLIN\n\nGreat! That's great! Dory, you did it!\n\nDORY\n\nOh, please. I'm just your little helper. Helping along, that's me.\n\nMARLIN\n\nWell, listen fellas, thank you.\n\nMOONFISH LEADER\n\nDon't mention it. And, uh, loosen up. Okay, buddy?\n\nDORY\n\nOh, you guys. You really nailed him. Bye.\n\nMOONFISH LEADER\n\nOh, hey ma'am, one more thing.\n\nDORY\n\nYes.\n\nMOONFISH LEADER\n\nWhen you come to this trench, swim through it, not over it.\n\nDORY\n\nTrench, through it, not over it. I'll remember. Hey, hey! Hey! Hey! Hey, wait up, partner.\n\nHold on. Wait! Wait-wait! I got, I gotta tell you something..whoa. Nice trench. Hello!\n\nOkay, let's go.\n\nMARLIN\n\nBad trench, bad trench. Come on, we're gonna swim over this thing.\n\nDORY\n\nWhoa, whoa, partner. Little red flag goin' up. Somethin's telling me we should swim through\n\nit, not over it.\n\nMARLIN\n\nAre you even looking at this thing? It's got death written all over it.\n\nDORY\n\nI'm sorry, but I really, really, really think we should swim through.\n\nMARLIN\n\nAnd I'm really, really done talking about this. Over we go.\n\nDORY\n\nCome on, trust me on this.\n\nMARLIN\n\nTrust you?\n\nDORY\n\nYes, trust. It's what friends do.\n\nMARLIN\n\nLook! Something shiny!\n\nDORY\n\n 28\n\nWhere?\n\nMARLIN\n\nOh, it just swam over the trench. Come on, we'll follow it.\n\nDORY\n\nOkay.\n\nDORY\n\nBoy, sure is clear up here.\n\nMARLIN\n\nExactly. And look at that, there's the current. We should be there in no time.\n\nDORY\n\nHey, little guy.\n\nMARLIN\n\nYou wanted to go through the trench.\n\nDORY\n\nI shall call him Squishy and he shall be mine and he shall be my Squishy. Come here,\n\nSquishy. Come here, little Squishy. [Baby talk]---oww!\n\nMARLIN\n\nDory! That's a jellyfish!\n\nDORY\n\nBad Squishy! Bad Squishy!\n\nMARLIN\n\nShoo! Shoo, shoo! Get away! Come here, let me see.\n\nDORY\n\nDon't touch it! Don't touch it!\n\nMARLIN\n\nI'm not gonna touch it. I just wanna look.\n\nDORY\n\nHeeey, how come it didn't sting you?\n\nMARLIN\n\nIt did. It's just that..\n\nDORY\n\nOw! Ow, oww!\n\nMARLIN\n\n..hold still. I live in this anemone and I'm, I'm, I'm used to these kind of stings.\n\nCome here.\n\nDORY\n\nOw, ow! Oww!\n\nMARLIN\n\nIt doesn't look bad, you're gonna be fine. But now we know, don't we?\n\nDORY\n\nYeah.\n\nMARLIN\n\nThat we don't wanna touch these again. Let's be thankful this time it was just a\n\nlittle one.[gasps]\n\nMARLIN/DORY\n\nAaaah!\n\nMARLIN\n\nDon't move! This is bad, Dory.\n\nDORY\n\nHey, watch this! Boing! Boing!\n\nMARLIN\n\n[gasps] Dory!\n\n 29\n\nDORY\n\nBoing-boing-boing! [singing] You can't catch me!\n\nMARLIN\n\nDory! Don't bounce on the tops! They will..not sting you. The tops don't sting you,\n\nthat's it!\n\nDORY\n\nOoh! Two in a row, beat that.\n\nMARLIN\n\nDory! All right, listen to me. I have an idea, a game.\n\nDORY\n\nA game?\n\nMARLIN\n\nA game.\n\nDORY\n\nA game?\n\nMARLIN\n\nYes.\n\nDORY\n\nAah! I love games! Pick me!\n\nMARLIN\n\nAll right, here's the game. Um, whoever can hop the fastest out of these jellyfish, wins.\n\nDORY\n\nOkay!\n\nMARLIN\n\nRules, rules, rules!\n\nDORY\n\nOkay!\n\nMARLIN\n\nYou can't touch the tentacles, only the tops.\n\nDORY\n\nSomething about tentacles, got it. On your mark, get set, go!\n\nMARLIN\n\nW-wait! Wait! Not something about them, it's all about them! Wait!\n\nDORY\n\nWeeee!\n\nMARLIN\n\nDory!\n\nDORY\n\nGotta go faster if you wanna win!\n\nMARLIN\n\n[gasps] Dory!\n\nDORY\n\nBoing! Boing! Boing-boing-boing-boing!\n\nMARLIN\n\nWait a minute--whoa! Dory!\n\nDORY\n\nWeeee!\n\nMARLIN\n\nSo, we're cheating death now. That's what we're doin'. We're havin' fun at the same time.\n\nI can do this, just be careful.\n\nDORY\n\nYeah, careful I don't make you cry when I win!\n\n 30\n\nMARLIN\n\nOh, I don't think so!\n\nDORY\n\nHa ha ha ha! Whooo! Give it up, old man. You can't fight evolution, I was built for speed.\n\nMARLIN\n\nThe question is, Dory, are you hungry?\n\nDORY\n\nHuh? Hungry?\n\nMARLIN\n\nYeah, 'cause you're about to eat my bubbles! Duck to the left! Right there! The clownfish\n\nis the winner! Woohoo! We did it! We're gonna...Dory? Oh no. Dory! Dory! Dory! [gasps]\n\nDory! Uggghhh!\n\nDORY\n\nUgh...am I disqualified?\n\nMARLIN\n\nNo, you're doing fine! You're, you're actually winning! But you gotta stay awake. Uh, where\n\ndoes P. Sherman live?\n\nDORY\n\nP..Sherman..Wallaby Way...Sydney...\n\nMARLIN\n\nThat's it! Oww! Ow! Stay awake! Stay awake! Ow! Stay awake! Stay--awake!\n\nDORY\n\nAwake...P..Sherman..\n\nMARLIN\n\nAwake...\n\nDORY\n\n..42 Wallaby Way...\n\nMARLIN\n\nAwake...wake up...Nemo...\n\n======================================================================================\n\nGILL\n\nYou miss your dad, don't you, Sharkbait?\n\nNEMO\n\nYeah.\n\nGILL\n\nWell, you're lucky to have someone out there who's lookin' for you.\n\nNEMO\n\nHe's not looking for me. He's scared of the ocean.\n\nGILL\n\nPeach, any movement?\n\nPEACH\n\nHe's had at least four cups of coffee, it's gotta be soon.\n\nGILL\n\nKeep on him.\n\nGILL\n\nMy first escape, landed on dental tools. I was aimin' for the toilet.\n\nNEMO\n\nToilet?\n\nGILL\n\nAll drains lead to the ocean, kid.\n\nNEMO\n\nWow. How many times have you tried to get out?\n\n 31\n\nGILL\n\nAah, I've lost count. Fish aren't meant to be in a box, kid. It does things to 'ya.\n\nBUBBLES\n\nBubbles! Bubbles, bubbles, bubbles---\n\nPEACH\n\nPotty break! Potty break! He just grabbed the Reader's Digest! We have 4.2 minutes.\n\nGILL\n\nThat's your cue, Sharkbait.\n\nBLOAT\n\nYou can do it, kid.\n\nGILL\n\nOkay, you gotta be quick. Once you get in, you swim down to the bottom of the chamber\n\nand I'll talk you through the rest.\n\nNEMO\n\nOkay.\n\nGILL\n\nGo on, it'll be a piece of kelp.\n\nNEMO\n\n[takes a deep breath]\n\nGILL\n\nNicely done! Can you hear me?\n\nNEMO\n\nYeah.\n\nGILL\n\nHere comes the pebble. Now, do you see a small opening?\n\nNEMO\n\nUh-huh.\n\nGILL\n\nOkay, inside it you'll see a rotating fan. Very carefully, wedge that pebble into the\n\nfan to stop it turning.\n\nNEMO\n\nAaah!\n\nGILL\n\nCareful, Sharkbait.\n\nNEMO\n\nI can't do it!\n\nPEACH\n\nGill, this isn't a good idea.\n\nGILL\n\nHe'll be fine. Try again.\n\nNEMO\n\nOkay.\n\nGILL\n\nThat's it, Sharkbait. Nice and steady.\n\nNEMO\n\nI got it! I got it!\n\nPEACH\n\n[sigh]\n\nBLOAT\n\nHe did it!\n\nGURGLE\n\nWhew!\n\n 32\n\nGILL\n\nThat's great, kid! Now, swim up the tube and out.\n\nNEMO\n\nOh no! Gill! Gill!\n\nGILL\n\nSharkbait!\n\nBLOAT\n\nOh my gosh!\n\nGILL\n\nGet 'im outta there! Get 'im outta there!\n\nBUBBLES\n\nHelp him!\n\nGURGLE\n\nWhat do we do!? What do we do!?\n\nPEACH\n\nOh no!\n\nGILL\n\nStay calm, kid! Just don't panic!\n\nNEMO\n\nHelp me!\n\nGILL\n\nSharkbait! Grab hold of this!\n\nNEMO\n\nNo! No!\n\nGILL\n\nFeed me more!\n\nGURGLE\n\nThat's it!\n\nGILL\n\nCome on, Sharkbait! Grab it!\n\nNEMO\n\nI got it!\n\nGILL\n\nPull!\n\nPEACH\n\nGill, don't make him go back in there.\n\nGILL\n\nNo. We're done.\n\n======================================================================================\n\nCRUSH\n\nDude.\n\nMARLIN\n\nOoh...\n\nCRUSH\n\nDude. Focus, dude. Dude.\n\nMARLIN\n\nOoooh...\n\nCRUSH\n\nOh, he lives! Hey, dude!\n\nMARLIN\n\nOoooh..what happened?\n\n 33\n\nCRUSH\n\nOh, saw the whole thing, dude. First you were like, 'whoa'! And then we were all like,\n\n'whoa'! And then you were like, 'whoa'.\n\nMARLIN\n\nWhat're you talking about?\n\nCRUSH\n\nYou, mini-man. Takin' on the jellies. You got serious thrill issues, dude.\n\nMARLIN\n\nOoh.\n\nCRUSH\n\nAwesome.\n\nMARLIN\n\nOoh..ooh, my stomach. Ooooh..\n\nCRUSH\n\nOh, man. No hurlin' on the shell, dude, okay, just waxed it.\n\nMARLIN\n\nSo Mr. Turtle...\n\nCRUSH\n\nWhoa, dude. Mr. Turtle is my father. Name's Crush.\n\nMARLIN\n\nCrush? Really? Okay Crush, listen I need to get to the East Australian Current. EAC?\n\nCRUSH\n\nHa ha ha, dude, ha ha, you're ridin' it, dude! Check it out!\n\nCRUSH\n\nOkay, grab shell, dude!\n\nMARLIN\n\nGrabbing--waaaaaaaaaaaaaaaaah!!! Aaaaaaaaaaaah!!! Aaaaaaaaaaaah!!! Whooooooaaaa!!!\n\nCRUSH\n\nHa ha! Righteous! Righteous! Yeah!\n\nMARLIN\n\nStop!\n\nCRUSH\n\nSo, what brings you on this fine day to the EAC?\n\nMARLIN\n\nWell, Dory and I need to get to Sydney. [gasps] Dory! Dory! Is she all right!?\n\nCRUSH\n\nOh. Oh, Little Blue. She is sub-level, dude.\n\nMARLIN\n\nDory, Dory! Dory!\n\nDORY\n\nHmm-mmm....\n\nMARLIN\n\nOh, Dory. I-I-I'm so sorry. This is all my fault, it's my fault...\n\nDORY\n\n..29, 30! Ready or not, here I come! There you are! Catch me if you can! Ha ha!\n\nHa ha ha ha!\n\nMARLIN\n\nHuh?\n\nSQUIRT\n\nWhoa!\n\nMARLIN\n\n[gasps] Oh my goodnes!\n\n 34\n\nCRUSH\n\nWhoa. Kill the motor, dude. Let us see what Squirt does flying solo.\n\nSQUIRT\n\nWhoa! Whoa! That was so cool! Hey dad, did you see that? Did you see me? Did you see\n\nwhat I did?\n\nCRUSH\n\nYou so totally rock, Squirt! So give me some fin..noggin..\n\nCRUSH/SQUIRT\n\n..dude!\n\nCRUSH\n\nOh, intro. Jellyman, Offspring. Offspring, Jellyman.\n\nSQUIRT\n\nJellies? Sweet.\n\nCRUSH\n\nTotally.\n\nMARLIN\n\nWell, apparently, I must've done something you all like. Heh, uh, dudes.\n\nSQUIRT\n\nYou rock, dude.\n\nMARLIN\n\nOw.\n\nCRUSH\n\nCurl away, my son. Aw, it's awesome, Jellyman. Little dudes are just eggs, leave 'em\n\non the beach to hatch, then coo-coo-ca-choo, they find their way back to the big 'ol blue.\n\nMARLIN\n\nAll by themselves?\n\nCRUSH\n\nYeah.\n\nMARLIN\n\nBut-but-but dude, how do you know when they're ready?\n\nCRUSH\n\nWell, you never really know. But when they'll know, you'll know, you know? Ha.\n\nDORY\n\nHey! Look, everybody!\n\nSQUIRT\n\nI know that dude. It's the Jellyman.\n\nDORY\n\nWell, go on, jump on him.\n\nTURTLE KIDS\n\nTurtle pile!\n\nMARLIN\n\nW-w-wai-wait--\n\nTURTLE KID 1\n\nAre you funny?\n\nTURTLE KID 2\n\nWhere's your shell?\n\nMARLIN\n\nHold on, I need to breath--\n\nTURTLE KID 3\n\nAre you running away?\n\nTURTLE KID 4\n\nDid you really cross the jellyfish forest?\n\n 35\n\nTURTLE KID 5\n\nDid they sting you?\n\nMARLIN\n\nOne at a time!\n\nTURTLE KID 6\n\nMr. Fish, did you die?\n\nDORY\n\nSorry. I was a little vague on the details.\n\nSQUIRT\n\nSo where are you going?\n\nMARLIN\n\nWell, you see my son was taken. My son was taken away from me.\n\nTURTLE KIDS\n\n[gasp]\n\nDORY\n\nNo way.\n\nSQUIRT\n\nWhat happened?\n\nMARLIN\n\nNo, no, no, kids. I don't wanna talk about it.\n\nTURTLE KIDS\n\nAwww! Please?\n\nSQUIRT\n\nPleeeease?\n\nMARLIN\n\n[sighs] Well, okay. I live on this reef, a long long way from here.\n\nDORY\n\nOh, boy. This is gonna be good, I can tell.\n\nMARLIN\n\nAnd my son, Nemo, see he was mad at me. And maybe he wouldn't have done it if I hadn't been\n\nso tough on him, I don't know. Anyway, he swam out in the open water to this boat and when\n\nhe was out there, these divers appeared and I tried to stop them but the boat was too fast.\n\nSo we swam out in the ocean to follow them...\n\nTURTLE KID\n\nThey couldn't stop them. And then Nemo's dad, he swims out to the ocean and they bump into..\n\nSMALL FISH\n\n..three ferocious sharks! He scares away the sharks by blowin' them up!\n\nBIG FISH\n\nGolly, that's amazing!\n\nSMALL FISH\n\nAnd then dives thousands of..\n\nLOBSTER\n\n..feet straight down into the dark. It's like wicked dark down there, you can see a thing.\n\nHow's it goin', Bob? And the only thing that they can see down there..\n\nSWORDFISH\n\n..is the light from this big horrible creature with razor sharp teeth. Nice parry, old man.\n\nAnd then he has to blast his way...\n\nDOLPHIN\n\nSo, these two little fish have been..searching the ocean for days. On the East Australian\n\nCurrent.\n\nFEMALE BIRD\n\nWhich means that he may be on his way here right now. That should put them in Sydney..\n\nMALE BIRD 1\n\n..Harbor in a matter of days. I mean, it sounds like this guy's gonna stop at..\n\n 36\n\nMALE BIRD 2\n\n..nothing until he finds his son. I sure hope he makes it.\n\nMALE BIRD 3\n\nThat's one dedicated father if you ask me.\n\nGULLS\n\nMine! Mine! Mine! Mine! Mine! Mine! Mine! Mine! Mine!\n\nNIGEL\n\nOh, would you just shut up! You're rats with wings!\n\nPELICAN\n\n..bloke's been lookin' for his boy Nemo.\n\nNIGEL\n\nNemo?\n\nPELICAN\n\nHe was taken off the reef by divers and this..\n\nNIGEL\n\nThere, take it! You happy!\n\nGULLS\n\nMine! Mine! Mine! Mine!\n\nNIGEL\n\nHey, hey, hey! Say that again! You said something about Nemo. What was it?\n\nGULLS\n\nMine! Mine! Mine!\n\nCRAB\n\nWhooooooaaa..watcha!\n\nGULL\n\nMine!\n\nPELICAN\n\nLast I heard, he's headin' towards the harbor.\n\nNIGEL\n\nHo ho! Brilliant!\n\n======================================================================================\n\nNEMO\n\n[sighs]\n\nDEB\n\nIs he doing okay?\n\nGURGLE\n\nI don't know, but whatever you do, don't mention D-A-R..\n\nNEMO\n\nIt's okay, I know who you're talking about.\n\nNEMO\n\nGill? Gill?\n\nGILL\n\nHey, Sharkbait.\n\nNEMO\n\nI'm sorry I couldn't stop the--\n\nGILL\n\nNo, I'm the one who should be sorry. I was so ready to get out, so ready to taste that\n\nocean. I was willing to put you in harm's way to get there. Nothing should be worth that.\n\nI'm sorry I couldn't get you back to your father, kid.\n\nNIGEL\n\nAll right! Hey, hey, hey, hey--!\n\n 37\n\nDENTIST\n\nWhat the!?\n\nPATIENT\n\nAAAAAAAAAH!!! Oooooh...\n\nDENTIST\n\nWell, uh, that's one way to pull a tooth. He he he he he! Huh, darn kids. Well, good\n\nthing I pulled the right one, eh, prime minister? He he he he!\n\nNIGEL\n\nHey, hey. Psst!\n\nPEACH\n\nOh, Nigel. You just missed an extraction.\n\nNIGEL\n\nOoh! Has he loosened the periodontal ligament yet--oh, what I'm talkin' about!? Nemo!\n\nWhere's Nemo? I gotta speak with him.\n\nNEMO\n\nWhat? What is it?\n\nNIGEL\n\nYour dad's been fighting the entire ocean looking for you.\n\nNEMO\n\nMy father? Really?\n\nGILL\n\nReally?\n\nNIGEL\n\nOh yeah. He's travelled hundreds of miles. He's been battling sharks and jellyfish and\n\nall sorts of--\n\nNEMO\n\nSharks? That can't be him.\n\nNIGEL\n\nAre you sure? What was his name? Some sort of sportfish or something: tuna, uh, trout..\n\nNEMO\n\nMarlin?\n\nNIGEL\n\nThat's it! Marlin! The little clownfish from the reef.\n\nNEMO\n\nIt's my dad! He took on a shark!\n\nNIGEL\n\nI heard he took on three.\n\nDEB/BLOAT/GURGLE\n\nThree!?\n\nGILL\n\nThree sharks!?\n\nBLOAT\n\nThat's gotta be forty eight hundred teeth!\n\nNIGEL\n\nYou see, kid, after you were taken by diver Dan over there, your dad followed the boat\n\nyou were on like a maniac.\n\nNEMO\n\nReally?\n\nNIGEL\n\nHe's swimming and he's swimming and he's giving it all he's got and then three gigantic\n\nsharks capture him and he blows them up! And then dives thousands of feet and gets chased\n\nby a monster with huge teeth! He ties this demon to a rock and what does he get for a\n\nreward? He gets to battle an entire jellyfish forest! And now he's riding with a bunch\n\nof sea turtles on the East Australian Current and the word is he's headed this way right\n\nnow, to Sydney!\n\n 38\n\nBLOAT\n\nWow! Ha ha ha!\n\nDEB\n\nOh, what a good daddy!\n\nGILL\n\nHe was lookin' for you after all, Sharkbait.\n\nGILL\n\n[gasps]\n\nGURGLE\n\nHe's swimming to the filter!\n\nGILL\n\n[gasps] Sharkbait!\n\nBLOAT\n\nNot again!\n\nGILL\n\nSharkbait!\n\nDEB\n\nNo!\n\nGURGLE\n\nYou've got your whole life ahead of you!\n\nBLOAT\n\nOh no!\n\nGILL\n\nWe'll help you, kid!\n\nBLOAT\n\nGotta get him out!\n\nDEB\n\nGimme that thing!\n\nDEB\n\nGet him outta there!\n\nGURGLE\n\nCome on, kid! Grab the end!\n\nALL\n\n[gasps]\n\nDEB\n\nSharkbait!\n\nBLOAT\n\nSharkbait! Are you okay!?\n\nGURGLE\n\nNo!\n\nGILL\n\nCan you hear me, Sharkbait!? Nemo! Can you hear me!?\n\nNEMO\n\nYeah, I can hear you.\n\nGILL\n\nSharkbait, you did it!\n\nGURGLE\n\nSharkbait, you're--covered with germs! Aaaaaaah!!!\n\nGILL\n\nThat took guts, kid.\n\nGILL\n\n 39\n\nAll right, gang. We have less than 48 hours before Darla gets here. This tank'll get\n\nplenty dirty in that time but we have to help it along any way we can. Jacques!\n\nJACQUES\n\nOui!\n\nGILL\n\nNo cleaning.\n\nJACQUES\n\nI shall resist.\n\nGILL\n\nEverybody else, be as gross as possible. Think dirty thoughts. We're gonna make this tank\n\nso filthy, the dentist'll have to clean it.\n\nBLOAT\n\n[belch]\n\nGILL\n\nGood work.\n\nNEMO\n\nHa ha ha ha!\n\n======================================================================================\n\nCRUSH\n\nAll right, we're here, dudes! Get ready! Your exit's comin' up, man!\n\nMARLIN\n\nWhere!? I don't see it!\n\nDORY\n\nRight there! I see it! I see it!\n\nMARLIN\n\nYou mean the swirling vortex of terror!?\n\nCRUSH\n\nThat's it, dude!\n\nMARLIN\n\nOf course it is.\n\nCRUSH\n\nOkay, first: find your exit buddy!\n\nCRUSH\n\nDo you have your exit buddy?\n\nDORY\n\nYes!\n\nCRUSH\n\nOkay, Squirt here will now give you a rundown of proper exiting technique!\n\nSQUIRT\n\nGood afternoon, we're gonna have a great jump today! Okay, crank a hard cutback as you hit\n\nthe wall! There's a screaming bottom turn, so watch out! Remember: rip it, roll it and\n\npunch it!\n\nMARLIN\n\nIt's like he's trying to speak to me, I know it! You know, you're really cute! But I don't\n\nknow what you're saying! Say the first thing again!\n\nCRUSH\n\nOkay, Jellyman! Go, go, go, go, go, go!\n\nMARLIN/DORY\n\nAaaaaaaaaah!!! Weeeeeeeeeeee!!! Whoooooooooooaaaaa!!! Aaaaaaaaaaah!!! Woohoooo!!!\n\nWhoooooaaa!!!\n\nDORY\n\nWhoooo!\n\nMARLIN\n\n 40\n\nHa ha ha ha! That was..fun! Ha ha! I actually enjoyed that!\n\nDORY\n\nHey, look! Turtles!\n\nCRUSH\n\nHa ha! Most excellent! Now, turn your fishy tails 'round and swim straight on through\n\nto Sydney! No worries, man!\n\nMARLIN\n\nNo worries! Thank you, dude Crush!\n\nTURTLE KIDS\n\nBye! Bye, Jellyman!\n\nCRUSH\n\nYou tell your little dude I said 'hi', okay?\n\nSQUIRT\n\nSee you later, dudes!\n\nDORY\n\nBye, everyone!\n\nMARLIN\n\nOh, Nemo would've loved this. Hey, ooh! Hey, Crush! Crush, I forgot! How old are you?\n\nCRUSH\n\nHundred and fifty, dude! And still young! Rock on!\n\nMARLIN\n\nHundred and fifty! Hundred and fifty, I gotta remember that.\n\nDORY\n\nWhoa. We goin' in there?\n\nMARLIN\n\nYup.\n\nDORY\n\nP. Sherman, 42 Wallaby Way, Sydney?\n\nMARLIN\n\nYup. We're gonna just swim straight.\n\nDORY\n\n[singing] Just keep swimming, just keep swimming.\n\nMARLIN\n\nDory?\n\n======================================================================================\n\nMARLIN\n\nBoy, this is taking a while.\n\nDORY\n\nHey, how about we play a game?\n\nMARLIN\n\nOkay.\n\nDORY\n\nUh, okay. I'm thinking of something, uh, orange. And it's small..\n\nMARLIN\n\nIt's me.\n\nDORY\n\nRight. Okay..\n\nDORY\n\n..orange, and uh, small..\n\nMARLIN\n\nIt's me.\n\n 41\n\nDORY\n\nAll righty, Mr. Smarty Pants.\n\nDORY\n\n..orange and small, and white stripes..\n\nMARLIN\n\nMe. And the next one's just a guess: me.\n\nDORY\n\nOkay, that's just scary.\n\nMARLIN\n\nW-w-wait, I have definitely seen this floating speck before. That means we've passed it\n\nbefore and that means we're going in circles and that means we're not going straight!\n\nDORY\n\nHey. Hey!\n\nMARLIN\n\nWe gotta get to the surface, come on! Let's figure it out up there. Let's go! Follow me!\n\nWha--?\n\nDORY\n\nWhoa, whoa, whoa! Hey! Relax. Take a deep breath. Now, let's ask somebody for directions.\n\nMARLIN\n\nOh, fine. Who do you wanna ask, the speck? There's nobody here!\n\nDORY\n\nWell, there has to be someone. It's the ocean, silly, we're not the only two in here.\n\nLet's see...okay, no one there. Uhh, nope. Nada. [gasps] There's somebody. Hey! Excuse--\n\nMARLIN\n\nDory! Dory! Dory! Okay, now it's my turn. I'm thinking of something dark and mysterious.\n\nIt's a fish we don't know. And if we ask it directions, it could ingest us and spit out\n\nour bones!\n\nDORY\n\nWhat is it with men and asking for directions?\n\nMARLIN\n\nLook, I don't wanna play the gender card right now. You wanna play a card? Let's play the\n\n'Let's Not Die' card.\n\nDORY\n\nYou wanna get outta here, don't you?\n\nMARLIN\n\nOf course, I do.\n\nDORY\n\nWell then, how are we gonna do that unless we give it a shot and hope for the best? Hmmm?\n\nHmmmm!? Come on, trust me on this.\n\nMARLIN\n\nAll right.\n\nDORY\n\nExcuse me! Woohoo! Little fella? Hello. Don't be rude, say 'hi'.\n\nMARLIN\n\nHa..hello.\n\nDORY\n\nHis son Bingo..\n\nMARLIN\n\nNemo.\n\nDORY\n\n..Nemo, was taken to, uh..\n\nMARLIN\n\nSydney.\n\nDORY\n\n 42\n\nSydney. Yes. And it's really, really important that we get there as fast as we can. So can\n\nyou help us out? Come on, little fella. Come on.\n\nMARLIN\n\nDory, I'm a little fella. I don't think that's a little fella.\n\nDORY\n\nOh. Oh, oh, big fella. Big fe--whale. Okay. Maybe he only speaks whale.\n\nMOOOOO-WEEEEEEE-NEEEEED...\n\nMARLIN\n\nUh, Dory..what're you doing?\n\nDORY\n\nTOOOOOOO-FIIIIIIND...\n\nMARLIN\n\nWhat're you doing?\n\nDORY\n\nHIS-SOOOOOOOOOOOON...\n\nMARLIN\n\nAre you sure you speak whale?\n\nDORY\n\nCAN-YOOOOOOOUUU-GIIIIIIIIIVE-USSSS-DIRECTIOOOOOOOONS-TOOOOOOOOO...\n\nMARLIN\n\nDory! Heaven knows what you're saying! See, he's swimming away.\n\nDORY\n\nCOOOME-BAAAAAAAAAAAAAACK!\n\nMARLIN\n\nHe's not coming back. You offended him.\n\nDORY\n\nMaybe a different dialect. MOOOOOOOOOOOOOO! MOOOOOAAAAAAAAAA..!\n\nMARLIN\n\nDory. Dory, this is not whale. You're speaking like..upset stomach.\n\nDORY\n\nMaybe I should try humpback.\n\nMARLIN\n\nNo, don't try humpback.\n\nDORY\n\nWAAAAAAAAAAAAAAOOOOOOO!!! WAAAAAAAAAOOOOOO!!!\n\nMARLIN\n\nOkay, you actually sound sick.\n\nDORY\n\nMaybe louder, huh? RAAAH!!! RAAAAH!!!\n\nMARLIN\n\nDon't do that!\n\nDORY\n\nToo much orca. Didn't it sound a little orca-ish?\n\nMARLIN\n\nIt doesn't sound orca! It sounds like nothing I've ever heard!\n\nDORY\n\nMOOOO..MOOOOOOOOOOOOOOO!!!\n\nMARLIN\n\nIt's just as well, he might be hungry.\n\nDORY\n\nDon't worry. Whales don't eat clownfish, they eat krill.\n\nKRILL\n\n 43\n\nSwim away!\n\nDORY\n\nOh, look. Krill.\n\nMARLIN\n\nMove, Dory! Move!\n\nDORY\n\nAah-aaah! Aaaaaaaaaah!\n\n======================================================================================\n\nGILL\n\nLook at that. Would you look at that? Filthy. Absolutely filthy. And it's all thanks to\n\nyou, kid. You made it possible. Jacques, I said no cleaning!\n\nJACQUES\n\nI am ashamed.\n\nPEACH\n\nHey, look. Scum angel.\n\nGURGLE\n\nAah! Aaaah! Ooh-ooh! Aaaaah!\n\nBUBBLES\n\nBubbles! I love the bubbles--! [coughs]\n\nDEB\n\nFlo! Flo! Has anybody seen Flo? Flo!\n\nPEACH\n\nNine o' clock and cue dentist.\n\nDENTIST\n\nHello, Barbara. Sorry I'm late.\n\nPEACH\n\nOkay. Okay, here we go. Here we go, okay.\n\nDENTIST\n\nLittle Davey Reynolds.\n\nPEACH\n\nOkay. Walks to the counter, drops the keys..\n\nGURGLE\n\nBloat, that's disgusting!\n\nBLOAT\n\nTastes pretty good to me. [belch]\n\nGURGLE\n\nEww! Don't you people realize we are swimming in our own--\n\nPEACH\n\nShhh! Here he comes.\n\nDENTIST\n\nCrikey, what a state. Oh. Barbara, what's my earliest appointment tomorrow?\n\nBARBARA\n\nUh, ten 'o clock, luv.\n\nDENTIST\n\nLeave it open, would you? I gotta clean the fish tank before Darla gets here.\n\nGILL\n\nHe he! Did you hear that, Sharkbait?\n\nNEMO\n\nYay! He's gonna clean the tank! He's gonna clean the tank! We're gonna be clean!\n\nGILL\n\nAre you ready to see your dad, kid?\n\n 44\n\nNEMO\n\nUh-huh.\n\nGILL\n\nOf course you are. Y'know, I wouldn't be surprised if he's out there in the harbor\n\nwaitin' for you right now.\n\nNEMO\n\nYeah.\n\n======================================================================================\n\nMARLIN\n\nAaaaaaaaaaaah! Ooof!\n\nDORY\n\nHa~~haaa~~haaaaaaah! Whooo!\n\nMARLIN\n\nAaaaaaaaaaaah!\n\nDORY\n\nHere comes a big one--whooooooo! Come on, you gotta try this!\n\nMARLIN\n\nWould you just stop it!?\n\nDORY\n\nWhy? What's wrong?\n\nMARLIN\n\nWe're in a whale! Don't you get it!?\n\nDORY\n\nA whale?\n\nMARLIN\n\nA whale! 'Cause you had to ask for help! And now we're stuck here!\n\nDORY\n\nWow. A whale. You know I speak whale.\n\nMARLIN\n\nNo, you're insane! You can't speak whale! I have to get out! I have to find my son!\n\nI have to tell him how old sea turtles are! [sobs]\n\nDORY\n\nWoo-ho-ho-ho-ho-ho-hoo! Hey. You okay?\n\nDORY\n\nThere, there. It's all right. It'll be okay.\n\nMARLIN\n\nNo. No, it won't.\n\nDORY\n\nSure it will, you'll see.\n\nMARLIN\n\nNo. I promised him I'd never let anything happen to him.\n\nDORY\n\nHuh. That's a funny thing to promise.\n\nMARLIN\n\nWhat?\n\nDORY\n\nWell, you can't never let anything happen to him. Then nothing would ever happen to him.\n\nNot much fun for little Harpo.\n\nDORY\n\nHmm..\n\nMARLIN\n\nWhat's going on?\n\n 45\n\nDORY\n\nI don't know. I'll ask him. MMMWWHAAAAAAAAA! HUUUWHAAAAAAAAA..\n\nMARLIN\n\nDory. Dory.\n\nMARLIN\n\n..AAAAAAAAAAT'SSS-GOOIIIIIIING..\n\nMARLIN\n\nDory.\n\nDORY\n\n..OOOOOOOOONNN?\n\nDORY\n\nI think he says we've stopped.\n\nMARLIN\n\nOf course, we've stopped. Just stop trying to speak whale, you're gonna make things worse.\n\n[gasps] What is that noise? Oh no. Look what you did. The water's going down!\n\nIt's-it's-it's going down!\n\nDORY\n\nReally? You sure about that?\n\nMARLIN\n\nLook, it's already half-empty!\n\nDORY\n\nHmm..I'd say it's half full.\n\nMARLIN\n\nStop that! It's half-empty!\n\nDORY\n\nOkay, that one was a little tougher. He either said we should go to the back of the throat\n\nor he wants a root beer float.\n\nMARLIN\n\nOf course he wants us to go there! That's eating us! How do I taste, Moby!? Huh!?\n\nDo I taste good!? You tell him I'm not interested in being lunch!\n\nDORY\n\nOkay. HEEEEEEEEE--\n\nMARLIN\n\nStop talking to him--waaaah!\n\nDORY\n\nAaaaaaaaaaaaaaaaaaah!!!\n\nMARLIN\n\nWhat is going on!?\n\nDORY\n\nI'll check! WHAAAAAAA--!\n\nMARLIN\n\nNo! No more whale! You can't speak whale!\n\nDORY\n\nYes, I can!\n\nMARLIN\n\nNo, you can't! You think you could do these things but you can't, Nemo!\n\nDORY\n\nOkay.\n\nMARLIN\n\nDory!\n\nDORY\n\nHe says it's time to let go! Everything's gonna be all right!\n\nMARLIN\n\n 46\n\nHow do you know!? How do you know something bad isn't gonna happen!?\n\nDORY\n\nI don't!\n\nMARLIN/DORY\n\nAAAAAAAAAAAAAAAAAAHHH!!! AAAAAAAAAAAAAAAHHH!!!\n\nMARLIN\n\nHa ha ha! We're alive!\n\nDORY\n\nLook! Sy-d-ney..Sydney! Uh, Sydney! Sydney again!\n\nMARLIN\n\nYou were right, Dory! We made it! We're gonna find my son!\n\nMARLIN\n\nTHAAAAAAAAAAAAAAANK-YOOOOOOOOOOOOOUUUU-SIIIRRRRRRRRRRRRRRRR!\n\nDORY\n\nWow. I wish I could speak whale.\n\nMARLIN\n\nOkay. All we gotta do is find the boat that took him.\n\nDORY\n\nRight!\n\nMARLIN\n\nCome on, Dory. We can do this!\n\n======================================================================================\n\nPEACH\n\n[yawn] Morning. [gasps] It's morning, everyone! Today's the day! The sun is shining, the\n\ntank is clean and we are getting out of--[gasps]--the tank is clean. The tank is clean!\n\nDEB\n\nBut how?\n\nGILL\n\nBoss must've installed it last night while we were sleepin'.\n\nNEMO\n\nWhat're we gonna do?\n\nGILL\n\nWhat's it say, Peach?\n\nPEACH\n\n[muffled] The AquaScum two-thousand..\n\nGILL\n\nI can't hear you, Peach.\n\nPEACH\n\n'The AquaScum 2003 is an all-purpose, self-cleaning maintenance free salt water purifier\n\nthat is guaranteed to even extend the life of your aquarium fish'.\n\nBLOAT\n\n[inflates] Stop it!\n\nPEACH\n\n'The AquaScum is programmed to scan your tank environment every 5 minutes'?\n\nGURGLE\n\nScan? What does that mean?\n\nGURGLE\n\nAaah!\n\nAQUASCUM\n\nTemperature: 82 degrees. PH balance: normal.\n\nALL\n\nOooooh.\n\n 47\n\nPEACH\n\nNice.\n\nGURGLE\n\nOoh..ah..curse you, AquaScum!\n\nBLOAT\n\nThat's it for the escape plan. It's ruined!\n\nNEMO\n\nThen what're we gonna do about--\n\nALL\n\n[gasps] Darla!\n\nGILL\n\nStay down, kid!\n\nBLOAT\n\nFalse alarm.\n\nGURGLE\n\nMy nerves can't take much more of this.\n\nBLOAT\n\nWhat're we gonna do when that little brat gets here?\n\nGILL\n\nI'm thinkin', I'm thinkin'.\n\nNEMO\n\nAaah! Oh! Gill!\n\nGILL\n\n[gasps] Nemo!\n\nNEMO\n\nHelp me! Help me!\n\nGILL\n\nHold on! I'm comin'!\n\nNEMO\n\nHelp me!\n\nGILL\n\nSwim down! Come on, kid! Swim down! Come on!\n\nBLOAT\n\nEverybody jump in!\n\nDEB\n\nSwim down!\n\nGILL\n\nThat's it!\n\nDENTIST\n\nWhat the!?\n\nALL\n\nYay!\n\nGILL\n\nGood work!\n\nNEMO\n\nGill!\n\nGILL\n\n[gasps] Nemo!\n\nBLOAT\n\nSharkbait!\n\nGILL\n\n 48\n\nRoll, kid! Lean! Lean!\n\nDENTIST\n\nWhoops. That would've been a nasty fall.\n\nNEMO\n\nGill! Don't let me go belly up!\n\nGILL\n\nJust calm down, Nemo.\n\nNEMO\n\nDon't let me go belly up!\n\nGILL\n\nYou won't go belly up, I promise. You're gonna be okay.\n\nALL\n\n[gasps] Darla!\n\n======================================================================================\n\nDORY\n\nAll right, do any of these boats look familiar to you?\n\nMARLIN\n\nNo, but the boat has to be here somewhere! Come on, Dory, we're gonna find it.\n\nDORY\n\nI'm totally excited. [yawn] Are you excited? [yawn]\n\nMARLIN\n\nDory, wake up, wake up. Come on.\n\nDORY\n\n[gasps] Duck!\n\nMARLIN\n\nThat's not a duck. It's a--pelican! Whooooaaaaah!\n\nDORY\n\nAaaaaaaaaaaah!\n\nMARLIN\n\nNo! I didn't come this far to be breakfast!\n\nPELICAN\n\nHey, hey, Nigel. Heh, would you look at that?\n\nNIGEL\n\nHuh? Wha-what?\n\nPELICAN\n\nSun's barely up and already Gerald's had more than he can handle.\n\nNIGEL\n\nYeah. Reckon somebody oughta help the poor guy.\n\nPELICANS\n\nYeah, yeah, right.\n\nNIGEL\n\nWell, don't everybody fly off at once.\n\nNIGEL\n\nAll right, Gerald, what is it? Fish got your tongue?\n\nDORY\n\nAaaaaaaaaaaaaah!!!\n\nNIGEL\n\nLove a duck!\n\nMARLIN\n\nI gotta find my son Nemo!\n\nNIGEL\n\n 49\n\n[gasps] Nemo? Hey, hey, hey! He's that fish! Y'know the one we were talking about!\n\nThe one that's been fighting the whole ocean! Hey, I know where your son i--huh?\n\nHey, wait! Come back! Stop!\n\nMARLIN\n\nDory, keep going! He's crazy!\n\nNIGEL\n\nI got something to tell 'ya!\n\nGULL\n\nMine.\n\nNIGEL\n\nOkay, don't make any sudden moves. Hop inside my mouth if you want to live.\n\nMARLIN\n\nHop in your mouth, huh? And how does that make me live?\n\nGULL\n\nMine.\n\nNIGEL\n\nBecause I can take you to your son.\n\nMARLIN\n\nYeah, right.\n\nNIGEL\n\nNo. I know your son. He's orange, he's got a gimpy fin on one side..\n\nMARLIN\n\nThat's Nemo!\n\nGULLS\n\nMine! Mine! Mine! Mine! Mine! Mine!\n\nDORY\n\nAaaaaaaaaaaaaah!!!\n\nNIGEL\n\nFasten your seatbelts!\n\nGULLS\n\nMine! Mine! Mine! Mine! Mine! Mine!\n\nDORY\n\nWhoooooo! Woohooooo!\n\nGULLS\n\nMine! Mine! Mine! Mine! Mine! Mine!\n\nDORY\n\nHa-haaaa! Ha ha ha ha!\n\nMARLIN\n\nAaaaaaaaaaaaaaaah!\n\nNIGEL\n\nEverybody hold on!\n\nMARLIN/DORY\n\nAaaaaaaaaaaaaaaaah!\n\nGULLS\n\nMine! Mine! Mine! Mine! Mine! Mine!\n\n======================================================================================\n\nBUBBLES\n\nAaaah! Too loud! Too loud for me!\n\nDARLA\n\n[singing] Twinkle, twinkle little star.\n\nPEACH\n\nFind a happy place, find a happy place, find a happy place!\n\n 50\n\nBARBARA\n\nDarla, you're uncle will see you now.\n\nDENTIST\n\nAll right, let's see those pearly whites.\n\nDARLA\n\nRAAAH! I'm a piranha. They're in the Amazon.\n\nDENTIST\n\nAnd a piranha's a fish, just like your present.\n\nDARLA\n\n[giggling] I get a fishy! Fishy, fishy, fishy!\n\nDENTIST\n\nOh no. Poor little guy.\n\nBLOAT\n\nHe's dead!\n\nGILL\n\nSharkbait!\n\nDARLA\n\nYay! Fishy, fishy, fishy!\n\nDENTIST\n\nHe he he! Must've left your present in the car, sweetie. Ha ha ha ha ha!\n\nDARLA\n\nAwwwww.\n\nDENTIST\n\nI'll go and get it.\n\nGILL\n\n[gasps] He's still alive!\n\nPEACH\n\nHe's not dead!\n\nBLOAT\n\nWhat's happening? Why is he playing dead?\n\nGILL\n\nHe's gonna get flushed down the toilet! He's gonna get outta here!\n\nDEB\n\nYay!\n\nBLOAT\n\nHe's gonna get flushed!\n\nGURGLE\n\nWhat a smart little guy!\n\nGILL\n\nOh no, not the trash can!\n\nBUBBLES\n\nNemo! No!\n\nNIGEL\n\nHey! Hey! I found his dad!\n\nMARLIN\n\nWhere's Nemo!? Where is he!?\n\nBLOAT\n\nDentist! Dentist!\n\nGILL\n\nHe's over there!\n\nMARLIN\n\n 51\n\nWhat's a dentist!? What is that!? [gasps] Nigel, get in there!\n\nNIGEL\n\nI can't go in there.\n\nMARLIN\n\nOh yes, you can! Charge!\n\nDARLA\n\nAaaaaaaaaaaah!\n\nDENTIST\n\nWhat the--!? Darla, sweetie! Look out!\n\nDARLA\n\nAaaaaaaah!\n\nDENTIST\n\nHold still!\n\nDARLA\n\nAaaaaaaah!\n\nDENTIST\n\nEasy! Easy!\n\nDARLA\n\nAaaaaaaah!\n\nDENTIST\n\nHold still! Nobody's going to hurt you! Oof!\n\nMARLIN\n\n[gasps] Nemo.\n\nDORY\n\n[gasps] Oh my goodness.\n\nDENTIST\n\nGotcha! Keep down!\n\nMARLIN\n\nNemo!\n\nNEMO\n\nDaddy?\n\nDENTIST\n\nOut with 'ya! And stay out!\n\nNEMO\n\nDaddy!?\n\nDARLA\n\nFishy? Fishy! Wake up! Wake up!\n\nDEB\n\nOh no!\n\nGILL\n\nQuick! To the top of Mt. Wannahockaloogie!\n\nDARLA\n\nWhy are you sleeping!?\n\nPEACH\n\nHurry!\n\nGILL\n\nBloat! Ring of Fire!\n\nDARLA\n\nFishy--aaaaaaaaaaaah! Aaaaaaaaaah!\n\nDENTIST\n\nWhat!? All the animals have gone mad!\n\n 52\n\nDARLA\n\nAaaaaaaah! Get it out!\n\nGURGLE\n\nSmack her in the head!\n\nBLOAT\n\nGo, Gill! Go!\n\nDARLA\n\nFish in my hair! Aaaaaaaah!\n\nNEMO\n\nGill.\n\nGILL\n\nSharkbait. Tell your dad..I said..hi. Go get 'em.\n\nDENTIST\n\nOoooh. [gasps]\n\nBLOAT\n\nHe did it! Ha ha!\n\nDEB\n\nYay!\n\nBUBBLES\n\nI'm so happy!\n\nGURGLE\n\nIs he gonna be okay, Gill?\n\nGILL\n\nDon't worry. All drains lead to the ocean.\n\nDARLA\n\nFishy!\n\nNEMO\n\nAaaaaaaaaaaaaah! Daddy!\n\n======================================================================================\n\nNIGEL\n\nI'm, I'm so sorry. Truly, I am.\n\nDORY\n\nHey..\n\nMARLIN\n\nDory. If it wasn't for you, I never even would have made it here. So, thank you.\n\nDORY\n\nHey! Hey, wait a minute. W-w-wait! Where are you going?\n\nMARLIN\n\nIt's over, Dory. We were too late. Nemo's gone and I'm going home now.\n\nDORY\n\nNo..no, you can't! Stop! Please don't go away. Please? No one's ever stuck with me for\n\nso long before. And if you leave, if you leave...I just, I remember things better with you.\n\nI do. Look, P. Sherman, 42..40..2..agh! I remember it, I do. It's there, I know it is\n\nbecause when I look at you, I can feel it. And I, I look at you and...I'm home. Please.\n\nI don't want them to go away. I don't wanna forget.\n\nMARLIN\n\nI'm sorry, Dory, but I do.\n\n======================================================================================\n\nCRAB 1\n\nManna from heavens!\n\nCRAB 2\n\nSweet nectar of life!\n\n 53\n\nCRAB 1/CRAB 2\n\nHey! Hey, hey! Hey!\n\nCRAB 1\n\nThis is our spot!\n\nCRAB 2\n\nGo on! Get outta here!\n\nCRAB 1/CRAB 2\n\nHey, hey! Hey! Hey, hey, hey!\n\nCRAB 1\n\nYeah, that's it fella! Just keep on swimmin', you got that!\n\nCRAB 2\n\nToo right, mate! Oh, Oh! I got a live one here!\n\nNEMO\n\nHey, have you seen my dad?\n\nCRAB 2\n\nGotcha! Hey! Hey! Come back here!\n\nCRAB 1\n\nYou let 'im go!\n\nCRAB 1/CRAB 2\n\nHey! Hey, hey, hey!\n\nNEMO\n\nDad! Dad! Dad!\n\nDORY\n\nAah! No!\n\nNEMO\n\nUm, excuse me. Are you all right?\n\nDORY\n\nI don't know where I am! I don't know what's going on, I think I lost somebody but I,\n\nI can't remember.\n\nNEMO\n\nIt's okay, it's okay. I'm looking for someone too. Hey, we can look together.\n\nDORY\n\nI'm Dory.\n\nNEMO\n\nI'm Nemo.\n\nDORY\n\nNemo? That's a nice name.\n\n======================================================================================\n\nNEMO\n\nDad!\n\nDORY\n\nDad!\n\nNEMO\n\nDad!\n\nDORY\n\nDad! Wait a minute, is it your dad or my dad?\n\nNEMO\n\nMy dad.\n\nDORY\n\nGot it. Dad!\n\nNEMO\n\nWhere are we, anyway?\n\n 54\n\nDORY\n\nDad! Dad! Oh. S-ss-syl--shi--Sydney. [gasps] 'P. Sherman, 42 Wallaby Way, Sydney'.\n\nDORY\n\nAaaaah! Nemo! It's you! Aaaaaah! You're Nemo!\n\nNEMO\n\n[muffled] Yes! Yes! I'm Nemo!\n\nDORY\n\nOh! You're Nemo! [gasps] You were dead! I saw you! And then I--[gasps], here you are!\n\nI found you! You're not dead! And your father--[gasps]! Your father!\n\nNEMO\n\nMy father!? You know my father!? Where is he!?\n\nDORY\n\n[gasps] This way! He went this way! Quick!\n\nDORY\n\nHey! Hey, hey! Hey!\n\nCRAB 1/CRAB 2\n\nHey! Hey, hey, hey!\n\nDORY\n\nHey! Have you seen an orange fish swim by? It looks just like him!\n\nNEMO\n\nBut bigger!\n\nCRAB 2\n\nYeah, I saw 'im, bluey! But I'm not tellin' you where he went. And there's no way you're\n\ngonna make me!\n\nGULL\n\nMine.\n\nCRAB\n\nHuh!? Aaaah! All right! I'll talk! I'll talk! He went to the fishing grounds! Aaaaah!\n\nGULLS\n\nMine!Mine! Mine! Mine! Mine! Mine!\n\n======================================================================================\n\nFISH\n\nHey! Look out!\n\nMARLIN\n\nSorry. Just trying to get home.\n\nNEMO\n\nDad! Dad!\n\nMARLIN\n\nNemo?\n\nNEMO\n\nDaddy!\n\nMARLIN\n\nNemo?\n\nNEMO\n\nDad!\n\nDORY\n\nNemo's alive!\n\nMARLIN\n\nDory? [gasps] Nemo!\n\nNEMO\n\nDaddy!\n\n 55\n\nMARLIN\n\nNemo! I'm coming, Nemo!\n\nNEMO\n\nDad!\n\nMARLIN\n\nNemo!\n\nNEMO\n\nDad!\n\nMARLIN\n\nOh, thank goodness! It's all right, son. It's gonna be okay.\n\nFISH\n\nTurn around! You're going the wrong way! Aaaaaaaaaaah!\n\nDORY\n\nAaaaaaaaaaaah! Look out!\n\nMARLIN\n\nMove! Move!\n\nFISH\n\nAaaaaaaaaaaah!\n\nDORY\n\nHelp! AAAAAAAAAAAAH!!!\n\nMARLIN\n\nDory!\n\nNEMO\n\nCome on!\n\nDORY\n\nHeeeeeeeelp!!! Help!\n\nNEMO\n\nDory!\n\nDORY\n\nHelp! Get us out! Aaaaaaaah!\n\nMARLIN\n\nNo, no, no! No! Dory!\n\nNEMO\n\nDad! I know what to do!\n\nMARLIN\n\nNemo! No!\n\nNEMO\n\nWe have to tell all the fish to swim down together!\n\nMARLIN\n\nGet out of there, now!\n\nNEMO\n\nI know this will work!\n\nMARLIN\n\nNo, I am not gonna lose you again!\n\nNEMO\n\nDad, there's no time! It's the only way we can save Dory! I can do this!\n\nMARLIN\n\nYou're right. I know you can.\n\nNEMO\n\nLucky fin!\n\nMARLIN\n\nNow go! Hurry!\n\n 56\n\nNEMO\n\nTell all of the fish to swim down!\n\nMARLIN\n\nWell!? You heard my son! Come on!\n\nNEMO\n\nDory!\n\nDORY\n\n[gasps]\n\nNEMO\n\nYou have to tell everybody to..\n\nMARLIN\n\n..swim down together! Do you understand what I'm saying to you!? Swim down!\n\nDORY\n\nEverybody swim down!\n\nNEMO\n\nCome on! You have to swim down!\n\nDORY\n\nSwim down, okay?\n\nNEMO\n\nSwim..\n\nMARLIN\n\ndown! Swim down! Swim down! Swim down!\n\nMARLIN\n\nDon't give up! Keep swimming! Just keep swimming!\n\nNEMO\n\nIt's working!\n\nFISH\n\nKeep swimming! Keep swimming! Keep swimming!\n\nMARLIN\n\nJust keep swimming! Keep swimming!\n\nNEMO\n\nCome on, dad!\n\nMARLIN\n\nYou're doing great, son!\n\nNEMO\n\nThat's my dad!\n\nMARLIN\n\nCome on! Let's get to the bottom! Keep swimming!\n\nDORY\n\n[singing] Just keep swimming, just keep swimming.\n\nMARLIN\n\nAlmost there! Keep swimming!\n\nFISH\n\nKeep swimming! Keep swimming! Keep swimming! Keep swimming! Yay!\n\nMARLIN\n\nOww!\n\nDORY\n\nHey!\n\nMARLIN\n\nDory! Where's Nemo!?\n\nDORY\n\n 57\n\n[gasps] There!\n\nMARLIN\n\nOh no. Nemo!\n\nMARLIN\n\nNemo? Nemo? It's okay. Daddy's here, daddy's got you.\n\nNEMO\n\n[coughs] Daddy?\n\nMARLIN\n\nOh, thank goodness.\n\nNEMO\n\nDad...I don't hate you.\n\nMARLIN\n\nNo, no, no. I'm so sorry, Nemo.\n\nMARLIN\n\nHey, guess what?\n\nNEMO\n\nWhat?\n\nMARLIN\n\nSea turtles? I met one! And he was a hundred and fifty years old.\n\nNEMO\n\nHundred and fifty?\n\nMARLIN\n\nYep.\n\nNEMO\n\n'Cause Sandy Plankton said they only live to be a hundred.\n\nMARLIN\n\nSandy Plankton? Do you think I would cross the entire ocean and not know as much as\n\nSandy Plankton!?\n\nNEMO\n\nHa ha ha ha!\n\nMARLIN\n\nHe was a hundred and fifty! Not one hundred! Who is this Sandy Plankton who knows everything?\n\n======================================================================================\n\nMARLIN\n\nTime for school! Time for school! Get up! Let's go! Go!\n\nMARLIN\n\nI'm gonna win!\n\nNEMO\n\nNo, you're not! I did it! Woohoo! Ha ha ha!\n\nMARLIN\n\nOh! My own son beats me!\n\nMR. RAY\n\nClimb aboard, explorers!\n\nMARLIN\n\nSo just then, the sea cucumber looks over to the mollusk and says : 'with fronds like these,\n\nwho needs anemones?'!\n\nBOB/TED/BILL\n\nHaaa-ha ha ha ha ha ha!\n\nMR. RAY\n\nWell, hello, Nemo! Who's this?\n\nNEMO\n\nExchange student.\n\n 58\n\nSQUIRT\n\nI'm from the EAC, dude!\n\nMR. RAY\n\nSweet.\n\nNEMO/SQUIRT\n\nTotally.\n\nBOB\n\nBut seriously, Marty, did you really do all the things you say you did?\n\nBRUCE\n\nUh, pardon me.\n\nBOB/TED/BILL\n\n[gasps]\n\nBRUCE\n\nHello.\n\nTED\n\nOhh!\n\nBRUCE\n\nDon't be alarmed.\n\nANCHOR\n\nOh, we just wanna make sure that our newest member got home safe.\n\nDORY\n\nThanks, guys.\n\nBRUCE\n\nWell, we'll see you next week.\n\nCHUM\n\nKeep up with the program, Dory.\n\nANCHOR\n\nRemember: fish are friends..\n\nDORY\n\n..not food! Bye!\n\nMR. RAY\n\nHold on! Here we go! Next up, knowledge!\n\nMARLIN\n\nBye, son! Have fun!\n\nNEMO\n\nBye, dad! Oh! Oh, Mr. Ray! Wait. I forgot something.\n\nNEMO\n\nLove you, dad.\n\nMARLIN\n\nI love you too, son.\n\nNEMO\n\nUh, dad, you can let go now.\n\nMARLIN\n\nSorry! Now go have an adventure!\n\nSQUIRT\n\nGoodbye! See you later, dudes!\n\nDORY\n\nBye, Elmo!\n\nMARLIN\n\nNemo.\n\nDORY\n\n 59\n\nNemo! Bye, Nemo!\n\nNEMO\n\nSee you after school, Dory! Bye, dad!\n\nMARLIN\n\nBye, son.\n\n======================================================================================\n\nDENTIST\n\nBarbara?\n\nBARBARA\n\nUh-huh?\n\nDENTIST\n\nI don't understand it. Here this thing has a lifetime guarantee and it breaks! Had to clean\n\nthe tank myself, take all the fish out, put 'em in bags and---where'd the fish go?\n\nGILL\n\nCome on, Peach!\n\nDEB\n\nHurry!\n\nGILL\n\nYou can do it!\n\nBLOAT\n\nYeah, that's it! You can do it!\n\nGURGLE\n\nJust a little further!\n\nPEACH\n\nThat's the shortest red light I've ever seen!\n\nBLOAT\n\nCome on, Peach!\n\nPEACH\n\nOooh--aaaaah!\n\nALL\n\nYay! We did it! Ha ha ha ha ha!\n\nBLOAT\n\nNow what?\n\n######################################################################################\n\n# FINDING NEMO, and all related media, characters, and stories #\n\n# are copyright 2003 Walt Disney Pictures and Pixar Animation Studios. #\n\n# The transcript below contains parts of a screenplay written by Andrew Stanton, #\n\n# Bob Peterson and David Reynolds. This transcript is provided for fans' enjoyment #\n\n# and reference and does not intend copyright infringement. The entire content of #\n\n# this transcript is property of Andrew Stanton, Bob Peterson and David Reynolds, #\n\n# Walt Disney Pictures and Pixar Animation Studios. #\n\n# No claim is lain on the ownership of the words contained within this transcript #\n\n# on the part of BaD_BURN. #\n\n# #\n\n# GIVE CREDIT WHERE CREDIT IS DUE. RETAIN THIS COMMENT BLOCK. #\n\n# #\n\n# The transcript is intended for teaching /educational purposes only. It falls under #\n\n# the U.S. Code 17/Sec. 107 - Limitations on exclusive rights: 'Fair Use'. #\n\n# Notwithstanding the provisions of sections 106 and 106A, the fair use of a #\n\n# copyrighted work, including such use by reproduction in copies or phonorecords or #\n\n# by any other means specified by that section, for purposes such as criticism, #\n\n# comment, news reporting, teaching (including multiple copies for classroom use), #\n\n# scholarship, or research, is not an infringement of copyright. #\n\n######################################################################################\n\n 60\n\n61\n\n62\n\n |
| 15 | script_frozendisney.txt | 0abeb1f8-b6c | Script | FROZEN\n\n Written by\n\n Jennifer Lee\n\n Final Shooting Draft\n\n 9/23/13\n\n OPEN ON: ICE.\n\n We're underwater looking up at it. A saw cuts through,\n\n heading right for us.\n\n EXT. SNOW-CAPPED MOUNTAINS -- DUSK\n\n ICE HARVESTERS, dressed in traditional Sami clothing, score a\n\n frozen lake. They SING.\n\n "The Frozen Heart (Ice Worker's Song)"\n\n ICE HARVESTERS\n\n BORN OF COLD AND WINTER AIR\n\n AND MOUNTAIN RAIN COMBINING,\n\n THIS ICY FORCE BOTH FOUL AND FAIR\n\n HAS A FROZEN HEART WORTH MINING.\n\n The men drag giant ice blocks through channels of water.\n\n ICE HARVESTERS (CONT'D)\n\n CUT THROUGH THE HEART, COLD AND CLEAR.\n\n STRIKE FOR LOVE AND STRIKE FOR FEAR.\n\n SEE THE BEAUTY SHARP AND SHEER.\n\n SPLIT THE ICE APART!\n\n AND BREAK THE FROZEN HEART.\n\n Hup! Ho! Watch your step! Let it go!\n\n A young Sami boy, KRISTOFF (8), and his reindeer calf, SVEN,\n\n share a carrot as they try to keep up with the men.\n\n ICE HARVESTERS (CONT'D)\n\n Hup! Ho! Watch your step! Let it go!\n\n Young Kristoff struggles to get a block of ice out of the\n\n water. He fails, ends up soaked. Sven licks his wet cheek.\n\n ICE HARVESTERS (CONT'D)\n\n BEAUTIFUL! POWERFUL! DANGEROUS! COLD!\n\n ICE HAS A MAGIC CAN'T BE CONTROLLED.\n\n A sharp ice floe overtakes the workers, threateningly. They\n\n fight it back.\n\n ICE HARVESTERS (CONT'D)\n\n STRONGER THAN ONE, STRONGER THAN TEN\n\n STRONGER THAN A HUNDRED MEN!\n\n Massive fjord horses drag heavy ice plows.\n\n 2\n\nFROZEN - J. Lee\n\n ICE HARVESTERS (CONT'D)\n\n BORN OF COLD AND WINTER AIR\n\n AND MOUNTAIN RAIN COMBINING\n\n The sun sets. Lanterns are lit.\n\n ICE HARVESTERS (CONT'D)\n\n THIS ICY FORCE BOTH FOUL AND FAIR\n\n HAS A FROZEN HEART WORTH MINING.\n\n CUT THROUGH THE HEART, COLD AND CLEAR.\n\n In the dark, Kristoff and Sven finally manage to get a single\n\n block of ice out of the water.\n\n ICE HARVESTERS (CONT'D)\n\n STRIKE FOR LOVE AND STRIKE FOR FEAR.\n\n THERE'S BEAUTY AND THERE'S DANGER HERE.\n\n SPLIT THE ICE APART!\n\n BEWARE THE FROZEN HEART.\n\n The workers pile onto the giant horse-drawn ice sled as it\n\n pulls away.\n\n Left behind, Kristoff and Sven push their ice block onto a\n\n dinky little sled then head off.\n\n We sweep up from them to the Northern Lights filling the\n\n sky...then move across the mountains...beneath the\n\n snowline...and descend upon...\n\n EXT. THE KINGDOM OF ARENDELLE -- NIGHT\n\n A humble castle, built of wood, nestled in a deep fjord.\n\n INT. CASTLE, NURSERY -- NIGHT\n\n ELSA (8) sleeps in her bed. Her little sister ANNA (5) pops\n\n up beside her.\n\n YOUNG ANNA\n\n Elsa. Psst. Elsa! Psst.\n\n Elsa doesn't stir. Anna sits on Elsa and bounces.\n\n YOUNG ANNA (CONT'D)\n\n Wake up. Wake up. Wake up.\n\n YOUNG ELSA\n\n (grumbling)\n\n Anna, go back to sleep.\n\n Anna rolls onto her back and spreads all her weight on Elsa.\n\n 3\n\nFROZEN - J. Lee\n\n YOUNG ANNA\n\n (drama queen-ish)\n\n I just can't. The sky's awake, so\n\n I'm awake, so we have to play.\n\n YOUNG ELSA\n\n ...Go play by yourself.\n\n Elsa shoves Anna off the bed.\n\n Anna lands butt to floor, sighs, defeated. But then she gets\n\n an idea. She hops back on the bed and lifts one of Elsa's\n\n eyelids.\n\n YOUNG ANNA\n\n (mischievously)\n\n Do you want to build a snowman?\n\n Elsa's eyes both pop open. She smiles.\n\n INT. CASTLE STAIRCASE -- NIGHT\n\n Anna, now wearing snow boots, pulls Elsa by the hand.\n\n YOUNG ANNA\n\n Come on, come on, come on, come on.\n\n Elsa tries to shush her, but Anna's too excited.\n\n INT. BALLROOM -- NIGHT\n\n The girls sneak into the ballroom. Elsa shuts the door.\n\n YOUNG ANNA\n\n Do the magic! Do the magic!\n\n Elsa laughs and waves her hands together. Snowflakes suddenly\n\n burst forth and dance between her palms, forming a snowball.\n\n Elsa throws the snowball high into the air. Snow bursts out\n\n and flurries around the room. Anna dances about, catching\n\n flakes in her palms and mouth.\n\n YOUNG ANNA (CONT'D)\n\n This is amazing!\n\n YOUNG ELSA\n\n Watch this!\n\n Elsa stomps her little slippered foot and a layer of ice\n\n suddenly coats the floor, forming a giant ice rink. Anna\n\n slides off, laughing.\n\n 4\n\nFROZEN - J. Lee\n\n PLAY MONTAGE:\n\n -Anna and Elsa roll giant snowballs and build a snowman\n\n together. Elsa moves his stick arms around.\n\n YOUNG ELSA (CONT'D)\n\n (goofy voice)\n\n Hi, I'm Olaf and I like warm hugs.\n\n Anna jumps up and hugs him.\n\n YOUNG ANNA\n\n I love you, Olaf.\n\n -Anna and Olaf appear to be dancing. REVEAL: Elsa is actually\n\n propelling them across the ice floor with her magic.\n\n -The girls slide down snowbanks together!\n\n -Anna fearlessly jumps off a snow peak into mid air.\n\n YOUNG ANNA (CONT'D)\n\n Catch me!\n\n Elsa makes another peak to catch Anna.\n\n YOUNG ELSA\n\n Gotcha!\n\n Anna keeps jumping. Elsa keeps casting magic.\n\n YOUNG ANNA\n\n (jumping faster)\n\n Again! Again!\n\n YOUNG ELSA\n\n (struggling to keep up)\n\n Slow down!\n\n Elsa suddenly slips.\n\n Her magic accidentally STRIKES Anna in the head. Anna tumbles\n\n down a snowbank and lands, unconscious.\n\n YOUNG ELSA (CONT'D)\n\n ANNA!\n\n Elsa runs to Anna and takes her in her arms. A streak of\n\n Anna's hair, where struck, turns white.\n\n YOUNG ELSA (CONT'D)\n\n MAMA! PAPA!\n\n The room around them fills with frightening ice spikes.\n\n 5\n\nFROZEN - J. Lee\n\n The parents burst through the frozen door. GASP at the sight\n\n of the room.\n\n KING\n\n Elsa, what have you done? This is\n\n getting out of hand!\n\n QUEEN\n\n (seeing Anna)\n\n Anna!\n\n The King and Queen rush to Anna and take her in their arms.\n\n ELSA\n\n It was an accident. I'm sorry,\n\n Anna.\n\n QUEEN\n\n (about Anna)\n\n She's ice cold.\n\n KING\n\n ...I know where we have to go.\n\n SLAM CUT TO:\n\n INT. DARK ROOM -- NIGHT\n\n The King sifts through a shelf to find an ancient book\n\n inscribed with Old Norse runes. He opens the book, scrambles\n\n to a page with an ancient map.\n\n EXT. ARENDELLE -- NIGHT\n\n Carrying the girls, the King and Queen ride their horses out\n\n of the kingdom. Snow streams from Elsa's hands, leaving a\n\n trail of ice behind them.\n\n EXT. FJORD MOUNTAIN FOREST -- NIGHT\n\n A sleepy Kristoff and Sven travel alone through the dark\n\n woods. All of a sudden, the King and Queen race by with the\n\n girls, leaving the wake of ice.\n\n KRISTOFF\n\n Ice?\n\n SLAM CUT TO:\n\n 6\n\nFROZEN - J. Lee\n\n EXT. BLACK MOUNTAINS -- NIGHT\n\n Kristoff rides Sven as they follow the trail of ice.\n\n YOUNG KRISTOFF\n\n Faster, Sven!\n\n EXT. THE VALLEY OF THE LIVING ROCK -- NIGHT\n\n Kristoff hops off Sven at the edge of a deep valley. They\n\n hide behind a rock and peek out.\n\n Down below, the King holds a frightened Elsa. The Queen holds\n\n the still unconscious Anna.\n\n KING\n\n Please, help. My daughter!\n\n Suddenly, a bunch of rocks tumble down the valley toward\n\n them. It looks as though they'll be crushed!\n\n But, luckily, the rocks stop at their feet. The rocks then\n\n unfold, revealing bright faces.\n\n YOUNG KRISTOFF\n\n Trolls...?\n\n The rock in front of Kristoff "wakes up." Meet BULDA.\n\n BULDA\n\n Shush. I'm trying to listen.\n\n She grabs Kristoff and Sven by hand and hoof and hugs them\n\n close. Sven licks her face and she eyes them both.\n\n BULDA (CONT'D)\n\n Cuties. I'm gonna keep you.\n\n Back below, the crowd parts for a troll as old as the Earth.\n\n They call him GRAND PABBIE. He approaches arthritically, but\n\n determined. He nods respectfully to the king.\n\n GRAND PABBIE\n\n Your Majesty.\n\n (referring to Elsa)\n\n Born with the powers or cursed?\n\n KING\n\n Born. And they're getting stronger.\n\n Grand Pabbie motions for the Queen to bring Anna to him. She\n\n does. He examines her.\n\n 7\n\nFROZEN - J. Lee\n\n GRAND PABBIE\n\n (about Anna)\n\n You are lucky it wasn't her heart.\n\n The heart is not so easily changed,\n\n but the head can be persuaded.\n\n KING\n\n Do what you must.\n\n GRAND PABBIE\n\n I recommend we remove all magic,\n\n even memories of magic to be\n\n safe.... But don't worry, I'll\n\n leave the fun.\n\n Grand Pabbie pulls out a glowing blue energy from Anna's\n\n head. We see her memories floating right above her. Grand\n\n Pabbie changes all of her magical memories to ordinary\n\n memories -- snowy play indoors with the girls in their\n\n nightgowns changes to outdoors on the winter fjords with the\n\n girls in winter gear. He puts the ordinary memories back in\n\n her head.\n\n GRAND PABBIE (CONT'D)\n\n She will be okay.\n\n YOUNG ELSA\n\n But she won't remember I have\n\n powers?\n\n KING\n\n It's for the best.\n\n PABBIE\n\n Listen to me, Elsa, your power will\n\n only grow.\n\n As he speaks, he conducts the Northern Lights to show a\n\n silhouette of an adult Elsa creating magical snowflakes.\n\n PABBIE (CONT'D)\n\n There is beauty in your magic....\n\n But also great danger.\n\n The snowflakes turn to sharp spikes.\n\n PABBIE (O.S.) (CONT'D)\n\n You must learn to control it.\n\n In the Northern Lights display, the sharp spikes cause human\n\n figures to panic and attack Elsa.\n\n PABBIE (CONT'D)\n\n Fear will be your enemy.\n\n 8\n\nFROZEN - J. Lee\n\n Elsa gasps and buries her face in the King's chest. The King\n\n wraps his arms around Elsa, protectively.\n\n KING\n\n No. We'll protect her. She can\n\n learn to control it. I'm sure.\n\n Over the King's words we...\n\n DISSOLVE TO:\n\n -The Arendelle castle gates shutting.\n\n KING (O.S.) (CONT'D)\n\n Until then, we'll lock the gates.\n\n We'll reduce the staff. We will\n\n limit her contact with people and\n\n keep her powers hidden from\n\n everyone... including Anna.\n\n -The castle shutters close.\n\n -Anna sits on her bed as Elsa's furniture disappears.\n\n -Anna rushes to the hall to see Elsa shut the door to her new\n\n room. Anna watches, confused and sad.\n\n DISSOLVE TO:\n\n INT. CASTLE WINDOW -- DAY\n\n We look out on a gentle snowfall. Little Anna skips up to the\n\n window. She lights up at the sight of the snow and rushes\n\n down the hall.\n\n INT. HALLWAY, ELSA'S DOOR -- DAY\n\n Anna knocks on Elsa's door and SINGS.\n\n "Do You Want to Build a Snowman?"\n\n YOUNG ANNA\n\n DO YOU WANT TO BUILD A SNOWMAN?\n\n COME ON LET'S GO AND PLAY.\n\n Anna peeks under the door.\n\n YOUNG ANNA (CONT'D)\n\n I NEVER SEE YOU ANYMORE.\n\n COME OUT THE DOOR.\n\n IT'S LIKE YOU'VE GONE AWAY.\n\n 9\n\nFROZEN - J. Lee\n\n -INT. ANNA'S ROOM -- Anna plays with two dolls, gives up, sad.\n\n YOUNG ANNA (CONT'D)\n\n WE USED TO BE BEST BUDDIES\n\n AND NOW WE'RE NOT.\n\n I WISH YOU WOULD TELL ME WHY.\n\n -ELSA'S DOOR. Anna peeks through the key hole.\n\n YOUNG ANNA (CONT'D)\n\n DO YOU WANT TO BUILD A SNOWMAN?\n\n -Anna calls through the keyhole.\n\n YOUNG ANNA (CONT'D)\n\n IT DOESN'T HAVE TO BE A SNOWMAN.\n\n YOUNG ELSA (O.S.)\n\n Go away, Anna.\n\n YOUNG ANNA\n\n (hearbroken)\n\n ...OKAY BYE.\n\n -BEHIND THE DOOR -- DAY. Elsa sits at the window looking out,\n\n longingly. Suddenly, her icy hands freeze the windowsill.\n\n -LATER. The King slips leather gloves onto Elsa's hands.\n\n KING\n\n The gloves will help.\n\n He pats her gloved hand.\n\n KING (CONT'D)\n\n See? You're good....\n\n (starting their mantra)\n\n Conceal it.\n\n YOUNG ELSA\n\n Don't feel it.\n\n YOUNG ELSA & KING\n\n Don't let it show.\n\n -INT. HALLWAY, ELSA'S DOOR -- DAY. Anna, now 9, knocks on\n\n Elsa's door.\n\n ANNA (9)\n\n DO YOU WANT TO BUILD A SNOWMAN?\n\n -INT. HALLWAY -- DAY. Alone, Anna rides a bicycle built for\n\n two in the hall by standing on the back seat.\n\n 10\n\nFROZEN - J. Lee\n\n ANNA (9) (CONT'D)\n\n OR RIDE OUR BIKE AROUND THE HALL?\n\n I THINK SOME COMPANY IS OVERDUE...\n\n -INT. PORTRAIT ROOM -- DAY. Anna runs around the portrait\n\n room, gaining momentum to flip over the arm of the couch.\n\n ANNA (9) (CONT'D)\n\n I'VE STARTED TALKING TO\n\n THE PICTURES ON THE WALLS.\n\n Anna lands PLOP on the cushions, then looks up at the\n\n painting above her of the courageous Joan of Arc.\n\n ANNA (9) (CONT'D)\n\n Hang in there, Joan.\n\n -INT. EMPTY LIBRARY -- DAY. Looks like no one's around.\n\n ANNA (9) (CONT'D)\n\n IT GETS A LITTLE LONELY\n\n ALL THESE EMPTY ROOMS.\n\n But then we find Anna, laying at the base of the grandfather\n\n clock, playing with her braids, bored out of her mind.\n\n ANNA (9) (CONT'D)\n\n JUST WATCHING THE HOURS TICK BY.\n\n Anna's eyes follow the grandfather clock's pendulum.\n\n ANNA (9) (CONT'D)\n\n TICK TOCK. TICK TOCK. TICK TOCK.\n\n -INT. ELSA'S ROOM -- NIGHT. Elsa (now 12) paces as she panics.\n\n The entire wall is frozen behind her.\n\n ELSA (12)\n\n I'm scared. It's getting stronger.\n\n KING\n\n Getting upset only makes it worse.\n\n The King goes to hug her.\n\n ELSA (12)\n\n No. Don't touch me. I don't want to\n\n hurt you.\n\n He and the Queen look at each other with alarmed sadness.\n\n -INT. LIBRARY -- DAY. Anna, now a teenager, slides past Elsa's\n\n room without stopping.\n\n 11\n\nFROZEN - J. Lee\n\n -INT. KING AND QUEEN'S QUARTERS -- DAY. Anna runs into the\n\n room and throws herself into her parents' arms.\n\n TEEN ANNA\n\n See you in two weeks.\n\n -INT. ELSA'S ROOM -- DAY. Elsa curtsies in front of her\n\n parents, formally, not touching them.\n\n TEEN ELSA\n\n Do you have to go?\n\n KING\n\n You'll be fine, Elsa.\n\n -EXT. DOCKS -- DAY. The King and Queen leave on a ship.\n\n -EXT. ROUGH SEAS -- NIGHT. Lightning flashes. The sea rages in\n\n a storm. The King and Queen's ship is lost in the waves.\n\n -INT. CASTLE -- DAY. A portrait of the King and Queen is\n\n covered in mourning cloth.\n\n -EXT. CEMETERY -- DAY. Anna looks small, standing before her\n\n people, beside burial stones.\n\n -INT. HALLWAY, ELSA'S DOOR. Anna, still in her mourning\n\n clothes, approaches and knocks.\n\n ANNA\n\n (singing)\n\n Elsa? PLEASE I KNOW YOU'RE IN THERE\n\n PEOPLE ARE ASKING WHERE YOU'VE BEEN\n\n THEY SAY HAVE COURAGE\n\n AND I'M TRYING TO\n\n I'M RIGHT OUT HERE FOR YOU.\n\n PLEASE LET ME IN.\n\n Anna slides down the door and sits with her head against it.\n\n ANNA (CONT'D)\n\n WE ONLY HAVE EACH OTHER.\n\n IT'S JUST YOU AND ME.\n\n WHAT ARE WE GONNA DO?\n\n (weak, internal)\n\n DO YOU WANT TO BUILD A SNOWMAN?\n\n We move through the door...\n\n -INT. ELSA'S ROOM -- DAY. Elsa is sitting in the exact same\n\n pose as Anna. Her bedroom is frozen with ice. Snowflakes hang\n\n in the air, suspended by grief.\n\n FADE OUT.\n\n 12\n\nFROZEN - J. Lee\n\n EXT. THE KINGDOM OF ARENDELLE -- MORNING\n\n A new dawn rises over the fjords.\n\n Ships pull up to the docks. Guests pile out.\n\n DOCK MASTER\n\n Welcome to Arendelle!\n\n A BOY tries to get away as his MOTHER tries to stuff him in\n\n his bunad jacket.\n\n BOY\n\n Why do I have to wear this?\n\n MOTHER\n\n Because the Queen has come of age.\n\n It's Coronation Day!\n\n BOY\n\n That's not my fault.\n\n They pass the May Pole being raised and a Sami ice harvester\n\n chatting with his reindeer. We recognize them as Kristoff and\n\n Sven, all grown up. Sven hops around excitedly like a dog and\n\n nuzzles Kristoff's chest.\n\n KRISTOFF\n\n What do you want, Sven?\n\n Kristoff leans in and speaks for Sven, as if he can.\n\n KRISTOFF (AS SVEN) (CONT'D)\n\n Give me a snack.\n\n KRISTOFF (CONT'D)\n\n What's the magic word?\n\n KRISTOFF (AS SVEN) (CONT'D)\n\n Please!\n\n Kristoff pulls a carrot out of his shirt pocket and hands it\n\n to Sven. Sven tries to bite the whole thing.\n\n KRISTOFF (CONT'D)\n\n Hey, hey, hey! Share!\n\n Sven takes a smaller bite. Kristoff then has a bite himself,\n\n not seeming to care that it's covered in reindeer slobber.\n\n We move on to PERSI and AGGIE, a super-excited couple who\n\n rush towards the castle.\n\n 13\n\nFROZEN - J. Lee\n\n PERSI\n\n I can't believe they're finally\n\n opening up the gates!\n\n AGGIE\n\n And for a whole day! Faster, Persi!\n\n They pass a tiny but menacing DUKE, who wears taps on his\n\n shoes to "enhance" his presence. Two THUG guards follow close\n\n behind him.\n\n DUKE\n\n Ah, Arendelle, our most mysterious\n\n trade partner. Open those gates so\n\n I may unlock your secrets and\n\n exploit your riches.\n\n (catching himself)\n\n ...Did I just say that out loud?\n\n We leave him and head down the bridge towards the castle\n\n gates, passing an Irishman and a Spanish Dignitary.\n\n IRISHMAN\n\n Oh, me sore eyes can't wait to see\n\n the Queen and the Princess. I bet\n\n they're absolutely lovely.\n\n SPANISH DIGNITARY\n\n I bet they are beautiful.\n\n We move past them, to a particular castle window.\n\n CUT TO:\n\n INT. CASTLE, ANNA'S BEDROOM -- DAY\n\n Anna, 18, snores. Drools. KNOCK. KNOCK.\n\n KAI (O.S.)\n\n Princess Anna...?\n\n Anna sits up. She's got major bedhead. She coughs. Snorts.\n\n Pulls a hair from her mouth.\n\n ANNA\n\n ...Huh? Yeah?\n\n KAI (O.S.)\n\n Sorry to wake you, ma'am but--\n\n ANNA\n\n No, you didn't. I've been up for\n\n hours.\n\n 14\n\nFROZEN - J. Lee\n\n She falls back asleep while sitting. She snores. Her head\n\n drops, startling her awake.\n\n ANNA (CONT'D)\n\n Who is it?\n\n KAI (O.S.)\n\n It's still me, ma'am. Time to get\n\n ready.\n\n ANNA\n\n Ready for what?\n\n KAI (O.S.)\n\n Your sister's coronation, ma'am.\n\n ANNA\n\n My sister's cor-neration...\n\n One eye opens enough to catch sight of her coronation dress.\n\n She bolts, wide awake in excitement.\n\n ANNA (CONT'D)\n\n Coronation Day! Ha ha!\n\n SLAM CUT TO:\n\n EXT. CASTLE HALL -- DAY\n\n Anna bursts out of her room, wearing her coronation dress.\n\n She finishes pinning ribbons in her hair. Seeing the hustle\n\n and bustle of preparations, she can't help but SING.\n\n "For the First Time in Forever"\n\n ANNA\n\n THE WINDOW IS OPEN!\n\n SO'S THAT DOOR!\n\n I DIDN'T KNOW THEY DID THAT ANYMORE.\n\n WHO KNEW WE OWNED 8000 SALAD PLATES...?\n\n -Anna slides along the floor of the ballroom in her socks.\n\n ANNA (CONT'D)\n\n FOR YEARS I HAVE ROAMED THESE EMPTY HALLS\n\n WHY HAVE A BALLROOM WITH NO BALLS?\n\n FINALLY, THEY'RE OPENING UP THE GATES!\n\n -She shakes hands with a suit of armor. Breaks it. Hides the\n\n evidence.\n\n 15\n\nFROZEN - J. Lee\n\n ANNA (CONT'D)\n\n THERE'LL BE REAL, ACTUAL PEOPLE -\n\n IT'LL BE TOTALLY STRANGE.\n\n BUT WOW AM I SO READY FOR THIS CHANGE!\n\n -Anna comes to a window and jumps out onto a window washer's\n\n pulley. She raises herself up to see the ships arriving.\n\n ANNA (CONT'D)\n\n FOR THE FIRST TIME IN FOREVER,\n\n THERE'LL BE MUSIC, THERE'LL BE LIGHT.\n\n FOR THE FIRST TIME IN FOREVER,\n\n I'LL BE DANCING THROUGH THE NIGHT.\n\n -Anna walks through the garden and follows a family of geese.\n\n ANNA (CONT'D)\n\n DON'T KNOW IF I'M ELATED OR GASSY,\n\n BUT I'M SOMEWHERE IN THAT ZONE\n\n 'CAUSE FOR THE FIRST TIME IN FOREVER,\n\n I WON'T BE ALONE.\n\n (speaking)\n\n I can't wait to meet everyone....\n\n (GASP) What if I meet THE ONE?\n\n -Anna twists herself in a velvet drape like it's a gown. She\n\n acts like she looks gorgeous, but she looks ridiculous.\n\n ANNA (CONT'D)\n\n TONIGHT, IMAGINE ME GOWN AND ALL-\n\n FETCHINGLY DRAPED AGAINST THE WALL.\n\n THE PICTURE OF SOPHISTICATED GRACE.\n\n -She notices the bust of a man across the room.\n\n ANNA (CONT'D)\n\n (google-eyed)\n\n I SUDDENLY SEE HIM STANDING THERE,\n\n A BEAUTIFUL STRANGER TALL AND FAIR.\n\n (mouth full of chocolate)\n\n I WANNA STUFF SOME CHOCOLATE IN MY\n\n FACE!\n\n -She grabs the bust of the man and swings it around.\n\n ANNA (CONT'D)\n\n BUT THEN WE LAUGH AND TALK ALL EVENING,\n\n WHICH IS TOTALLY BIZARRE.\n\n NOTHING LIKE THE LIFE I'VE LED SO FAR.\n\n The bust goes flying and lands on the top of the cake.\n\n -Anna bursts into the portrait room, bounces on the\n\n furniture, and interacts with the paintings.\n\n 16\n\nFROZEN - J. Lee\n\n ANNA (CONT'D)\n\n FOR THE FIRST TIME IN FOREVER,\n\n THERE'LL BE MAGIC, THERE'LL BE FUN.\n\n FOR THE FIRST TIME IN FOREVER,\n\n I COULD BE NOTICED BY SOMEONE.\n\n AND I KNOW IT IS TOTALLY CRAZY\n\n TO DREAM I'D FIND ROMANCE.\n\n BUT FOR THE FIRST TIME IN FOREVER,\n\n AT LEAST I'VE GOT A CHANCE!\n\n -INT. LIBRARY. ELSA, now a very poised 21, watches out the\n\n window as the coronation guests arrive.\n\n ELSA\n\n DON'T LET THEM IN.\n\n DON'T LET THEM SEE.\n\n BE THE GOOD GIRL\n\n YOU ALWAYS HAVE TO BE.\n\n Elsa moves to a painting of her father's coronation. She\n\n takes off her gloves and mimics the painting by holding a\n\n candlestick and ornament in place of an orb and scepter.\n\n ELSA (CONT'D)\n\n CONCEAL. DON'T FEEL.\n\n PUT ON A SHOW.\n\n MAKE ONE WRONG MOVE\n\n AND EVERYONE WILL KNOW.\n\n The candlestick and ornament ice over. Elsa gasps, slams them\n\n back down onto the table. She tries to reassure herself.\n\n ELSA (CONT'D)\n\n BUT IT'S ONLY FOR TODAY.\n\n We cut between Anna's excitement and Elsa's nerves.\n\n ANNA\n\n IT'S ONLY FOR TODAY!\n\n ELSA\n\n IT'S AGONY TO WAIT.\n\n ANNA\n\n IT'S AGONY TO WAIT!!!\n\n ELSA\n\n TELL THE GUARDS TO OPEN UP THE GATE.\n\n ANNA\n\n THE GATE!!!\n\n -Finally, the gates are open! Anna moves through the crowd,\n\n admiring the people around her.\n\n 17\n\nFROZEN - J. Lee\n\n ANNA (CONT'D) ELSA\n\n FOR THE FIRST TIME IN DON'T LET THEM IN\n\n FOREVER. DON'T LET THEM SEE\n\n ANNA ELSA\n\n I'M GETTING WHAT I'M DREAMING BE THE GOOD GIRL\n\n OF YOU ALWAYS HAVE TO BE\n\n ANNA ELSA\n\n A CHANCE TO LEAVE MY SISTER'S CONCEAL.\n\n WORLD CONCEAL. DON'T FEEL.\n\n A CHANCE TO FIND TRUE LOVE DON'T LET THEM KNOW.\n\n -Anna hurries over the bridge and into the village square.\n\n ANNA (CONT'D)\n\n I KNOW IT ALL ENDS TOMORROW,\n\n SO IT HAS TO BE TODAY!!\n\n `CAUSE FOR THE FIRST TIME IN\n\n FOREVER. . .\n\n FOR THE FIRST TIME IN FOREVER!\n\n NOTHING'S IN MY WAY!!!\n\n -Anna SLAMS right into the breast of a HORSE!\n\n She falls back and lands in a small wooden boat. It tips off\n\n of the dock. She's heading overboard. But just then, the\n\n horse slams his hoof into the boat and steadies it.\n\n ANNA (CONT'D)\n\n (frustrated)\n\n Hey!\n\n HANS\n\n I'm so sorry. Are you hurt?\n\n The rider, HANS, sure is handsome and regal.\n\n ANNA\n\n (gentler)\n\n Hey. I-ya, no. No. I'm okay.\n\n HANS\n\n Are you sure?\n\n ANNA\n\n Yeah, I just wasn't looking where I\n\n was going. But I'm okay.\n\n He hops down from his horse and steps into the boat.\n\n ANNA (CONT'D)\n\n I'm great, actually.\n\n 18\n\nFROZEN - J. Lee\n\n HANS\n\n Oh, thank goodness.\n\n He offers her a hand and their eyes meet. Chemistry. He helps\n\n her to her feet.\n\n HANS (CONT'D)\n\n (bowing)\n\n Prince Hans of the Southern Isles.\n\n ANNA\n\n (curtseying)\n\n Princess Anna of Arendelle.\n\n HANS\n\n Princess...? My Lady.\n\n He drops to his knees, head bowed. The horse bows too,\n\n curling his hoof up and out of the boat.\n\n The boat tips. Hans tumbles on top of Anna. Awkward.\n\n ANNA\n\n Hi...again.\n\n The horse slams his foot back into the boat to stabilize it.\n\n Anna and Hans tumble the other way. Anna lands on top of him.\n\n HANS\n\n Oh boy.\n\n ANNA\n\n Ha. This is awkward. Not you're\n\n awkward, but just because we're--\n\n I'm awkward. You're gorgeous.\n\n (did she just say that?)\n\n Wait, what?\n\n Hans quickly gets to his feet and helps Anna up again.\n\n HANS\n\n I'd like to formally apologize for\n\n hitting the Princess of Arendelle\n\n with my horse...and for every\n\n moment after.\n\n ANNA\n\n No. No-no. It's fine. I'm not THAT\n\n Princess. I mean, if you'd hit my\n\n sister Elsa, that would be-- yeash!\n\n `Cuz, you know...\n\n (patting the horse)\n\n Hello.\n\n (MORE)\n\n 19\n\nFROZEN - J. Lee\n\n ANNA (CONT'D)\n\n (to Hans)\n\n But, lucky you, it's-it's just me.\n\n HANS\n\n Just you?\n\n Hans smiles, amused. She smiles back. The bells RING. She\n\n doesn't notice at first; she's too busy drinking in Hans's\n\n handsomeness.\n\n ANNA\n\n ...The bells. The coronation. I-I-I\n\n better go. I have to...I better go.\n\n She hurries off, stops, turns back. Gives Hans a little wave.\n\n ANNA (CONT'D)\n\n Bye!\n\n As she rushes off again, Hans waves back. The horse waves\n\n too, once again taking his hoof out of the boat.\n\n HANS\n\n Oh no.\n\n The boat falls, with Hans in it. SPLASH! It lands upside down\n\n in the water. Hans raises it up off of him, gasping for air.\n\n CUT TO:\n\n INT. CHURCH CHAPEL -- DAY\n\n Elsa stands at the alter. Anna stands off to one side. She\n\n peeks out to the audience.\n\n Hans waves at her from the pews. He's changed his clothes.\n\n The crown is placed on Elsa's head. The scepter and orb are\n\n presented to Elsa on a pillow. She slowly reaches for them.\n\n BISHOP\n\n (a whisper)\n\n Your Majesty, the gloves.\n\n Elsa hesitates. She breathes nervously, removes her gloves,\n\n places them on the pillow. Her hands shake. She takes the orb\n\n and scepter, then turns to the people.\n\n BISHOP (CONT'D)\n\n (formal, in Old Norse)\n\n Sehm hon HELL-drr IN-um HELL-gum\n\n AYG-num ok krund ee THES-um HELL-\n\n gah STAHTH, ehk teh frahm FUR-ear U-\n\n thear...\n\n 20\n\nFROZEN - J. Lee\n\n The scepter and orb start to freeze over.\n\n BISHOP (CONT'D)\n\n ...Queen Elsa of Arendelle.\n\n CROWD\n\n Queen Elsa of Arendelle.\n\n Just in time. Elsa manages to set the orb and scepter back\n\n down on the pillow before anyone notices the ice. She picks\n\n up her gloves and slips them on. She made it.\n\n CUT TO:\n\n INT. GREAT HALL -- NIGHT\n\n Springy music fills the Great Hall. Guests dance. Eat. Laugh.\n\n TRUMPETS SOUND.\n\n KAI\n\n (announcing)\n\n Queen Elsa of Arendelle.\n\n Elsa enters, poised and looking surprisingly content. She\n\n stands under a formal awning.\n\n KAI (CONT'D)\n\n Princess Anna of Arendelle!\n\n Anna runs into the room, waves awkwardly. Kai ushers her over\n\n to stand right next to Elsa.\n\n ANNA\n\n Here? Are you sure?\n\n She and Elsa sneak awkward peeks at each other.\n\n ELSA\n\n ...Hi.\n\n ANNA\n\n Hi me...? Oh. Um. Hi.\n\n ELSA\n\n ...You look beautiful.\n\n ANNA\n\n Thank you. You look beautifuller. I\n\n mean, not fuller. You don't look\n\n fuller, but more beautiful.\n\n 21\n\nFROZEN - J. Lee\n\n ELSA\n\n Thank you.\n\n They look out at the celebration.\n\n ELSA (CONT'D)\n\n So, this is what a party looks\n\n like?\n\n ANNA\n\n It's warmer than I thought.\n\n ELSA\n\n And what is that amazing smell?\n\n They both close their eyes and inhale.\n\n ANNA AND ELSA (TOGETHER)\n\n ...Chocolate.\n\n Their eyes pop open. They laugh.\n\n Elsa looks back out at the party. Anna looks at Elsa. She\n\n wants to say so much, but she can't think of where to start.\n\n Just as she finds her way, Kai interrupts.\n\n KAI\n\n Your Majesty. The Duke of\n\n Weaseltown.\n\n DUKE\n\n Weselton. The Duke of Weselton.\n\n (to Elsa)\n\n Your Majesty, as your closest\n\n partner in trade, it seems only\n\n fitting that I offer you your first\n\n dance as queen.\n\n The Duke does a funny flitter of his feet, a hitch-kick, and\n\n a deep bow.\n\n DUKE (CONT'D)\n\n (whispers to himself)\n\n One, two, three. Jump.\n\n As he holds out his hand, head down, his toupee dips forward.\n\n Anna giggles. Elsa looks at Anna, stifles a giggle herself.\n\n ELSA\n\n (to the Duke)\n\n Thank you...only I don't dance.\n\n 22\n\nFROZEN - J. Lee\n\n DUKE\n\n (offended)\n\n Oh...?\n\n ELSA\n\n But my sister does.\n\n ANNA\n\n What?\n\n DUKE\n\n Lucky you....\n\n ANNA\n\n Oh, I don't think--\n\n The Duke grabs Anna's arm and yanks her away before she can\n\n protest.\n\n DUKE\n\n If you swoon, let me know, I'll\n\n catch you.\n\n Anna looks back at Elsa, desperately.\n\n ELSA\n\n Sorry.\n\n OUT ON THE DANCE FLOOR: The Duke showboats, but he's just\n\n awful. Anna tries to make the best of it.\n\n DUKE\n\n Like an agile peacock... CLUCK-\n\n CLUGGLE-CLUCK!\n\n He lands on her feet.\n\n ANNA\n\n Ow. Ow.\n\n DUKE\n\n Speaking of, so great to have the\n\n gates open. Why did they shut them\n\n in the first place? Do you know the\n\n reason? Hmm?\n\n He gets in her face, suspicious.\n\n ANNA\n\n ...No.\n\n 23\n\nFROZEN - J. Lee\n\n DUKE\n\n Oh, all right. Hang on. They don't\n\n call me the little dipper for\n\n nothing.\n\n He dips Anna back. Elsa peeks through the crowd, can barely\n\n hold in her laughter. Anna shoots Elsa funny, help-me looks.\n\n DUKE (CONT'D)\n\n (groove fully on)\n\n Like a chicken...with the face of a\n\n monkey...I fly.\n\n JUMP CUT TO:\n\n MOMENTS LATER...\n\n Anna limps back to Elsa.\n\n DUKE (O.S.)\n\n Let me know when you're ready for\n\n another round, M'Lady.\n\n ELSA\n\n Well, he was sprightly.\n\n ANNA\n\n (rubbing her sore feet)\n\n Especially for a man in heels.\n\n ELSA\n\n Are you okay?\n\n ANNA\n\n (loving Elsa's attention)\n\n I've never been better. This is so\n\n nice. I wish it could be like this\n\n all the time.\n\n ELSA\n\n (sincere)\n\n Me too....\n\n But then Elsa catches herself. She stiffens up, looks away.\n\n ELSA (CONT'D)\n\n But it can't.\n\n ANNA\n\n Why not? If--\n\n ELSA\n\n It just can't.\n\n 24\n\nFROZEN - J. Lee\n\n Anna's smile drops. She tries not to get emotional.\n\n ANNA\n\n Excuse me for a minute.\n\n She walks away. Elsa watches her go, saddened.\n\n Moving through the crowd, Anna gets bumped by a bowing man's\n\n butt. She falls. Just before she hits the floor, Hans catches\n\n her. He smiles perfectly.\n\n HANS\n\n Glad I caught you.\n\n ANNA\n\n Hans.\n\n He smoothly sets his drink down on a passing tray. He lifts\n\n her up and leads her in a romantic dance.\n\n DISSOLVE TO:\n\n LATER: Anna and Hans drink and chat.\n\n ANNA (CONT'D)\n\n I often had the whole parlor to\n\n myself to slide... Oops. Sorry.\n\n She hits him in the face by mistake with her hand. He laughs.\n\n DISSOLVE TO:\n\n -THE CASTLE DOORS: Anna and Hans stroll out of the castle.\n\n ANNA (CONT'D)\n\n ...Your physique helps I'm sure.\n\n DISSOLVE TO:\n\n -THE ROSE GARDEN... Hans notices her white streak.\n\n HANS\n\n (about her white streak)\n\n What's this?\n\n ANNA\n\n I was born with it, although I\n\n dreamt I was kissed by a troll.\n\n HANS\n\n I like it.\n\n DISSOLVE TO:\n\n 25\n\nFROZEN - J. Lee\n\n EXT. BALCONY -- NIGHT\n\n Anna teaches Hans how to eat krumkake.\n\n ANNA\n\n Yeah, the whole thing! You got it.\n\n They laugh as the krumkake crumbles in his face.\n\n ANNA(CONT'D)\n\n Okay wait, wait. So you have how\n\n many brothers?\n\n HANS\n\n Twelve older brothers. Three of\n\n them pretended I was invisible...\n\n literally...for two years.\n\n ANNA\n\n That's horrible.\n\n HANS\n\n It's what brothers do.\n\n ANNA\n\n ...And sisters. Elsa and I were\n\n really close when we were little.\n\n But then, one day she just shut me\n\n out, and I never knew why.\n\n He takes her hand. Leans in close.\n\n HANS\n\n I would never shut you out.\n\n ANNA\n\n Okay, can I just say something\n\n crazy?\n\n HANS\n\n I love crazy.\n\n "Love is an Open Door"\n\n ANNA\n\n (singing)\n\n ALL MY LIFE HAS BEEN A SERIES OF\n\n DOORS IN MY FACE.\n\n AND THEN SUDDENLY I BUMP INTO YOU.\n\n HANS\n\n I was thinking the same thing,\n\n because like. . .\n\n (MORE)\n\n 26\n\nFROZEN - J. Lee\n\n HANS (CONT'D)\n\n I'VE BEEN SEARCHING MY WHOLE LIFE\n\n TO FIND MY OWN PLACE.\n\n AND MAYBE IT'S THE PARTY TALKING,\n\n OR THE CHOCOLATE FONDUE.\n\n ANNA\n\n BUT WITH YOU-\n\n HANS\n\n BUT WITH YOU,\n\n I FOUND MY PLACE.\n\n ANNA\n\n I SEE YOUR FACE.\n\n BOTH\n\n AND IT'S NOTHING LIKE I'VE EVER\n\n KNOWN BEFORE.\n\n They jump to the neighboring balcony and enter a door.\n\n They come out on top of one of the castle's towers.\n\n BOTH (CONT'D)\n\n LOVE IS AN OPEN DOOR!\n\n LOVE IS AN OPEN DOOR!\n\n Cut to them sliding across an empty hallway in their socks.\n\n BOTH (CONT'D)\n\n LOVE IS AN OPEN DOOR\n\n ANNA\n\n WITH YOU!\n\n HANS\n\n WITH YOU!\n\n ANNA\n\n WITH YOU!\n\n HANS\n\n WITH YOU!\n\n BOTH\n\n LOVE IS AN OPEN DOOR.\n\n They hop up on the castle roof and watch a shooting star.\n\n HANS\n\n I MEAN IT'S CRAZY.\n\n ANNA\n\n What?\n\n 27\n\nFROZEN - J. Lee\n\n HANS\n\n WE FINISH EACH OTHER'S-\n\n ANNA\n\n SANDWICHES!\n\n HANS\n\n That's what I was gonna say!\n\n They slide down the back of the roof out of sight.\n\n We next find them strutting on a bridge ledge.\n\n ANNA\n\n I'VE NEVER MET SOMEONE-\n\n BOTH\n\n WHO THINKS SO MUCH LIKE ME.\n\n BOTH (SPOKEN) (CONT'D)\n\n Jinx.. . .jinx again.\n\n Are they doing the robot? No. They're imitating the\n\n mechanical figures on the clock tower.\n\n BOTH (CONT'D)\n\n OUR MENTAL SYNCHRONIZATION\n\n CAN HAVE BUT ONE EXPLANATION,\n\n HANS\n\n YOU-\n\n ANNA\n\n AND I-\n\n HANS\n\n WERE-\n\n ANNA\n\n JUST-\n\n BOTH\n\n MEANT TO BE.\n\n Anna and Hans dance on top of the lighthouse and cast dancing\n\n shadows across the sails of ships in the docks.\n\n ANNA\n\n SAY GOODBYE-\n\n HANS\n\n SAY GOODBYE-\n\n 28\n\nFROZEN - J. Lee\n\n BOTH\n\n TO THE PAIN OF THE PAST.\n\n BOTH (CONT'D)\n\n WE DON'T HAVE TO FEEL IT ANYMORE!\n\n LOVE IS AN OPEN-\n\n They play hide and seek amongst the stable doors.\n\n BOTH (CONT'D)\n\n DOOR! LOVE IS AN OPEN DOOR!\n\n They climb to the waterfall looking out over the kingdom.\n\n Anna raises up her hands to frame the moon. Hans puts his\n\n hands on top of hers. Together their hands form a heart.\n\n BOTH (CONT'D)\n\n LIFE CAN BE SO MUCH MORE-\n\n ANNA\n\n WITH YOU!\n\n HANS\n\n WITH YOU!\n\n ANNA\n\n WITH YOU!\n\n HANS\n\n WITH YOU!\n\n BOTH\n\n LOVE IS AN OPEN\n\n HANS\n\n DOOR.\n\n ANNA\n\n DOOR.\n\n HANS\n\n Can I say something crazy...? Will\n\n you marry me?\n\n ANNA\n\n Can I just say something even\n\n crazier? Yes.\n\n CUT TO:\n\n 29\n\nFROZEN - J. Lee\n\n INT. BALL -- NIGHT\n\n Anna pushes through the crowd towards Elsa, Hans in tow.\n\n ANNA\n\n Oops! Pardon. Sorry. Can we just\n\n get around you there? Thank you.\n\n Oh, there she is. Elsa!\n\n Elsa turns to Anna. Anna curtseys awkwardly.\n\n ANNA (CONT'D)\n\n I mean...Queen.... Me again. Um.\n\n May I present Prince Hans of the\n\n Southern Isles.\n\n HANS\n\n (bowing)\n\n Your Majesty.\n\n Elsa gives a polite but reserved curtsey.\n\n ANNA\n\n We would like--\n\n HANS\n\n --your blessing--\n\n ANNA\n\n --of--\n\n ANNA/HANS\n\n --our marriage!\n\n ELSA\n\n Marriage...?\n\n ANNA\n\n Yes!\n\n ELSA\n\n I'm sorry, I'm confused.\n\n ANNA\n\n Well, we haven't worked out all the\n\n details ourselves. We'll need a few\n\n days to plan the ceremony. Of\n\n course we'll have soup, roast, and\n\n ice cream and then--\n\n Wait. Would we live here?\n\n ELSA\n\n Here?\n\n 30\n\nFROZEN - J. Lee\n\n HANS\n\n Absolutely!\n\n ELSA\n\n Anna--\n\n ANNA\n\n Oh, we can invite all twelve of\n\n your brothers to stay with us--\n\n ELSA\n\n What? No, no, no, no, no.\n\n ANNA\n\n Of course we have the room. I don't\n\n know. Some of them must--\n\n ELSA\n\n Wait. Slow down. No one's brothers\n\n are staying here. No one is getting\n\n married.\n\n ANNA\n\n Wait, what?\n\n ELSA\n\n May I talk to you, please. Alone.\n\n Anna sees Hans's worried face. Hooks arms with him.\n\n ANNA\n\n No. Whatever you have to say, you-\n\n you can say to both of us.\n\n ELSA\n\n Fine. You can't marry a man you\n\n just met.\n\n ANNA\n\n You can if it's true love.\n\n ELSA\n\n Anna, what do you know about true\n\n love?\n\n ANNA\n\n More than you. All you know is how\n\n to shut people out.\n\n ELSA\n\n You asked for my blessing, but my\n\n answer is no. Now, excuse me.\n\n 31\n\nFROZEN - J. Lee\n\n HANS\n\n Your Majesty, if I may ease your--\n\n ELSA\n\n (flustered)\n\n No, you may not. And I-I think you\n\n should go.\n\n Elsa walks away. As she passes the Royal Handler--\n\n ELSA (CONT'D)\n\n The party is over. Close the gates.\n\n ANNA\n\n What? Elsa, no. No, wait!\n\n Anna grabs Elsa's hand. She pulls off Elsa's glove. Elsa\n\n gasps, spins around and reaches for the glove in panic.\n\n ELSA\n\n Give me my glove!\n\n Anna holds the glove away from Elsa.\n\n ANNA\n\n (desperate)\n\n Elsa, please. Please. I can't live\n\n like this anymore.\n\n Elsa fights tears.\n\n ELSA\n\n (weak)\n\n ...Then leave.\n\n Elsa sees Anna's hurt face. It's too much. She can't hold it\n\n in. She turns and rushes away.\n\n ANNA\n\n (heartbroken)\n\n ...What did I ever do to you?!\n\n The party goes silent as everyone watches the sisters.\n\n ELSA\n\n Enough, Anna.\n\n ANNA\n\n No. Why? Why do you shut me out?!\n\n Why do you shut the world out?!\n\n What are you so afraid of?!\n\n ELSA\n\n I said, enough!\n\n 32\n\nFROZEN - J. Lee\n\n Ice shoots from Elsa's hand, spikes across the floor! Guests\n\n cry out in shock, back away.\n\n DUKE\n\n (ducking behind his men)\n\n ...Sorcery. I knew there was\n\n something dubious going on here.\n\n ANNA\n\n Elsa...?\n\n Elsa rushes out of the room.\n\n CUT TO:\n\n EXT. COURTYARD -- NIGHT\n\n Elsa bursts out of the castle door. The CITIZENS CHEER!\n\n CROWD\n\n There she is. Your Majesty! Long\n\n live the Queen! Queen Elsa.... Come\n\n drink with us.\n\n Elsa ducks through the crowd, holding her bare hand.\n\n BOWING TOWNSMAN\n\n Queen Elsa.\n\n TOWNSWOMAN WITH BABY\n\n Your Majesty? Are you all right?\n\n Elsa backs away from the baby. She knocks into the fountain,\n\n grabs its edge. The waters freeze at her touch.\n\n GASPS of shock and fear sweep over the crowd.\n\n The Duke and thugs come out the door.\n\n DUKE\n\n There she is! Stop her!\n\n ELSA\n\n (to the Duke)\n\n Please, just stay away from me.\n\n Stay away!\n\n Magic accidentally shoots from her hand and turns the\n\n staircase into ice. The thugs and the Duke fall.\n\n DUKE\n\n Monster.... Monster!\n\n 33\n\nFROZEN - J. Lee\n\n The crowd panics.\n\n A snowstorm begins. Elsa flees.\n\n Anna runs out of the palace doors, carrying the glove.\n\n ANNA\n\n Elsa!\n\n Hans follows closely behind her.\n\n GATES TO THE KINGDOM: Elsa runs out of the gates and down to\n\n the water's edge. The shoreline freezes under her feet.\n\n Anna calls to her from the gates.\n\n ANNA (CONT'D)\n\n Elsa! Wait, please!\n\n Elsa glances back at Anna, but turns away. She tentatively\n\n steps out onto the fjord. It freezes instantly. She breaks\n\n into a run, as the water freezes over with each step.\n\n ANNA (CONT'D)\n\n Elsa, stop!\n\n Anna rushes out onto the fjord ice, slips, falls.\n\n HANS\n\n Anna!\n\n Hans rushes to Anna's side.\n\n Elsa reaches the far shore. She doesn't look back. She just\n\n scrambles into the mountains.\n\n ANNA\n\n No.\n\n HANS\n\n (shocked)\n\n Look.... The fjord.\n\n The ice spreads out until the entire fjord is frozen, locking\n\n the ships in place.\n\n INT. CASTLE COURTYARD -- NIGHT\n\n Snow falls. Hans and Anna move through the panicking crowd.\n\n CROWD WALLAH\n\n Snow? It's...snow...in July.\n\n 34\n\nFROZEN - J. Lee\n\n HANS\n\n ...Are you all right?\n\n ANNA\n\n (in shock)\n\n No.\n\n HANS\n\n Did you know?\n\n ANNA\n\n No.\n\n Nearby, the Duke flutters about in fright.\n\n DUKE\n\n Look! It's snowing! It's snowing!\n\n The Queen has cursed this land! She\n\n must be stopped!\n\n (to his thugs)\n\n You have to go after her.\n\n Anna rushes up to the Duke.\n\n ANNA\n\n Wait, no!\n\n The Duke hides behind his thugs and points out at Anna.\n\n DUKE\n\n You! Is there sorcery in you, too?\n\n Are you a monster, too?\n\n ANNA\n\n No. No. I'm completely ordinary.\n\n HANS\n\n That's right she is...\n\n (realizing how that\n\n sounds)\n\n ...in the best way.\n\n ANNA\n\n ...And my sister's not a monster.\n\n DUKE\n\n She nearly killed me.\n\n HANS\n\n You slipped on ice.\n\n DUKE\n\n Her ice!\n\n 35\n\nFROZEN - J. Lee\n\n ANNA\n\n It was an accident. She was scared.\n\n She didn't mean it. She didn't mean\n\n any of this.... Tonight was my\n\n fault. I pushed her. So I'm the one\n\n that needs to go after her.\n\n DUKE\n\n Yes. Fine. Do.\n\n HANS\n\n What?\n\n ANNA\n\n (to the Royal Handler)\n\n Bring me my horse, please.\n\n HANS\n\n Anna, no. It's too dangerous.\n\n ANNA\n\n Elsa's not dangerous. I'll bring\n\n her back, and I'll make this right.\n\n The Royal Handler brings Anna her horse and a cloak.\n\n HANS\n\n I'm coming with you.\n\n ANNA\n\n No, I need you here to take care of\n\n Arendelle.\n\n He sees the desperation in her eyes.\n\n HANS\n\n ...On my honor.\n\n She throws on the cloak and hops right onto the horse,\n\n coronation dress and all.\n\n ANNA\n\n (to the crowd)\n\n I leave Prince Hans in charge!\n\n HANS\n\n (before letting her go)\n\n Are you sure you can trust her? I\n\n don't want you getting hurt.\n\n ANNA\n\n She's my sister; she would never\n\n hurt me.\n\n 36\n\nFROZEN - J. Lee\n\n She snaps the reins and rides out. Hans watches after her.\n\n The snow picks up and overtakes our view. We push through a\n\n blizzard...lose our way...then finds ourselves...\n\n EXT. HIGH UP IN THE MOUNTAINS -- NIGHT\n\n Well above the snow-line, a small figure climbs the highest\n\n peak. It's Elsa. Finally, she stops, looks around. Catches\n\n her breath and sings...\n\n "Let It Go"\n\n ELSA\n\n THE SNOW GLOWS WHITE\n\n ON THE MOUNTAIN TONIGHT,\n\n NOT A FOOTPRINT TO BE SEEN.\n\n A KINGDOM OF ISOLATION\n\n AND IT LOOKS LIKE I'M THE QUEEN.\n\n THE WIND IS HOWLING\n\n LIKE THIS SWIRLING STORM INSIDE.\n\n COULDN'T KEEP IT IN,\n\n HEAVEN KNOWS I TRIED. . .\n\n DON'T LET THEM IN,\n\n DON'T LET THEM SEE,\n\n BE THE GOOD GIRL YOU ALWAYS HAVE TO\n\n BE.\n\n CONCEAL,\n\n DON'T FEEL,\n\n DON'T LET THEM KNOW.\n\n WELL, NOW THEY KNOW.\n\n Elsa takes off her glove and throws it into the air.\n\n ELSA (CONT'D)\n\n LET IT GO. LET IT GO.\n\n CAN'T HOLD IT BACK ANYMORE.\n\n Elsa creates a snowman, just like the one she made with Anna\n\n when they were children.\n\n ELSA (CONT'D)\n\n LET IT GO. LET IT GO.\n\n TURN AWAY AND SLAM THE DOOR.\n\n I DON'T CARE WHAT THEY'RE GOING TO\n\n SAY.\n\n LET THE STORM RAGE ON.\n\n THE COLD NEVER BOTHERED ME ANYWAY.\n\n Elsa lets her cape fly back into the wind.\n\n 37\n\nFROZEN - J. Lee\n\n ELSA (CONT'D)\n\n IT'S FUNNY HOW SOME DISTANCE\n\n MAKES EVERYTHING SEEM SMALL.\n\n AND THE FEARS THAT ONCE CONTROLLED ME\n\n CAN'T GET TO ME AT ALL.\n\n IT'S TIME TO SEE\n\n WHAT I CAN DO,\n\n TO TEST THE LIMITS AND BREAK THROUGH.\n\n NO RIGHT, NO WRONG,\n\n NO RULES FOR ME...I'M FREE!\n\n Elsa creates ice steps and climbs them.\n\n ELSA (CONT'D)\n\n LET IT GO! LET IT GO!\n\n I AM ONE WITH THE WIND AND SKY.\n\n LET IT GO! LET IT GO!\n\n YOU'LL NEVER SEE ME CRY.\n\n HERE I STAND AND HERE I'LL STAY.\n\n Elsa slams her foot down and forms a giant snowflake.\n\n ELSA (CONT'D)\n\n LET THE STORM RAGE ON....\n\n In a flurry of creative release, she raises the snowflake on\n\n ice beams, builds walls, archways, a glistening chandelier,\n\n and an intricate ceiling that leaves the sky visible.\n\n ELSA (CONT'D)\n\n MY POWER FLURRIES THROUGH THE AIR\n\n INTO THE GROUND.\n\n MY SOUL IS SPIRALING IN FROZEN\n\n FRACTALS ALL AROUND.\n\n AND ONE THOUGHT CRYSTALLIZES LIKE\n\n AN ICY BLAST-\n\n Standing firmly in her mighty ice palace, Elsa removes her\n\n crown and throws it.\n\n ELSA (CONT'D)\n\n I'M NEVER GOING BACK,\n\n (back to resolve)\n\n THE PAST IS IN THE PAST!\n\n She takes down her hair and creates a new dress made of ice.\n\n ELSA (CONT'D)\n\n LET IT GO! LET IT GO!\n\n AND I'LL RISE LIKE THE BREAK OF DAWN.\n\n LET IT GO! LET IT GO!\n\n The sun rises. Elsa struts onto out onto a balcony and into\n\n the light. She's free.\n\n 38\n\nFROZEN - J. Lee\n\n ELSA (CONT'D)\n\n THAT PERFECT GIRL IS GONE.\n\n HERE I STAND IN THE LIGHT OF DAY.\n\n LET THE STORM RAGE ON!!\n\n THE COLD NEVER BOTHERED ME ANYWAY.\n\n She turns and slams her ice palace door on us.\n\n CUT TO:\n\n EXT. THE FJORD FOREST -- DAY\n\n Anna rides her horse through two feet of snow. She shivers.\n\n ANNA\n\n (shivering)\n\n Elsa! Elsa! It's me, Anna...your\n\n sister who didn't mean to make you\n\n freeze the summer. I'm sorry. It's\n\n all my f-f-f-f-f-f-fault.\n\n DISSOLVE TO:\n\n LATER: Anna and the horse struggle through a wooded area.\n\n ANNA (CONT'D)\n\n (hearing a wolf howl)\n\n Of course, none of this would have\n\n happened if she'd just told me her\n\n secret...ha...she's a stinker.\n\n A branch of a nearby tree snaps and startles the horse. Anna\n\n goes flying off, lands face down in the snow. She sits up.\n\n Spits out snow. Sees the horse running away.\n\n ANNA (CONT'D)\n\n Oh no. No. No. No. Come back. No.\n\n No. No. No.... Oooo-kay.\n\n He doesn't come back. Anna grabs onto a branch of a leaning\n\n conifer, tries to pull herself to her feet, but the tree\n\n snaps upright and releases all its snow onto her. GROAN.\n\n DISSOLVE TO:\n\n EXT. MOUNTAIN -- NIGHT\n\n The Northern Lights shine as Anna struggles, out of breath,\n\n reaching the top of a hill.\n\n 39\n\nFROZEN - J. Lee\n\n ANNA\n\n Snow, it had to be snow, she\n\n couldn't have had tr-tr-tropical\n\n magic that covered the f-f-fjords\n\n in white sand and warm --\n\n She sees smoke rising up in the distance.\n\n ANNA (CONT'D)\n\n Fire! WHOA!\n\n Anna goes tumbling down the hill. She lands with a crash in\n\n an icy stream at the bottom.\n\n ANNA (CONT'D)\n\n (from inside the snowball)\n\n Cold, cold, cold, cold, cold...\n\n EXT. A SMALL BUILDING AND STABLE -- NIGHT\n\n Anna shuffles up to the building, her dress frozen stiff. She\n\n shakes the snow off a sign and reads:\n\n ANNA\n\n Wandering Oaken's Trading Post.\n\n Snow drops off a smaller sign. She reads it, happily.\n\n ANNA (CONT'D)\n\n Ooh! And Sauna...\n\n INT. WANDERING OAKEN'S TRADING POST & SAUNA -- NIGHT\n\n Anna steps cautiously through the door--which hits her frozen\n\n butt and knocks her into the center of the shop. She looks\n\n around, sees only summer supplies.\n\n OAKEN (O.S.)\n\n Hoo hoo.\n\n Anna turns to see a bright-faced fellow sitting low behind\n\n the counter, fingers tapping tip to tip.\n\n OAKEN (CONT'D)\n\n Big summer blow out. Half off\n\n swimming suits, clogs, and a sun\n\n balm of my own invention, yah?\n\n ANNA\n\n Oh, great. For now, how about\n\n boots. Winter boots...and dresses?\n\n 40\n\nFROZEN - J. Lee\n\n OAKEN\n\n (slight disappointment)\n\n That would be in our winter\n\n department.\n\n The winter department contains one outfit, a pick ax, and a\n\n lonely pair of boots.\n\n ANNA\n\n Oh. Um, I was just wondering; has\n\n another young woman, the Queen\n\n perhaps, I don't know, passed\n\n through here?\n\n She brings the clothes and boots to the counter.\n\n OAKEN\n\n Only one crazy enough to be out in\n\n this storm is you, dear?\n\n The front door suddenly blows open and in walks a mass of a\n\n man covered in ice. Underneath is KRISTOFF.\n\n OAKEN (CONT'D)\n\n You and this fellow.... Hoo hoo.\n\n Big summer blow out.\n\n Kristoff walks right up to Anna.\n\n KRISTOFF\n\n (in her face)\n\n Carrots.\n\n ANNA\n\n Huh?\n\n KRISTOFF\n\n Behind you.\n\n ANNA\n\n Oh, right. Excuse me.\n\n Anna moves out of Kristoff's way. He grabs a bunch of\n\n carrots, tosses them on the counter, then moves through the\n\n place, gathering other supplies.\n\n OAKEN\n\n (to Kristoff)\n\n A real howler in July, yah? Where\n\n ever could it be coming from?\n\n KRISTOFF\n\n The North Mountain.\n\n 41\n\nFROZEN - J. Lee\n\n ANNA\n\n (to herself)\n\n North Mountain.\n\n Kristoff brings his supplies to the counter. Oaken counts on\n\n his fingertips.\n\n OAKEN\n\n That'll be forty.\n\n KRISTOFF\n\n Forty? No, ten.\n\n OAKEN\n\n (sweet as pie)\n\n Oh dear, that's no good. See these\n\n are from our winter stock, where\n\n supply and demand have a big\n\n problem.\n\n KRISTOFF\n\n You want to talk about a supply and\n\n demand problem? I sell ice for a\n\n living.\n\n Kristoff motions out the window, where we see the blocks of\n\n ice on his sled, covered in snow.\n\n ANNA\n\n Ooh, that's a rough business to be\n\n in right now. I mean, that is\n\n really...\n\n (he shoots her a look)\n\n Ahem. That's unfortunate.\n\n OAKEN\n\n Still forty. But I will throw in a\n\n visit to Oaken's sauna. Hoo hoo!\n\n Hi, family.\n\n Kristoff and Anna turn to see a naked family waving through\n\n the window of the steaming sauna.\n\n NAKED FAMILY\n\n Hoo hoo!\n\n KRISTOFF\n\n ...Ten's all I got. Help me out.\n\n OAKEN\n\n (isolating the carrots)\n\n Ten will get you this and no more.\n\n Kristoff seethes. Stalemate.\n\n 42\n\nFROZEN - J. Lee\n\n ANNA\n\n Okay, just tell me one thing; what\n\n was happening on the North\n\n Mountain? Did it seem magical?\n\n Kristoff pulls down his scarf and gives Anna a firm answer.\n\n KRISTOFF\n\n Yes! Now, back up while I deal with\n\n this crook here.\n\n Oaken stands up, revealing his seven-foot stature.\n\n OAKEN\n\n What did you call me?\n\n EXT. WANDERING OAKEN'S TRADING POST AND SAUNA -- NIGHT\n\n Oaken stomps out the door, carrying Kristoff with one arm.\n\n KRISTOFF\n\n Okay. Okay, I'm- Ow! Whoa!\n\n Oaken throws Kristoff, who face-plants in the snow.\n\n OAKEN\n\n Bye bye.\n\n Oaken slams the door. Kristoff sits up. His reindeer, Sven,\n\n canters over, snorts, and nudges him, expectantly.\n\n KRISTOFF\n\n No Sven, I didn't get your carrots.\n\n Sven huffs in his face. Kristoff turns away and sees\n\n something. He points to a dilapidated barn.\n\n KRISTOFF (CONT'D)\n\n But I did find us a place to sleep.\n\n And it's free.\n\n INT. WANDERING OAKEN'S TRADING POST AND SAUNA -- NIGHT\n\n Anna stands watching Oaken and all his great height as he\n\n squeezes behind the counter and sits down low again.\n\n OAKEN\n\n (teddy bear)\n\n I'm sorry about this violence. I\n\n will add a quart of lutefisk, so\n\n we'll have good feelings. Just the\n\n outfit and boots, yah?\n\n 43\n\nFROZEN - J. Lee\n\n Anna looks between Kristoff's supplies and the door.\n\n CUT TO:\n\n INT. OAKEN'S STABLES - NIGHT\n\n Kristoff, now unfrozen, relaxes on a bed of hay, playing his\n\n lute and singing to (and for) Sven.\n\n "Reindeer(s) are Better than People"\n\n KRISTOFF\n\n REINDEERS ARE BETTER THAN PEOPLE.\n\n SVEN, DON'T YOU THINK THAT'S TRUE?\n\n KRISTOFF (AS SVEN) (CONT'D)\n\n (throwing his voice)\n\n YEAH, PEOPLE WILL BEAT YOU & CURSE\n\n YOU & CHEAT YOU.\n\n EVERY ONE OF EM'S BAD, EXCEPT YOU.\n\n (speaking)\n\n Oh, thanks, Buddy.\n\n (singing, as Kristoff)\n\n BUT PEOPLE SMELL BETTER THAN\n\n REINDEERS.\n\n SVEN, DON'T YOU THINK I'M RIGHT?\n\n (As Sven)\n\n THAT'S ONCE AGAIN TRUE,\n\n FOR ALL EXCEPT YOU.\n\n (As Kristoff)\n\n YOU GOT ME. LET'S CALL IT A NIGHT.\n\n (As Sven)\n\n GOOD NIGHT.\n\n (As Kristoff)\n\n DON'T LET THE FROSTBITE BITE.\n\n The door opens. Anna enters.\n\n ANNA\n\n Nice duet.\n\n Kristoff sits up with a start...sees who it is.\n\n KRISTOFF\n\n Oh, it's just you. What do you\n\n want?\n\n ANNA\n\n I want you to take me up the North\n\n Mountain.\n\n 44\n\nFROZEN - J. Lee\n\n KRISTOFF\n\n I don't take people places.\n\n He lays back down, closes his eyes.\n\n ANNA\n\n Let me rephrase that...\n\n A sack of supplies lands in Kristoff's lap.\n\n KRISTOFF\n\n Umph.\n\n He sits up. Looks in the bag.\n\n ANNA\n\n Take me up the North Mountain....\n\n Please.\n\n He eyes her. He clearly doesn't take orders.\n\n ANNA (CONT'D)\n\n Look, I know how to stop this\n\n winter.\n\n He considers, lies back down, pulls his hat over his eyes.\n\n KRISTOFF\n\n We leave at dawn.... And you forgot\n\n the carrots for Sven.\n\n A bag of carrots hits Kristoff in the face.\n\n KRISTOFF (CONT'D)\n\n Ugh!\n\n ANNA\n\n Oops. Sorry. Sorry. I'm sorry. I\n\n didn't--\n\n (catching herself)\n\n We leave now. Right now.\n\n She steps back outside and waits, anxiously. Annoyed,\n\n Kristoff offers Sven a carrot. Sven has a bite. Then Kristoff\n\n has a bite, contemplating.\n\n SLAM CUT TO:\n\n EXT. MOUNTAIN HIGH -- NIGHT\n\n Sven races, top speed, up a narrow cliff, pulling the sled,\n\n which skids precariously. Kristoff mans the reins. Anna sits\n\n beside him.\n\n 45\n\nFROZEN - J. Lee\n\n KRISTOFF\n\n (trying to scare Anna)\n\n Hang on! We like to go fast!\n\n ANNA\n\n (fearless)\n\n I like fast!\n\n Anna leans back and puts her feet up on the dashboard.\n\n KRISTOFF\n\n Whoa, whoa! Get your feet down.\n\n He pushes her feet down.\n\n KRISTOFF (CONT'D)\n\n This is fresh lacquer. Seriously,\n\n were you raised in a barn?\n\n Kristoff spits on the dash to clean it. The spit flies back\n\n and hits Anna in the face.\n\n ANNA\n\n (grossed out)\n\n Ew. No, I was raised in a castle.\n\n She wipes off her face.\n\n KRISTOFF\n\n So tell me, what made the Queen go\n\n all ice-crazy?\n\n ANNA\n\n ...Oh well, it was all my fault. I\n\n got engaged but then she freaked\n\n out because I'd only just met him,\n\n you know, that day. And she said\n\n she wouldn't bless the marriage--\n\n KRISTOFF\n\n Wait. You got engaged to someone\n\n you just met?\n\n ANNA\n\n Yeah. Anyway, I got mad and so she\n\n got mad and then she tried to walk\n\n away, and I grabbed her glove--\n\n KRISTOFF\n\n Hang on. You mean to tell me you\n\n got engaged to someone you just\n\n met?!\n\n 46\n\nFROZEN - J. Lee\n\n ANNA\n\n Yes. Pay attention. But the thing\n\n is she wore the gloves all the\n\n time, so I just thought, maybe she\n\n has a thing about dirt.\n\n KRISTOFF\n\n Didn't your parents ever warn you\n\n about strangers?\n\n Anna eyes Kristoff up and down, then slides away from him.\n\n ANNA\n\n Yes, they did.... But Hans is not a\n\n stranger.\n\n KRISTOFF\n\n Oh yeah? What's his last name?\n\n ANNA\n\n ...Of-the-Southern-Isles?\n\n KRISTOFF\n\n What's his favorite food?\n\n ANNA\n\n ...Sandwiches.\n\n KRISTOFF\n\n Best friend's name?\n\n ANNA\n\n Probably John.\n\n KRISTOFF\n\n Eye color.\n\n ANNA\n\n Dreamy.\n\n KRISTOFF\n\n Foot size...?\n\n ANNA\n\n Foot size doesn't matter.\n\n KRISTOFF\n\n Have you had a meal with him yet?\n\n What if you hate the way he eats?\n\n What if you hate the way he picks\n\n his nose?\n\n ANNA\n\n Picks his nose?\n\n 47\n\nFROZEN - J. Lee\n\n KRISTOFF\n\n And eats it.\n\n ANNA\n\n Excuse me, sir. He's a prince.\n\n KRISTOFF\n\n All men do it.\n\n ANNA\n\n Ew. Look it doesn't matter; it's\n\n true love.\n\n KRISTOFF\n\n Doesn't sound like true love.\n\n ANNA\n\n Are you some sort of love expert?\n\n KRISTOFF\n\n No. But I have friends who are.\n\n ANNA\n\n You have friends who are love\n\n experts.... I'm not buying it.\n\n Sven suddenly stops, ears perked in alarm.\n\n KRISTOFF\n\n (to Anna)\n\n Stop talking.\n\n ANNA\n\n No, no, no. I'd like to meet these--\n\n Kristoff clamps his hand over Anna's mouth.\n\n KRISTOFF\n\n I mean it. SHHH.\n\n Kristoff stands, looks into the dark woods surrounding them.\n\n Sensing something behind them, he holds up his lantern. Its\n\n light reflects off...EYES. Several.\n\n KRISTOFF(CONT'D)\n\n Sven, go. Go!\n\n Sven takes off.\n\n ANNA\n\n What are they?\n\n KRISTOFF\n\n Wolves.\n\n 48\n\nFROZEN - J. Lee\n\n Flashes of white dart through the woods. Kristoff hops into\n\n the back of the sled, grabs a torch. Lights it.\n\n ANNA\n\n Wolves. What do we do?\n\n KRISTOFF\n\n I've got this. You just...don't\n\n fall off and don't get eaten.\n\n ANNA\n\n But I wanna help.\n\n KRISTOFF\n\n No.\n\n ANNA\n\n Why not?\n\n KRISTOFF\n\n Because I don't trust your\n\n judgement.\n\n ANNA\n\n Excuse me?!\n\n A wolf jumps at them, but Kristoff kicks it off.\n\n KRISTOFF\n\n Who marries a man she just met?\n\n Anna grabs the lute, swings it right at Kristoff's head.\n\n ANNA\n\n It's true love!\n\n He screams, as she...BAM!...swings past Kristoff and knocks a\n\n wolf away.\n\n KRISTOFF\n\n (shocked)\n\n Whoa.\n\n Just then Kristoff is yanked off the sled by another wolf.\n\n The torch goes flying. Anna catches it, shocked.\n\n ANNA\n\n Christopher!\n\n Kristoff grabs onto a loose rope hanging from the back of the\n\n sled and holds on for dear life as he's dragged behind.\n\n KRISTOFF\n\n It's Kristoff!\n\n 49\n\nFROZEN - J. Lee\n\n A wolf jumps on Kristoff's back.\n\n KRISTOFF (CONT'D)\n\n AH!\n\n Anna thinks fast, uses the torch to light a blanket on fire.\n\n ANNA\n\n Duck!\n\n Anna throws the flaming blanket right at him. He ducks. The\n\n blanket hits the wolves. They tumble off Kristoff.\n\n KRISTOFF\n\n You almost set me on fire!\n\n Anna reaches out a hand, pulls Kristoff back onto the sled.\n\n ANNA\n\n But I didn't.\n\n Sven cries out. There is a massive gorge ahead.\n\n ANNA (CONT'D)\n\n Get ready to jump, Sven!\n\n KRISTOFF\n\n You don't tell him what to do!\n\n Kristoff shoves a satchel into her arms then scoops her up.\n\n KRISTOFF (CONT'D)\n\n I do!\n\n Kristoff tosses Anna onto Sven, then unhooks Sven's harness\n\n from the sled.\n\n KRISTOFF (CONT'D)\n\n Jump, Sven!\n\n Sven jumps the gorge with Anna on his back.\n\n Kristoff goes flying off behind them, still on the sled.\n\n Anna and Sven land safely on the other side of the gorge.\n\n Kristoff's sled loses momentum. It's not going to make it. He\n\n leaps off. He flaps his arms, claws at the air.\n\n He slams into the snowy edge of the cliff. Hanging by his\n\n hands, he looks down to see his sled hit the ground far below\n\n and burst into flames.\n\n 50\n\nFROZEN - J. Lee\n\n KRISTOFF (CONT'D)\n\n (shocked sadness)\n\n ...But I just paid it off.\n\n Suddenly, he starts to slip. He claws at the loose snow, but\n\n it's clearly hopeless. He's going down.\n\n KRISTOFF (CONT'D)\n\n Uh-oh. No, no, no.\n\n To make matters worse, an AXE comes flying right at his face.\n\n KRISTOFF (CONT'D)\n\n AH! NO, NO, NO!\n\n The axe slams into the snow, inches from his nose.\n\n ANNA (O.S.)\n\n Grab on!\n\n Kristoff grabs on.\n\n ANNA (CONT'D)\n\n Pull, Sven! Pull!\n\n REVEAL: The axe is tied to a rope, then wrapped around Sven.\n\n Anna helps Sven pull Kristoff to safety.\n\n Kristoff rolls onto his back, exhausted. Anna peeks down at\n\n the burning sled.\n\n ANNA (CONT'D)\n\n Whoa.... I'll replace your sled and\n\n everything in it.\n\n Kristoff groans.\n\n ANNA (CONT'D)\n\n And I understand if you don't want\n\n to help me anymore.\n\n Anna walks off, sadly. Sven comes over and nuzzles Kristoff.\n\n KRISTOFF\n\n Of course I don't want to help her\n\n anymore. In fact, this whole thing\n\n has ruined me for helping anyone\n\n ever again.\n\n KRISTOFF (AS SVEN) (CONT'D)\n\n But she'll die on her own.\n\n KRISTOFF (AS SELF) (CONT'D)\n\n I can live with that.\n\n 51\n\nFROZEN - J. Lee\n\n Through their conversation, they watch Anna go the wrong\n\n way...turn, go the other wrong way, turn, trip...\n\n KRISTOFF (AS SVEN) (CONT'D)\n\n But you won't get your new sled if\n\n she's dead.\n\n KRISTOFF (CONT'D)\n\n (knowing he's got a point)\n\n ...You know sometimes I really\n\n don't like you.\n\n Sven licks Kristoff happily.\n\n KRISTOFF (AS SELF) (CONT'D)\n\n (to Anna)\n\n Hold up. We're coming?!\n\n ANNA\n\n (excited)\n\n You are?!\n\n (catching herself)\n\n I mean, sure. I'll let you tag\n\n along.\n\n DISSOLVE TO:\n\n EXT. SHARP MOUNTAIN RIDGE -- DAWN\n\n Kristoff, Sven and Anna walk on a narrow rim of a mountain.\n\n DISSOLVE TO:\n\n EXT. MOUNTAIN FOREST CLEARING -- DAY\n\n As they step out of the thick trees, Anna catches sight of\n\n something far below.\n\n ANNA\n\n Arendelle.\n\n KRISTOFF\n\n It's completely frozen.\n\n ANNA\n\n ...But it'll be fine. Elsa will\n\n thaw it.\n\n KRISTOFF\n\n Will she?\n\n 52\n\nFROZEN - J. Lee\n\n ANNA\n\n (uncertain)\n\n ...Yeah. Now come on. This way to\n\n the North Mountain?\n\n She points straight ahead.\n\n KRISTOFF\n\n More like this way.\n\n He points her finger up towards a perilously mighty mountain.\n\n DISSOLVE TO:\n\n INT. FROZEN WILLOW TREES -- DAY\n\n Anna, Kristoff, and Sven walk beneath frozen willows. The\n\n hanging branches glisten like Christmas lights. Sven knocks\n\n them with his antlers. They tinkle like chimes.\n\n ANNA\n\n I never knew winter could be so\n\n beautiful.\n\n Suddenly, a voice comes in from nowhere. We'll call that\n\n voice OLAF.\n\n OLAF (O.S.)\n\n YEAH...It really is beautiful,\n\n isn't it? But it's so white. You\n\n know, how about a little color?\n\n Must we bleach the joy out of it\n\n all? I'm thinking like maybe some\n\n crimson, chartreuse...\n\n While this is going on, Anna and Kristoff look around for the\n\n source of the rambling. They look at Sven - could he actually\n\n be talking? Sven looks back at them, his antlers tangled in\n\n branches, just as baffled as they are.\n\n In the meantime, a nose-less snowman, Olaf, wanders up behind\n\n them.\n\n OLAF (CONT'D)\n\n How `bout yellow--no, not yellow.\n\n Yellow and snow? Brrrr...no go.\n\n He stops between Kristoff and Anna. They look down at him.\n\n How did he get there? He suddenly looks up at Anna.\n\n OLAF (CONT'D)\n\n Am I right?\n\n 53\n\nFROZEN - J. Lee\n\n Anna SCREAMS! Reflexes take over and she kicks Olaf's head,\n\n sending it flying off his body and into Kristoff's arms.\n\n OLAF (CONT'D)\n\n (cheery, to Kristoff)\n\n Hi!\n\n KRISTOFF\n\n You're creepy.\n\n Kristoff tosses the head back to Anna and they commence a\n\n game of hot potato.\n\n ANNA\n\n I don't want it!\n\n KRISTOFF\n\n Backatchya!\n\n OLAF\n\n Please don't drop me.\n\n ANNA\n\n Don't!\n\n KRISTOFF\n\n Come on, it's just a head.\n\n ANNA\n\n No!\n\n Olaf's body runs at Anna, arms waving.\n\n OLAF (O.S.)\n\n All right, we got off to a bad\n\n start.\n\n ANNA\n\n Ew, ew, the body!\n\n Anna slams Olaf's head back on the body, upside down. Olaf\n\n smiles happily, then looks confused.\n\n OLAF\n\n Wait, what am I looking at right\n\n now? Why are you hanging off the\n\n earth like a bat?\n\n ANNA\n\n (sympathetic)\n\n ...Okay. Wait one second.\n\n Anna kneels in front of Olaf and rights his head.\n\n 54\n\nFROZEN - J. Lee\n\n OLAF\n\n Oooh! Thank you!\n\n ANNA\n\n You're welcome.\n\n OLAF\n\n Now I'm perfect.\n\n She looks over his innocent face, gets an idea.\n\n ANNA\n\n Well, almost.\n\n She digs into Kristoff's satchel, holds up a carrot just as\n\n Olaf turns toward her. The carrot accidentally slams all the\n\n way through his head.\n\n OLAF\n\n Woo! Head rush!\n\n ANNA\n\n Oh! Too hard. I'm sorry! I-I, I was\n\n just.... Are you okay?\n\n Olaf sees a tiny piece of carrot sticking out between his\n\n eyes. He lights up.\n\n OLAF\n\n Are you kidding me? I am wonderful!\n\n I've always wanted a nose.\n\n (going cross-eyed to look\n\n at his tiny nose)\n\n So cute. It's like a little baby\n\n unicorn.\n\n Anna reaches behind Olaf to the bulk of the carrot sticking\n\n out the back of his head, and pushes it forward.\n\n OLAF (CONT'D)\n\n What? Hey! Whoa.\n\n (seeing his now big nose)\n\n Oh, I love it even more! Hah....\n\n All right, let's start this thing\n\n over. Hi everyone. I'm Olaf. And I\n\n like warm hugs.\n\n Olaf opens his arms wide to Anna. That triggers a memory. It\n\n takes her a moment to place it, but then she does.\n\n ANNA\n\n Olaf?...That's right, Olaf.\n\n 55\n\nFROZEN - J. Lee\n\n OLAF\n\n ...And you are?\n\n ANNA\n\n Oh, um...I'm Anna.\n\n OLAF\n\n And who's the funky-looking donkey\n\n over there?\n\n ANNA\n\n That's Sven.\n\n OLAF\n\n Uh-huh. And who's the reindeer?\n\n ANNA\n\n ...Sven.\n\n Olaf looks from Kristoff to Sven, confused.\n\n OLAF\n\n Oh. They're--oh, okay....\n\n (accepting it)\n\n Makes things easier for me.\n\n Sven tries to bite Olaf's nose.\n\n OLAF (CONT'D)\n\n Ha. Aw, look at him tryin' to kiss\n\n my nose.\n\n (gushes)\n\n I like you, too!\n\n ANNA\n\n Olaf, did Elsa build you?\n\n OLAF\n\n Yeah. Why?\n\n Curious, Kristoff takes one of Olaf's twig arms off, studies\n\n it. It seems to be moving in sync with his other arm.\n\n ANNA\n\n Do you know where she is?\n\n KRISTOFF\n\n (studying the arm)\n\n Fascinating...\n\n OLAF\n\n Yeah. Why?\n\n 56\n\nFROZEN - J. Lee\n\n ANNA\n\n Do you think you could show us the\n\n way?\n\n OLAF\n\n Yeah. Why?\n\n KRISTOFF\n\n (bending the arm)\n\n How does this work?\n\n Olaf's dismembered arm slaps Kristoff across the face.\n\n OLAF\n\n Stop it, Sven. Trying to focus\n\n here.\n\n (to Anna)\n\n Yeah, why?\n\n KRISTOFF\n\n I'll tell you why. We need Elsa to\n\n bring back summer.\n\n OLAF\n\n (shocked)\n\n Summer?\n\n (sinking into wistfulness)\n\n Oh, I don't know why but I've\n\n always loved the idea of summer,\n\n and sun, and all things hot.\n\n KRISTOFF\n\n Really? I'm guessing you don't have\n\n much experience with heat.\n\n OLAF\n\n Nope. But sometimes I like to close\n\n my eyes and imagine what it'd be\n\n like when summer does come.\n\n DISSOLVE TO:\n\n OLAF'S FANTASY WORLD -- PERFECT SUMMER DAY\n\n Olaf walks through a grassy meadow with the sun shining\n\n behind him. He SINGS.\n\n "In Summer"\n\n OLAF\n\n BEES'LL BUZZ / KIDS'LL BLOW\n\n DANDELION FUZZ / AND I'LL BE DOING\n\n WHATEVER SNOW DOES IN SUMMER.\n\n 57\n\nFROZEN - J. Lee\n\n -Olaf now lies in the sand on a beach.\n\n OLAF (CONT'D)\n\n A DRINK IN MY HAND / MY SNOW UP\n\n AGAINST THE BURNING SAND / PROB'LY\n\n GETTING GORGEOUSLY TANNED IN\n\n SUMMER.\n\n -Olaf sails in a boat.\n\n OLAF (CONT'D)\n\n I'LL FINALLY SEE A SUMMER BREEZE /\n\n BLOW AWAY A WINTER STORM /\n\n -Olaf floats in the water. All his pieces begin to separate.\n\n OLAF (CONT'D)\n\n AND FIND OUT WHAT HAPPENS TO SOLID\n\n WATER / WHEN IT GETS WARM.\n\n -Olaf tumbles on a sandy beach with sand-snowmen.\n\n OLAF (CONT'D)\n\n AND I CAN'T WAIT TO SEE / WHAT MY\n\n BUDDIES ALL THINK OF ME / JUST\n\n IMAGINE HOW MUCH COOLER I'LL BE IN\n\n SUMMER . . !\n\n -Olaf and the seagull break out into a tap-dance.\n\n OLAF (CONT'D)\n\n DA DA . . . DA DOO / AH BAH BAH BAH\n\n BAH BAH BOO.\n\n -Olaf and another snowman drink hot chocolate in a hot tub.\n\n OLAF (CONT'D)\n\n THE HOT AND THE COLD ARE BOTH SO\n\n INTENSE / PUT `EM TOGETHER, IT JUST\n\n MAKES SENSE!\n\n -Olaf tap dances with a gaggle of seagulls.\n\n OLAF (CONT'D)\n\n RATDADAT DAD DADA DOO . . .\n\n -Olaf bounds down a grassy hill.\n\n OLAF (CONT'D)\n\n WINTER'S A GOOD TIME TO STAY IN AND\n\n CUDDLE / BUT PUT ME IN SUMMER AND\n\n I'LL BE A...\n\n He stops at a puddle, looks down at it. Smiles. Hops over it.\n\n 58\n\nFROZEN - J. Lee\n\n OLAF (CONT'D)\n\n HAPPY SNOWMAN!\n\n -Olaf runs with a checkered blanket that he spreads out. He\n\n relaxes and stares at the blue sky.\n\n OLAF (CONT'D)\n\n WHEN LIFE GETS ROUGH I LIKE TO HOLD\n\n ON TO MY DREAM / OF RELAXING IN THE\n\n SUMMER SUN JUST LETTING OFF STEAM!\n\n Sven, Anna, Kristoff and Olaf have a picnic.\n\n OLAF (CONT'D)\n\n OH THE SKY WILL BE BLUE / AND YOU\n\n GUYS'LL BE THERE TOO / WHEN I\n\n FINALLY DO WHAT FROZEN THINGS DO IN\n\n SUMMER!\n\n KRISTOFF\n\n I'm gonna tell him.\n\n ANNA\n\n Don't you dare.\n\n OLAF\n\n IN SUMMER!\n\n Olaf sings the final note. We swing around him and return to:\n\n REALITY. He then straightens up and smiles.\n\n OLAF (CONT'D)\n\n So, come on! Elsa's this way. Let's\n\n go bring back summer!\n\n Olaf grabs Anna's hand and pulls her along up the mountain.\n\n ANNA\n\n (laughing)\n\n I'm coming!\n\n Sven hops along, happily following them. Kristoff watches all\n\n of them like they're nuts.\n\n KRISTOFF\n\n Somebody's got to tell him.\n\n DISSOLVE TO:\n\n 59\n\nFROZEN - J. Lee\n\n EXT. ARENDELLE, VILLAGE -- DAY\n\n A layer of solid ice coats everything. People huddle around\n\n weak fires. Anxiety runs high amongst the villagers and\n\n guests. We pass two CITIZENS fighting over a woodpile.\n\n CITIZEN ONE\n\n No. No. You've got the bark facing\n\n down. The bark needs to be face-up.\n\n CITIZEN TWO\n\n Bark down is drier.\n\n CITIZEN ONE\n\n Bark up.\n\n CITIZEN TWO\n\n Bark down.\n\n CITIZEN ONE\n\n Bark up.\n\n Like a light in the dark, Hans moves through the crowd.\n\n HANS\n\n Cloak. Does anyone need a cloak?\n\n GERDA\n\n Arendelle is indebted to you, Your\n\n Highness.\n\n HANS\n\n The castle is open. There's soup\n\n and hot glogg in the Great Hall.\n\n He hands the stack of cloaks to a guard.\n\n HANS (CONT'D)\n\n Here. Pass these out.\n\n Just then the Duke approaches Hans.\n\n DUKE\n\n Prince Hans, are we just expected\n\n to sit here and freeze while you\n\n give away all of Arendelle's\n\n tradable goods?\n\n HANS\n\n (tall and confident)\n\n Princess Anna has given her orders\n\n and--\n\n 60\n\nFROZEN - J. Lee\n\n DUKE\n\n And that's another thing; has it\n\n dawned on you that your princess\n\n may be conspiring with a wicked\n\n sorceress to destroy us all?\n\n Hans's nice eyes turn to threatening slits.\n\n HANS\n\n Do not question the Princess. She\n\n left me in charge, and I will not\n\n hesitate to protect Arendelle from\n\n treason.\n\n DUKE\n\n (flabbergasted, offended)\n\n Treason?!\n\n Suddenly they hear the alarmed whinny of Anna's horse. It\n\n returns alone, bucking and kicking. Hans grabs its reins.\n\n HANS\n\n Whoa! Whoa! Whoa, boy. Easy. Easy.\n\n CROWD\n\n (various)\n\n Princess Anna's horse. What\n\n happened to her? Where is she?\n\n Hans steadies the horse, looks up at the mountain. He sees\n\n all the panicked faces of the kingdom looking to him.\n\n HANS\n\n ...Princess Anna is in trouble.\n\n (calling out)\n\n I need volunteers to go with me to\n\n find her!\n\n Volunteers, some from Arendelle, some from other lands, rush\n\n up to offer their services.\n\n DUKE\n\n I volunteer two men, my Lord!\n\n (quietly to his thugs)\n\n Be prepared for anything, and\n\n should you encounter the Queen, you\n\n are to put an end to this winter.\n\n Do you understand?\n\n His two thugs sneer.\n\n CUT TO:\n\n 61\n\nFROZEN - J. Lee\n\n EXT. THE NORTH MOUNTAIN -- DAY\n\n Anna, Kristoff, Sven, and Olaf move through hostile terrain.\n\n Wind-swept icicles face horizontal.\n\n KRISTOFF\n\n So how exactly are you planning to\n\n stop this weather?\n\n ANNA\n\n (confident)\n\n Oh, I am gonna talk to my sister.\n\n KRISTOFF\n\n That's your plan? My ice business\n\n is riding on you talking to your\n\n sister.\n\n ANNA\n\n Yup.\n\n Kristoff, so stunned by her casual plan, doesn't look where\n\n he's going and ends up with an ice-spike to the nose. He\n\n stops short, GULP, moves carefully around the spike.\n\n KRISTOFF\n\n So you're not at all afraid of her?\n\n ANNA\n\n Why would I be?\n\n OLAF\n\n (oblivious)\n\n Yeah. I bet Elsa's the nicest,\n\n gentlest, warmest person ever.\n\n Olaf backs right into an icicle. It runs through his torso.\n\n OLAF (CONT'D)\n\n Oh, look at that. I've been\n\n impaled.\n\n He laughs it off.\n\n DISSOLVE TO:\n\n EXT. STEEP MOUNTAIN FACE -- DAY\n\n Anna and Kristoff hit what looks like a dead end. The face of\n\n the mountain goes straight up.\n\n ANNA\n\n What now?\n\n 62\n\nFROZEN - J. Lee\n\n Kristoff looks around, sighs. Digs in his rucksack.\n\n KRISTOFF\n\n ...It's too steep. I've only got\n\n one rope, and you don't know how to\n\n climb mountains.\n\n ANNA (O.S.)\n\n Says who?\n\n Sven nudges Kristoff, who looks up to see Anna trying to\n\n climb the cliff's flat face.\n\n KRISTOFF\n\n (finding her ridiculous)\n\n What are you doing?\n\n ANNA\n\n (straining)\n\n ...I'm going to see my sister.\n\n KRISTOFF\n\n You're going to kill yourself.\n\n Kristoff watches her searching for footholds and hand-holds.\n\n KRISTOFF (CONT'D)\n\n I wouldn't put my foot there.\n\n ANNA (O.S.)\n\n You're distracting me.\n\n KRISTOFF\n\n Or there. How do you know Elsa even\n\n wants to see you?\n\n ANNA (O.S.)\n\n I'm just blocking you out cause I\n\n gotta concentrate here.\n\n KRISTOFF\n\n You know, most people who disappear\n\n into the mountains want to be\n\n alone.\n\n ANNA (O.S.)\n\n Nobody wants to be alone. Except\n\n maybe you--\n\n KRISTOFF\n\n I'm not alone.... I have friends,\n\n remember?\n\n Anna kicks a foot above her head to catch a foot hold.\n\n 63\n\nFROZEN - J. Lee\n\n ANNA\n\n You mean the love experts?\n\n KRISTOFF\n\n Yes, the love experts!\n\n Anna realizes she's stuck.\n\n ANNA\n\n ...Please tell me I'm almost there.\n\n REVEAL: she's only about six feet up. Her muscles shake.\n\n ANNA (CONT'D)\n\n ...Does the air seem a bit thin to\n\n you up here?\n\n Kristoff smiles, getting a kick out of her.\n\n KRISTOFF\n\n Hang on.\n\n He pulls the rope from his bag. Just then Olaf steps out from\n\n behind a rock and waves to Kristoff.\n\n OLAF\n\n Hey, Sven? Not sure if this is\n\n going to solve the problem, but I\n\n found a staircase that leads\n\n exactly where you want it to go.\n\n ANNA\n\n Ha ha. Thank goodness. Catch!\n\n Anna drops off the cliff. Kristoff catches her.\n\n ANNA (CONT'D)\n\n Thanks! That was like a crazy trust\n\n exercise.\n\n She hops down, brushes off her dress, and bounds off.\n\n Kristoff watches after her, digging her fearless pluck.\n\n EXT. BASE OF THE ICE PALACE -- DAY\n\n Anna, Kristoff, and Olaf approach Elsa's elegant ice palace.\n\n ANNA\n\n Whoa.\n\n KRISTOFF\n\n (in awe)\n\n Now that's ice. I might cry.\n\n 64\n\nFROZEN - J. Lee\n\n ANNA\n\n Go ahead. I won't judge.\n\n Anna climbs the steps with Olaf. Sven tries to follow. His\n\n hooves slip out. He scrambles but can't get traction.\n\n Kristoff runs to his aide.\n\n KRISTOFF\n\n All right, take it easy. I gotcha.\n\n Kristoff settles Sven back down the stairs and pats him.\n\n KRISTOFF (CONT'D)\n\n You stay right here, buddy.\n\n Sven obediently plops his reindeer butt down and wags his\n\n tail. Kristoff climbs the stairs, admiring the ice details.\n\n KRISTOFF (CONT'D)\n\n ...Flawless.\n\n Anna arrives at the door. Hesitates.\n\n OLAF\n\n ...Knock....\n\n (she doesn't)\n\n Just knock....\n\n (she doesn't. To Kristoff)\n\n Why isn't she knocking...? Do you\n\n think she knows how to knock?\n\n Anna finally KNOCKS. The sound echoes inside. The ice doors\n\n slide open.\n\n ANNA\n\n Ha. It opened. That's a first.\n\n Anna goes to step in. Kristoff follows. She gets a thought,\n\n stops him.\n\n ANNA (CONT'D)\n\n You should probably wait out here.\n\n KRISTOFF\n\n What?\n\n ANNA\n\n Last time I introduced her to a\n\n guy, she froze everything.\n\n KRISTOFF\n\n But, it's a palace made of ice. Ice\n\n is my life.\n\n 65\n\nFROZEN - J. Lee\n\n OLAF\n\n Bye, Sven.\n\n Olaf starts to head inside. Anna stops him.\n\n ANNA\n\n You too, Olaf.\n\n OLAF\n\n Me?\n\n ANNA\n\n Just give us a minute.\n\n OLAF\n\n Okay.\n\n As Anna walks inside. Olaf starts counting.\n\n OLAF (CONT'D)\n\n One...two...\n\n Kristoff joins in.\n\n OLAF AND KRISTOFF\n\n Three...four...\n\n INT. ELSA'S PALACE -- DAY\n\n Anna walks into a great foyer. The place is beautiful, but\n\n also eerie.\n\n ANNA\n\n Elsa? It's me...Anna?!\n\n Anna slips. Steadies herself.\n\n ELSA (O.S.)\n\n Anna.\n\n Elsa steps out of the shadows onto a balcony. She sees Anna,\n\n looks to her longingly.\n\n Anna can't help but be struck by Elsa's beauty.\n\n ANNA\n\n Elsa, you look different.... It's a\n\n good different.... And this place\n\n is amazing.\n\n 66\n\nFROZEN - J. Lee\n\n ELSA\n\n (cautious, polite)\n\n Thank you, I never knew what I was\n\n capable of.\n\n Anna starts to climb the stairs.\n\n ANNA\n\n ...I'm so sorry about what\n\n happened. If I'd known--\n\n Elsa backs up, away from Anna.\n\n ELSA\n\n (on guard)\n\n No, it's okay. You don't have to\n\n apologize.... But you should\n\n probably go, please.\n\n ANNA\n\n But I just got here.\n\n ELSA\n\n ...You belong in Arendelle.\n\n ANNA\n\n So do you.\n\n Anna takes another step up. Elsa backs up more.\n\n ELSA\n\n No, I belong here. Alone. Where I\n\n can be who I am without hurting\n\n anybody.\n\n ANNA\n\n ...Actually, about that--\n\n OLAF (O.S.)\n\n 58...59...60.\n\n ELSA\n\n Wait. What is that?\n\n Olaf comes running in the front door. He waves.\n\n OLAF\n\n Hi, I'm Olaf and I like warm hugs.\n\n ELSA\n\n (shocked)\n\n Olaf?\n\n Olaf stops beside Anna, looks up at Elsa, intimidated.\n\n 67\n\nFROZEN - J. Lee\n\n OLAF\n\n (bashful)\n\n You built me. You remember that?\n\n ELSA\n\n (astonished)\n\n And you're alive?\n\n OLAF\n\n Um...I think so?\n\n Anna kneels down beside Olaf.\n\n ANNA\n\n He's just like the one we built as\n\n kids.... We were so close. We can\n\n be like that again.\n\n Elsa smiles, but then a memory returns to her.\n\n FLASH CUT TO:\n\n FLASHBACK: Young Anna is struck by Elsa's powers.\n\n YOUNG ELSA\n\n Anna!\n\n Young Anna falls unconscious. Young Elsa races to her.\n\n FLASH CUT TO:\n\n THE PRESENT: Elsa's face sinks in pain.\n\n ELSA\n\n No, we can't.\n\n Elsa turns and heads up the second story steps.\n\n ELSA (CONT'D)\n\n Goodbye, Anna.\n\n ANNA\n\n Elsa, wait--\n\n ELSA\n\n (calling back)\n\n I'm just trying to protect you.\n\n Elsa continues to flee. Anna pursues.\n\n ANNA\n\n You don't have to protect me. I'm\n\n not afraid. Please don't shut me\n\n out again.\n\n 68\n\nFROZEN - J. Lee\n\n Anna SINGS.\n\n "First Time in Forever, Reprise"\n\n ANNA (CONT'D)\n\n PLEASE DON'T SLAM THE DOOR.\n\n YOU DON'T HAVE TO KEEP YOUR\n\n DISTANCE ANYMORE.\n\n `CAUSE FOR THE FIRST TIME IN\n\n FOREVER,\n\n I FINALLY UNDERSTAND.\n\n FOR THE FIRST TIME IN FOREVER,\n\n WE CAN FIX THIS HAND IN HAND.\n\n WE CAN HEAD DOWN THIS MOUNTAIN\n\n TOGETHER.\n\n YOU DON'T HAVE TO LIVE IN FEAR.\n\n `CAUSE FOR THE FIRST TIME IN\n\n FOREVER,\n\n I WILL BE RIGHT HERE.\n\n They arrive on the top floor, Elsa's main living space. Elsa\n\n turns back to Anna, grateful, but determined.\n\n ELSA\n\n Anna,\n\n PLEASE GO BACK HOME.\n\n YOUR LIFE AWAITS.\n\n GO ENJOY THE SUN\n\n AND OPEN UP THE GATES.\n\n ANNA\n\n Yeah, but--\n\n ELSA\n\n I know!\n\n YOU MEAN WELL,\n\n BUT LEAVE ME BE.\n\n YES, I'M ALONE BUT I'M ALONE AND\n\n FREE.\n\n Elsa opens up the balcony doors.\n\n ELSA (CONT'D)\n\n JUST STAY AWAY AND YOU'LL BE SAFE\n\n FROM ME.\n\n ANNA\n\n ACTUALLY, WE'RE NOT.\n\n ELSA\n\n WHAT DO YOU MEAN YOU'RE NOT?\n\n 69\n\nFROZEN - J. Lee\n\n ANNA\n\n I GET THE FEELING YOU DON'T KNOW?\n\n ELSA\n\n WHAT DO I NOT KNOW?\n\n ANNA\n\n ARENDELLE'S IN DEEP DEEP DEEP DEEP\n\n SNOW.\n\n ELSA\n\n What?\n\n Elsa looks past Anna's shoulder out white-peaked mountains.\n\n ANNA\n\n You kind of set off an eternal\n\n winter...everywhere.\n\n ELSA\n\n Everywhere?\n\n ANNA\n\n It's okay, you can just unfreeze\n\n it.\n\n ELSA\n\n No, I can't. I don't know how.\n\n ANNA\n\n Sure you can. I know you can.\n\n Snow starts to swirl around the room.\n\n ANNA (CONT'D)\n\n CUZ FOR THE FIRST TIME IN FOREVER,\n\n ELSA\n\n (panicking)\n\n I'M SUCH A FOOL!\n\n I CAN'T BE FREE!\n\n ANNA\n\n YOU DON'T HAVE TO BE AFRAID.\n\n ELSA\n\n NO ESCAPE\n\n FROM THE STORM INSIDE OF ME!\n\n The snow picks up. Anna tries to fight through it.\n\n ANNA\n\n WE CAN WORK THIS OUT TOGETHER.\n\n 70\n\nFROZEN - J. Lee\n\n ELSA\n\n I CAN'T CONTROL THE CURSE!\n\n ANNA\n\n WE'LL REVERSE THE STORM YOU'VE\n\n MADE.\n\n ELSA\n\n ANNA, PLEASE, YOU'LL ONLY MAKE IT\n\n WORSE!\n\n ANNA\n\n DON'T PANIC.\n\n ELSA\n\n THERE'S SO MUCH FEAR!\n\n ANNA\n\n WE'LL MAKE THE SUN SHINE BRIGHT.\n\n ELSA\n\n YOU'RE NOT SAFE HERE!\n\n ANNA\n\n WE CAN FACE THIS THING TOGETHER...\n\n But as Anna sings, we lose sight of her in the thickening\n\n blizzard taking over the room.\n\n ELSA\n\n NO!\n\n ANNA (O.S.)\n\n WE CAN CHANGE THIS WINTER WEATHER,\n\n AND EVERYTHING WILL BE...\n\n Anna's voice disappears in the storm as Elsa cries out.\n\n ELSA\n\n I CAN'T!\n\n Elsa's fear, so strong, sucks the blizzard back into her and\n\n then it bursts out, unwittingly, like a sharp snowflake.\n\n Anna is STRUCK right in the heart. She grasps her chest in\n\n pain and stumbles back. She falls to her knees.\n\n Elsa gasps when she sees Anna. Just then, Olaf and Kristoff\n\n rush into the room to Anna's side.\n\n KRISTOFF\n\n Anna. Are you okay?\n\n 71\n\nFROZEN - J. Lee\n\n ANNA\n\n I'm okay.... I'm fine.\n\n Anna gets to her feet, determined to hide the pain.\n\n ELSA\n\n (scared)\n\n Who's this? Wait, it doesn't\n\n matter. You have to go.\n\n ANNA\n\n No, I know we can figure this out\n\n together--\n\n ELSA\n\n (desperate)\n\n How? What power do you have to stop\n\n this winter? To stop me?\n\n Anna doesn't have the answer. Kristoff sees spiky ice shadows\n\n creeping down the walls. Puts a protective arm around Anna.\n\n KRISTOFF\n\n Anna, I think we should go.\n\n ANNA\n\n (close to tears)\n\n No. I'm not leaving without you,\n\n Elsa.\n\n ELSA\n\n (heartbroken but decisive)\n\n Yes, you are.\n\n Elsa waves her arms and builds a giant, menacing snowman.\n\n We'll call him MARSHMALLOW.\n\n SLAM CUT TO:\n\n EXT. ICE PALACE -- DAY\n\n Marshmallow holds Anna and Kristoff by the scruff of their\n\n necks in one hand and Olaf in the other.\n\n ANNA\n\n Stop. Put us down!\n\n OLAF\n\n (to Marshmallow)\n\n You are a lot stronger than I think\n\n you realize.\n\n Marshmallow tosses Kristoff and Anna down the steps.\n\n 72\n\nFROZEN - J. Lee\n\n MARSHMALLOW\n\n (like a bouncer)\n\n Go away!\n\n Anna and Kistoff slide past Sven, who's got his tongue stuck\n\n to the ice railing.\n\n OLAF (O.S.)\n\n Heads up!\n\n Olaf's head smashes into a snowbank nearby.\n\n ANNA\n\n Olaf!\n\n OLAF\n\n Watch out for my butt!\n\n Anna and Kristoff duck as the rest of Olaf slams into the\n\n snowbank.\n\n Marshmallow turns to go back into the castle.\n\n Incensed, Anna tries to march back up the stairs.\n\n ANNA\n\n It is not nice to throw people!\n\n Kristoff grabs her, pulls her back.\n\n KRISTOFF ANNA\n\n All right feisty pants. Calm Let me at him. I want to get\n\n down. Woaw. Just let the snow him. I.... Okay. I'm Calm.\n\n man be.\n\n Anna backs down...for a moment. Then she grabs a snowball and\n\n throws it at Marshmallow.\n\n The tiny little ball hits Marshmallow's back, not making even\n\n the slightest dent. But it's enough to infuriate him. He\n\n ROARS. Spikes shoot out of his joints.\n\n KRISTOFF\n\n Uh-oh. Now you made him mad!\n\n OLAF\n\n ...I'll distract him. You guys go.\n\n Kristoff pushes Anna along. Sven runs off in the opposite\n\n direction. Olaf's belly and butt fall and follow Sven.\n\n OLAF (CONT'D)\n\n No, no, not you guys.\n\n 73\n\nFROZEN - J. Lee\n\n Marshmallow goes charging after Anna and Kristoff as Olaf's\n\n head falls and lands face down in snow.\n\n OLAF (CONT'D)\n\n (muffled)\n\n This just got a whole lot harder.\n\n Anna and Kristoff leap and slide down a steep slope. They\n\n tumble to a stop at the bottom just as Marshmallow lands hard\n\n right behind them.\n\n They're off again...through a maze of conifers that sag under\n\n the weight of the snow, Marshmallow hot on their trail.\n\n KRISTOFF\n\n This way!\n\n Anna grabs a branch of a sagging trees and releases all of\n\n the snow. The tree snaps upright, knocking Marshmallow back.\n\n KRISTOFF (CONT'D)\n\n (impressed)\n\n Ho-ho-ho!\n\n ANNA\n\n I got him!\n\n Anna and Kristoff burst out of the conifer forest and almost\n\n run right off a cliff. They stop short, toes on the edge.\n\n KRISTOFF\n\n Whoa, stop!\n\n ANNA\n\n It's a hundred foot drop.\n\n KRISTOFF\n\n It's two hundred.\n\n Kristoff ties the rope around Anna and pulls tight.\n\n ANNA\n\n Ow.\n\n He drops to his knees and starts digging a u-shape in the\n\n snow with a pick axe.\n\n ANNA (CONT'D)\n\n What's that for?\n\n KRISTOFF\n\n I'm digging a snow anchor.\n\n 74\n\nFROZEN - J. Lee\n\n ANNA\n\n (not trusting)\n\n Okay. What if we fall?\n\n KRISTOFF\n\n There's twenty feet of fresh powder\n\n down there; it'll be like landing\n\n on a pillow.... Hopefully.\n\n They hear an angry ROAR coming closer.\n\n KRISTOFF (CONT'D)\n\n Okay, Anna. On three.\n\n Anna preps for the jump like a boxer getting ready to fight.\n\n ANNA\n\n Okay. You tell me when...\n\n KRISTOFF\n\n One...\n\n ANNA\n\n ...I'm ready to go....\n\n KRISTOFF\n\n Two...\n\n ANNA\n\n (pumped up)\n\n ...I was BORN ready! Yes!\n\n KRISTOFF\n\n Calm down.\n\n A huge tree flies through the air toward them.\n\n ANNA (O.S.)\n\n TREE!\n\n Anna jumps and pulls Kristoff over the edge with her. They\n\n hang upside down over the cliff by the rope. The rope catches\n\n their fall.\n\n KRISTOFF\n\n Whoa! That happened.\n\n Back up top, Olaf emerges from the woods. He's a complete\n\n mess, all his body parts are in the wrong places. He huffs\n\n and puffs, struggling to run.\n\n OLAF\n\n Ah. Ah. Man, am I out of shape.\n\n 75\n\nFROZEN - J. Lee\n\n He stops. Puts his body back together in the right order.\n\n OLAF (CONT'D)\n\n There we go. Hey, Anna! Sven!\n\n Where'd ya guys go? We totally lost\n\n Marshmallow back there!\n\n Marshmallow steps up behind Olaf. Olaf turns to face him.\n\n OLAF (CONT'D)\n\n (happily)\n\n Hey. We were just talking about\n\n you. All good things, all good\n\n things.\n\n Marshmallow roars and approaches Kristoff's snow anchor.\n\n OLAF (CONT'D)\n\n NO!\n\n Olaf jumps onto Marshmallow's leg trying to stop him, but not\n\n making much of a difference.\n\n OLAF (CONT'D)\n\n This is not making much of a\n\n difference!\n\n Marshmallow flicks Olaf off his leg and right over the cliff.\n\n OLAF (CONT'D)\n\n WHOA!\n\n Olaf passes Anna and Kristoff.\n\n ANNA\n\n Olaf!\n\n OLAF\n\n Hang in there, guys!\n\n Marshmallow starts yanking Kristoff and Anna's rope up.\n\n ANNA\n\n Wait, what?\n\n Kristoff's head hits the cliff.\n\n KRISTOFF\n\n Aargghh!\n\n Kristoff passes out and hangs like a rag doll.\n\n ANNA\n\n Kristoff!\n\n 76\n\nFROZEN - J. Lee\n\n Marshmallow pulls them up. He roars and breathes snow all\n\n over them.\n\n MARSHMALLOW\n\n Don't come back!\n\n ANNA\n\n (grossed out by his snow\n\n breath)\n\n Ugh. We won't.\n\n Anna whips out a knife and cuts the rope. Kristoff comes to\n\n just as they fall. They both SCREAM!\n\n SLAM!\n\n REVEAL: Anna opens her eyes to find herself buried up to her\n\n shoulders in the soft thick snow. She laughs.\n\n ANNA (CONT'D)\n\n Hey, you were right. Just like a\n\n pillow.\n\n She looks up to see Olaf's upper half hanging onto Kristoff's\n\n boots, which are sticking out of the snow.\n\n OLAF\n\n (shaking the boots)\n\n I can't feel my legs! I can't feel\n\n my legs!\n\n Suddenly, Kristoff's head pops up. He spits out snow.\n\n KRISTOFF\n\n Those are my legs.\n\n Olaf's bottom goes running by.\n\n OLAF\n\n (to Kristoff)\n\n Ooh. Hey, do me a favor, grab my\n\n butt.\n\n Kristoff grabs Olaf's head and puts it on his body.\n\n OLAF (CONT'D)\n\n Oh, that feels better.\n\n Sven walks up and sniffs Olaf's nose.\n\n OLAF (CONT'D)\n\n Hey, Sven!\n\n 77\n\nFROZEN - J. Lee\n\n Olaf turns to Anna and Kristoff just as Sven goes to bite off\n\n his nose -- and misses.\n\n OLAF (CONT'D)\n\n He found us.\n\n (to Sven, funny voice)\n\n Who's my cute little reindeer?\n\n KRISTOFF\n\n Don't talk to him like that.\n\n Kristoff goes over to help Anna, who is stuck in the snow.\n\n KRISTOFF (CONT'D)\n\n Here.\n\n He lifts her out easily.\n\n ANNA\n\n (impressed)\n\n Whoa!\n\n KRISTOFF\n\n You okay?\n\n ANNA\n\n Thank you.\n\n They meet eyes. Wait. Is that chemistry?\n\n ANNA (CONT'D)\n\n ...Um.... How's your head?\n\n She touches the spot where he banged his head.\n\n KRISTOFF\n\n (in pain)\n\n Ah! Ooh!\n\n He catches himself. Waves off the pain with a giggle.\n\n KRISTOFF (CONT'D)\n\n I mean, It's fine. Ah...I'm good.\n\n Ha. I've got a thick skull.\n\n OLAF\n\n I don't have a skull.... Or bones.\n\n KRISTOFF\n\n ...So....\n\n The awkwardness is killing him.\n\n 78\n\nFROZEN - J. Lee\n\n KRISTOFF (CONT'D)\n\n (shy)\n\n Now what?\n\n ANNA\n\n (shy)\n\n Now what?\n\n (then...panicking)\n\n Now what?! Oh! What am I gonna do?\n\n She threw me out. I can't go back\n\n to Arendelle with the weather like\n\n this. And then there's your ice\n\n business--\n\n KRISTOFF\n\n Hey, hey, don't worry about my ice\n\n business...\n\n (noticing something)\n\n Worry about your hair?!\n\n She thinks he means it looks bad. She smooths it down.\n\n ANNA\n\n What? I just fell off a cliff. You\n\n should see your hair.\n\n KRISTOFF\n\n No, yours is turning white.\n\n She grabs her braid as a tendril turns white.\n\n ANNA\n\n White? It's what?\n\n KRISTOFF\n\n It's because she struck you; isn't\n\n it?\n\n ANNA\n\n Does it look bad?\n\n KRISTOFF\n\n (thinking)\n\n ...No.\n\n Olaf's head pops up. He's holding his head up off his body to\n\n join the conversation.\n\n OLAF\n\n You hesitated.\n\n KRISTOFF\n\n No, I didn't. Anna, you need help.\n\n Now, come on.\n\n 79\n\nFROZEN - J. Lee\n\n He heads towards the sunset. Sven and Olaf follow.\n\n OLAF\n\n Okay! Where are we going?\n\n KRISTOFF\n\n To see my friends.\n\n ANNA\n\n (catching up)\n\n The love experts?\n\n OLAF\n\n Love experts?!\n\n KRISTOFF\n\n Yes. And don't worry; they'll be\n\n able to fix this.\n\n ANNA\n\n How do you know?\n\n He looks her over, remembering the moment he saw the trolls\n\n heal her as a child.\n\n KRISTOFF\n\n ...Because I've seen them do it\n\n before.\n\n As they round the bend, the sun sets and Olaf turns to Sven.\n\n OLAF\n\n I like to consider myself a love\n\n expert.\n\n CUT TO:\n\n INT. ELSA'S PALACE -- DAY\n\n Elsa paces, distraught. She talks to herself.\n\n ELSA\n\n (mantra-style)\n\n Get it together. Control it. Don't\n\n feel. Don't feel. Don't FEEL!\n\n She hears ice cracking. Stops. Looks around. She's left a\n\n sharp wake of ice spikes behind her on the floor. They grow\n\n up the wall, taking over the castle.\n\n DISSOLVE TO:\n\n 80\n\nFROZEN - J. Lee\n\n EXT. BLACK MOUNTAINS -- NIGHT\n\n The Northern Lights are bright. Olaf stares at them in awe as\n\n he rides on Sven's back.\n\n OLAF\n\n Look, Sven. The sky's awake.\n\n Behind Olaf and Sven, Anna walks with Kristoff. She shivers.\n\n KRISTOFF\n\n Are you cold?\n\n ANNA\n\n ...A little.\n\n He reaches like he might put an arm around her, but decides\n\n against it. He looks around as if he doesn't know what to do,\n\n then gets a thought.\n\n KRISTOFF\n\n Wait. Come here.\n\n He takes her hand and pulls her around a bend into a rock-\n\n lined pass.\n\n Steam vents, powered by the volcanic activity, dot the path.\n\n He holds her hands over one of them.\n\n ANNA\n\n Oooh.... That's nice.\n\n They continue on the path, walking from vent to vent.\n\n KRISTOFF\n\n (taking a deep breath)\n\n So, about my friends...well, I say\n\n friends, they're more like\n\n family.... Anyway, when I was a\n\n kid, it was just me and\n\n Sven...until they took me in.\n\n ANNA\n\n (moved)\n\n They did?\n\n KRISTOFF\n\n (nervous ramble)\n\n Yeah. I don't want to scare you,\n\n they can be a little bit\n\n inappropriate...and loud...very\n\n loud...they're also stubborn at\n\n times, and a little overbearing.\n\n And heavy. Really, really heavy.\n\n (MORE)\n\n 81\n\nFROZEN - J. Lee\n\n KRISTOFF (CONT'D)\n\n But they're fine.. You'll get it.\n\n They mean well.\n\n Anna touches Kristoff's arm, reassuringly.\n\n ANNA\n\n Kristoff, they sound wonderful.\n\n Kristoff smiles, appreciating her sincerity.\n\n KRISTOFF\n\n Okay then....\n\n Mustering the courage, Kristoff steps forward and with a wave\n\n of the arms announces--\n\n KRISTOFF (CONT'D)\n\n Meet my family.\n\n REVEAL: he's surrounded by rocks.\n\n KRISTOFF (CONT'D)\n\n (to the rocks)\n\n Hey, guys!\n\n As Kristoff and Sven move through the rocks, waving and\n\n greeting, Olaf and Anna stand frozen, dumbfounded.\n\n ANNA\n\n (to herself)\n\n ...They're rocks.\n\n OLAF\n\n (realizing)\n\n He's crazy.\n\n (covertly, to Anna)\n\n I'll distract them while you run.\n\n (Loud and slow to a rock)\n\n Hi, Sven's family! It's nice to\n\n meet you!\n\n (quietly to Anna)\n\n Anna, because I love you, I insist\n\n you run.\n\n (to the rock)\n\n I understand you're love experts!\n\n (to Anna)\n\n Why aren't you running?\n\n Anna snaps out of her shock and starts backing away.\n\n ANNA\n\n Okay. Um...I'm gonna go--\n\n Just then the rocks around her start rolling.\n\n 82\n\nFROZEN - J. Lee\n\n ANNA (CONT'D)\n\n (panicking)\n\n Kristoff!\n\n Olaf lights up and chases the rocks, who surround Kristoff\n\n and unfold as trolls.\n\n BULDA\n\n KRISTOFF'S HOME!\n\n TROLLS (VARIOUS)\n\n Kristoff! Kristoff's home! It's\n\n been too long! Kristoff's home!\n\n Olaf jumps around all excitedly.\n\n OLAF\n\n (excitedly)\n\n Kristoff's home.\n\n He then stops, confused, and looks to one of the trolls.\n\n OLAF (CONT'D)\n\n Wait? Kristoff?\n\n Anna watches, shocked and confused.\n\n The trolls all want Kristoff's attention. One troll yanks him\n\n down with a boulder's strength.\n\n TROLL ONE\n\n Oh, lemme look at you!\n\n Another troll tries to pull off his clothes.\n\n TROLL TWO\n\n Oh, take off your clothes,\n\n Kristoff; I wash them.\n\n KRISTOFF\n\n (holding up his pants)\n\n Ah! No. I'm gonna keep my clothes\n\n on, thank you.\n\n KRISTOFF (CONT'D)\n\n Great to see you all. Where's\n\n grandpa?\n\n MUSHROOM KID TROLL\n\n He's napping. But look, I grew a\n\n mushroom.\n\n TROLL SCOUT KID\n\n And I earned my fire crystal.\n\n 83\n\nFROZEN - J. Lee\n\n KIDNEY STONE TROLL\n\n I passed a kidney stone.\n\n PICK ME UP TROLL\n\n Pick me up.\n\n The kid troll jumps up on Kristoff's arm. Kristoff sinks\n\n under the weight of him.\n\n Anna still stares, confused, then realizes...\n\n ANNA\n\n Trolls? They're trolls.\n\n Silence. All troll eyes turn to Anna. Blink. Blink.\n\n BULDA\n\n ...He's brought a girl!\n\n TROLLS (TOGETHER)\n\n He's brought a girl!\n\n Suddenly Anna is surrounded by trolls. They body-surf/roll\n\n Anna over to Kristoff. She falls into his arms.\n\n ANNA\n\n What's going on?\n\n KRISTOFF\n\n I've learned to just roll with it.\n\n Bulda climbs on top of her husband, Cliff, to get a good look\n\n at Anna. She studies her like she's a piece of cattle.\n\n BULDA\n\n Let me see. Bright eyes. Working\n\n nose. Strong teeth. Yes, yes, yes.\n\n She'll do nicely for our Kristoff.\n\n ANNA\n\n Wait. Oh. Um. No.\n\n KRISTOFF\n\n You've got the wrong idea. That's\n\n not why I brought her here.\n\n ANNA\n\n Right. We're not. I'm not--\n\n Anna laughs, uncomfortable, not knowing what to say.\n\n 84\n\nFROZEN - J. Lee\n\n BULDA\n\n (to Anna)\n\n What's the issue, dear? Why are you\n\n holding back from such a man?\n\n Bulda SINGS.\n\n "Fixer-Upper"\n\n TROLLS (VARIOUS)\n\n IS IT THE CLUMPY WAY HE WALKS?\n\n OR THE GRUMPY WAY HE TALKS?\n\n OR THE PEAR-SHAPED, SQUARE-SHAPED\n\n WEIRDNESS OF HIS FEET?\n\n AND THOUGH WE KNOW HE WASHES WELL\n\n HE ALWAYS ENDS UP SORTA SMELLY.\n\n BUT YOU'LL NEVER MEET A FELLA WHO'S\n\n AS SENSITIVE AND SWEET.\n\n TROLLS (CHORUS) (CONT'D)\n\n SO HE'S A BIT OF A FIXER UPPER,\n\n SO HE'S GOT A FEW FLAWS-\n\n HIS PECULIAR BRAIN, DEAR.\n\n HIS THING FOR THE REINDEER\n\n THAT OUTSIDE A FEW OF NATURE'S\n\n LAWS.\n\n SO HE'S A BIT OF A FIXER UPPER,\n\n BUT THIS WE'RE CERTAIN OF-\n\n YOU CAN FIX THIS FIXER UPPER UP\n\n WITH A LITTLE BIT OF LOVE.\n\n KRISTOFF\n\n Can we just stop talking about\n\n this?! We've got a real, actual\n\n problem here.\n\n BULDA\n\n I'll say--\n\n (To Anna)\n\n IS IT THE WAY THAT HE RUNS SCARED?\n\n TROLLS (VARIOUS)\n\n OR THAT HE'S SOCIALLY IMPAIRED?\n\n KID TROLL\n\n OR THAT HE ONLY LIKES TO TINKLE IN\n\n THE WOODS?\n\n TROLLS (VARIOUS)\n\n ARE YOU HOLDING BACK YOUR FONDNESS\n\n DUE TO HIS UNMANLY BLONDENESS?\n\n OR THE WAY HE COVERS UP THAT HE'S\n\n THE HONEST GOODS?\n\n 85\n\nFROZEN - J. Lee\n\n TROLLS (CHORUS) (CONT'D)\n\n HE'S JUST A BIT OF A FIXER UPPER-\n\n HE'S GOT A COUPLE A' BUGS.\n\n KRISTOFF\n\n No, I don't.\n\n TROLLS\n\n HIS ISOLATION\n\n IS CONFIRMATION\n\n OF HIS DESPERATION FOR HEALING\n\n HUGS.\n\n SO HE'S A BIT OF A FIXER UPPER,\n\n BUT WE KNOW WHAT TO DO.\n\n THE WAY TO FIX UP THIS FIXER UPPER\n\n IS TO FIX HIM UP WITH YOU.\n\n The girl trolls sweep Anna away. The boys take Kristoff.\n\n KRISTOFF\n\n (to the male trolls)\n\n Enough! She's engaged to someone\n\n else. Okay?!\n\n TROLLS beat. Blink. Blink. The boy trolls turn, huddle...\n\n TROLLS (VARIOUS)\n\n SO SHE'S A BIT OF A FIXER UPPER,\n\n THAT'S A MINOR THING.\n\n THIS QUOTE "ENGAGEMENT"\n\n IS A FLEX ARRANGEMENT.\n\n KID TROLL\n\n AND BY THE WAY, I DON'T SEE NO\n\n RING.\n\n TROLLS (VARIOUS)\n\n SO SHE'S A BIT OF A FIXER UPPER,\n\n HER BRAIN'S A BIT BETWIXT.\n\n GET THE FIANCE\n\n OUT OF THE WAY\n\n AND THE WHOLE THING WILL BE FIXED!\n\n GIRL TROLLS\n\n WE AREN'T SAYING YOU CAN CHANGE HIM\n\n TROLLS (VARIOUS)\n\n 'CAUSE PEOPLE DON'T REALLY CHANGE.\n\n WE'RE ONLY SAYING THAT LOVE'S A\n\n FORCE\n\n THAT'S POWERFUL AND STRANGE.\n\n PEOPLE MAKE BAD CHOICES\n\n IF THEY'RE MAD OR SCARED OR\n\n STRESSED.\n\n (MORE)\n\n 86\n\nFROZEN - J. Lee\n\n TROLLS (VARIOUS) (CONT'D)\n\n BUT THROW A LITTLE LOVE THEIR WAY\n\n (THROW A LITTLE LOVE THEIR WAY)\n\n AND YOU'LL BRING OUT THEIR BEST!\n\n TRUE LOVE BRINGS OUT THE BEST!\n\n Kristoff looks over at Anna. She actually looks shockingly\n\n beautiful dressed in moss, lit by shimmering crystals.\n\n ALL TROLLS\n\n EVERYONE'S A BIT OF A FIXER UPPER,\n\n THAT'S WHAT IT'S ALL ABOUT\n\n FATHER, SISTER, BROTHER\n\n WE NEED EACH OTHER\n\n TO RAISE US UP AND ROUND US OUT\n\n By this time Kristoff and Anna are being ushered into a pit\n\n by the sheer force of numbers.\n\n TROLLS\n\n EVERYONE'S A BIT OF A FIXER UPPER,\n\n BUT WHEN PUSH COMES TO SHOVE-\n\n THE ONLY FIXER UPPER FIXER THAT CAN\n\n FIX A FIXER UPPER IS\n\n TRUE\n\n TRUE\n\n TRUE\n\n TRUE\n\n LOVE\n\n During this last bit Anna and Kristoff are looking at each\n\n other differently. Hmmm. Maybe those trolls are right?\n\n Sparks! Chemistry!\n\n TROLL PRIEST\n\n Do you, Anna, take Kristoff to be\n\n your trollfully wedded--\n\n ANNA\n\n Wait, what?!\n\n TROLL PRIEST\n\n You're getting married.\n\n TROLLS\n\n LOVE!\n\n Just then, Anna collapses. Kristoff catches her. She's\n\n shivering something fierce.\n\n KRISTOFF\n\n Anna?\n\n He pulls off her cape and hat.\n\n 87\n\nFROZEN - J. Lee\n\n KRISTOFF (CONT'D)\n\n She's as cold as ice.\n\n Just then Grand Pabbie pushes his way through the crowd.\n\n Trolls clear the way for Pabbie. He stops at the edge of the\n\n pit.\n\n GRAND PABBIE\n\n There's strange magic here!\n\n KRISTOFF\n\n Grand Pabbie!\n\n GRAND PABBIE\n\n Bring her to me, Kristoff.\n\n Kristoff helps Anna over. Pabbie looks into her weak eyes.\n\n GRAND PABBIE (CONT'D)\n\n Anna, your life is in danger. There\n\n is ice in your heart, put there by\n\n your sister. If not removed, to\n\n solid ice will you freeze, forever.\n\n ANNA\n\n What...? No.\n\n KRISTOFF\n\n So remove it, Grand Pabbie.\n\n GRAND PABBIE\n\n I can't. If it was her head, that\n\n would be easy. But only an act of\n\n true love can thaw a frozen heart.\n\n ANNA\n\n An act of true love?\n\n BULDA\n\n (googley, to her hubby)\n\n A true love's kiss, perhaps?\n\n A bunch of trolls give each other kisses.\n\n Anna shivers again, collapsing into Kristoff's arms. More of\n\n her hair turns white.\n\n KRISTOFF\n\n Anna, we've got to get you back to\n\n Hans.\n\n ANNA\n\n (still weak)\n\n ...Hans.\n\n 88\n\nFROZEN - J. Lee\n\n KRISTOFF\n\n Help us out, Sven.\n\n Kristoff grabs Sven's antlers. Sven pulls them out.\n\n Kristoff helps Anna onto Sven and hops up behind her.\n\n KRISTOFF (CONT'D)\n\n Come on, Olaf!\n\n Sven takes off. Olaf grabs Sven's tail, rides with them.\n\n OLAF\n\n I'm coming! Let's go kiss Hans! Who\n\n is this Hans?!\n\n CUT TO:\n\n EXT. ELSA'S PALACE - DAWN\n\n Hans and the men tread cautiously towards the castle.\n\n HANS\n\n We are here to find Princess Anna.\n\n Be on guard, but no harm is to come\n\n to the Queen. Do you understand?\n\n The Duke's thugs exchange a look. Suddenly, a mass of snow\n\n rises from the ground behind Hans. It's Marshmallow, Elsa's\n\n snow guard.\n\n MARSHMALLOW\n\n Go away!\n\n He slams a fist inches from Hans. Hans deftly dodges out of\n\n the way. All of the guards take up arms against Marshmallow,\n\n who quickly knocks them over.\n\n Marshmallow throws down a guard and his horse, who topple\n\n over Hans. Marshmallow raises his foot to stomp on Hans, but\n\n Hans barrel-rolls himself to safety. He sees his sword,\n\n leaps, and grabs it.\n\n Just then, Elsa peeks out the front doors.\n\n The Duke's two thugs see her.\n\n DUKE'S THUG\n\n The Queen.\n\n The thugs charge up the stairs.\n\n 89\n\nFROZEN - J. Lee\n\n INT. ELSA'S PALACE -- DAY\n\n They guards burst through the ice doors.\n\n Elsa flees to the top floor of her palace. The guards pursue.\n\n They trap her on the top floor, raise their crossbows.\n\n ELSA\n\n (scared)\n\n No. Please.\n\n One of the thugs shoots an arrow right at Elsa. At the last\n\n moment she creates an ice wall. It stops the arrow, inches\n\n from her face.\n\n The thugs reposition to take another shot.\n\n ELSA (CONT'D)\n\n Stay away!\n\n Elsa shoots ice at the thugs. They duck out of the way and\n\n continue the attack.\n\n THUG\n\n Get her! Get her!\n\n Elsa fights for her life.\n\n BACK OUTSIDE: Hans is nearly crushed by Marshmallow. He rolls\n\n away. Jumps to his feet. And with agile might, he slices\n\n Marshmallow's leg off with his sword. Marshmallow stumbles\n\n back, off balance. And falls off over the cliff, but not\n\n before striking Hans. Hans goes over the edge.\n\n REVEAL: Hans clings to the ice steps. His men help him up and\n\n they rush into the ice palace.\n\n INT. ICE PALACE -- DAY\n\n Elsa is surrounded. It's do or die. In two swift moves, Elsa\n\n traps one thug in a cage of spikes that threaten his neck.\n\n The other she pushes back with a wall of ice....up against\n\n the balcony doors...which BURST and CRACK.\n\n OUT ONTO THE BALCONY.... The balcony doors shatter. The thug\n\n is pushed to the edge. He's inches away from falling to his\n\n death.\n\n BACK INSIDE: Hans and his men run in. See the destruction and\n\n the thugs near death.\n\n 90\n\nFROZEN - J. Lee\n\n HANS\n\n Queen Elsa! Don't be the monster\n\n they fear you are.\n\n Elsa snaps out of her rage. She sees the men, frightened,\n\n moments from death. She stops. Elsa looks to Hans,\n\n overwhelmed, frightened.\n\n The wall retreats from the thug on the balcony. The ice\n\n spikes lower from the second thug's neck. He takes advantage\n\n and aims his crossbow at Elsa's back.\n\n Seeing it. Hans runs and pushes the crossbow up just as the\n\n arrow releases. The arrow hits the ice chandelier, hanging\n\n directly above Elsa.\n\n The chandelier comes CRASHING DOWN.\n\n Elsa dives out of the way but she falls in the blast.\n\n All we see is ice smashing like glass, and all we hear is the\n\n sound of it shattering as it rings out.\n\n CUT TO BLACK.\n\n FADE IN ON:\n\n Elsa's face as her eyes flutter open.\n\n She sits up. She's surrounded by stone.\n\n INT. ARENDELLE, DUNGEON -- DAY\n\n Elsa looks to the nearby window. Tries to rush to it. She's\n\n pulled taut by giant shackles that fit like iron gloves.\n\n She's chained to the wall.\n\n Elsa strains to looks out a window...\n\n INSET WINDOW: Arendelle is outside, frozen solid and getting\n\n further buried under the ice and snow that is falling.\n\n ELSA\n\n No....What have I done?\n\n Hans enters. He hangs a torch by the door.\n\n ELSA (CONT'D)\n\n Why did you bring me here?\n\n HANS\n\n I couldn't just let them kill you.\n\n 91\n\nFROZEN - J. Lee\n\n ELSA\n\n But I'm a danger to Arendelle. Get\n\n Anna.\n\n HANS\n\n Anna has not returned....\n\n Elsa looks to the storm with worry.\n\n HANS (CONT'D)\n\n If you would just stop the winter,\n\n bring back summer...please.\n\n Elsa meets his eyes, desperate.\n\n ELSA\n\n Don't you see...I can't.\n\n Hans sees the sincerity in her eyes.\n\n ELSA (CONT'D)\n\n You have to tell them to let me go.\n\n Hans walks to the door. He takes the torch.\n\n HANS\n\n I will do what I can.\n\n He opens the door and leaves.\n\n Elsa, distraught, hears cracking. She looks down as her\n\n shackles begin to freeze over. The storm outside picks up.\n\n CUT TO:\n\n EXT. THE FJORDS -- DAY\n\n Sven charges down the mountain with Kristoff and Anna on his\n\n back. Olaf slides along beside them, penguin-style.\n\n Anna shivers in Kristoff's arms. She's weakening. Kristoff\n\n takes off his hat and puts it on her head.\n\n KRISTOFF\n\n Just hang in there.\n\n (to Sven)\n\n Come on, buddy, faster!\n\n They arrive at the walls of Arendelle. Olaf slides past them,\n\n out of control.\n\n OLAF\n\n I'll meet you guys at the castle!\n\n 92\n\nFROZEN - J. Lee\n\n KRISTOFF\n\n Stay out of sight, Olaf!\n\n OLAF\n\n I will!\n\n He disappears into the village streets.\n\n OLAF (O.S.) (CONT'D)\n\n Hello!\n\n TOWNSWOMAN (O.S.)\n\n Ah! It's alive!\n\n CUT TO:\n\n EXT. CASTLE COURTYARD -- DAY\n\n Guards see Kristoff and Anna approaching.\n\n GUARD\n\n It's Princess Anna!\n\n Sven skids to a stop outside the gates. Kristoff slides off,\n\n holding Anna, and carries her to the gate.\n\n KRISTOFF\n\n I've got you.\n\n Anna looks up at him, gratefully.\n\n ANNA\n\n ...Are you g-gonna be okay?\n\n KRISTOFF\n\n (touched, reassuring)\n\n Don't worry about me.\n\n Just then the castle gates open. Gerda, Kai, and a handmaid\n\n rush to help Anna.\n\n GERDA\n\n Anna! Oh, you had us worried sick.\n\n KAI\n\n My Lady. You are freezing.\n\n GERDA\n\n You poor girl, you're freezing.\n\n Let's get you inside.\n\n 93\n\nFROZEN - J. Lee\n\n KRISTOFF\n\n Get her warm and find Prince Hans,\n\n immediately.\n\n KAI\n\n We will. Thank you.\n\n Anna is swept away from Kristoff and into the palace grounds.\n\n KRISTOFF\n\n Make sure she's safe!\n\n Kristoff is shut out as the castle gates close on him.\n\n Kristoff stands there with Sven for a beat, staring with\n\n worry at the closed gates.\n\n Finally, he sighs, turns and walks off. Sven reluctantly\n\n follows.\n\n CUT TO:\n\n INT. LIBRARY -- DAY\n\n Hans stands with the dignitaries and guards.\n\n HANS\n\n I'm going back out to look for\n\n Princess Anna.\n\n FRENCH DIGNITARY\n\n You cannot risk going out there\n\n again.\n\n HANS\n\n If anything happens to her--\n\n SPANISH DIGNITARY\n\n If anything happens to the\n\n Princess, you are all Arendelle has\n\n left.\n\n Hans hesitates, realizing how much this kingdom has come to\n\n depend on him. Is he really all they have left?\n\n Just then the door opens and Gerda and Kai bring in Anna.\n\n KAI\n\n He's in here. Prince Hans.\n\n HANS\n\n Anna.\n\n 94\n\nFROZEN - J. Lee\n\n Hans rushes to Anna. She falls into his arms.\n\n HANS (CONT'D)\n\n You're so cold.\n\n ANNA\n\n (weak, but desperate)\n\n Hans, you have to kiss me.\n\n HANS\n\n What?\n\n ANNA\n\n Now. Here we go.\n\n She tries to kiss him, but is too weak to pull herself up in\n\n his arms.\n\n GERDA\n\n We'll give you two some privacy.\n\n Everyone shuffles out, leaving Hans and Anna alone.\n\n HANS\n\n What happened out there?\n\n ANNA\n\n Elsa struck me with her powers.\n\n HANS\n\n You said she'd never hurt you.\n\n ANNA\n\n I was wrong.\n\n Anna crumbles, weak.\n\n HANS\n\n Anna.\n\n Hans carries her to a couch, sets her down.\n\n ANNA\n\n (shivering more)\n\n She froze my heart and only an act\n\n of true love can save me.\n\n HANS\n\n (understanding)\n\n A true love's kiss.\n\n He takes her chin in his hand and gives her a tender smile.\n\n He leans in slowly...gently...\n\n 95\n\nFROZEN - J. Lee\n\n Then he stops.\n\n HANS (CONT'D)\n\n Oh, Anna. If only there was someone\n\n out there who loved you.\n\n ANNA\n\n What?\n\n Hans gets up, leaving her there.\n\n ANNA (CONT'D)\n\n ...You said you did.\n\n He goes to the window and shuts the curtains.\n\n HANS\n\n As thirteenth in line in my own\n\n kingdom, I didn't stand a chance. I\n\n knew I'd have to marry into the\n\n throne somewhere--\n\n ANNA\n\n What are you talking about?\n\n HANS\n\n (putting out the candles)\n\n As heir, Elsa was preferable, of\n\n course. But no one was getting\n\n anywhere with her. But you-\n\n ANNA\n\n Hans?\n\n HANS\n\n You were so desperate for love you\n\n were willing to marry me, just like\n\n that.\n\n Hans crosses the room, grabs a pitcher of water from a table\n\n and goes to the fireplace.\n\n HANS (CONT'D)\n\n I figured, after we married, I'd\n\n have to stage a little accident for\n\n Elsa.\n\n Hans pours the water on the fireplace, putting out the fire.\n\n Anna tries to stop him. She falls to the floor, weak.\n\n ANNA\n\n Hans. No, stop.\n\n 96\n\nFROZEN - J. Lee\n\n HANS\n\n But then she doomed herself, and\n\n you were dumb enough to go after\n\n her.\n\n ANNA\n\n Please.\n\n HANS\n\n (chuckles)\n\n All that's left now is to kill Elsa\n\n and bring back summer.\n\n Hans approaches Anna.\n\n ANNA\n\n ...You're no match for Elsa.\n\n He bends down, takes her chin in his hand again, this time\n\n not so gently.\n\n HANS\n\n No, you're no match for Elsa. I, on\n\n the other hand, am the hero who is\n\n going to save Arendelle from\n\n destruction.\n\n She wrenches her face out of his hands.\n\n ANNA\n\n (anger)\n\n You won't get away with this.\n\n Hans rises and crosses to the door.\n\n HANS\n\n Oh, I already have.\n\n Hans leaves and shuts her in, locking the door. Anna\n\n struggles to the door, yanks on the locked handle.\n\n ANNA\n\n (hoarse and weak)\n\n Please, somebody help.\n\n The rest of her hair turns white and she crumbles to the\n\n floor.\n\n CUT TO:\n\n 97\n\nFROZEN - J. Lee\n\n INT. COUNCIL CHAMBER -- NIGHT\n\n The Duke looks out the window at the growing snowstorm. He\n\n rubs his arms and shivers.\n\n DUKE\n\n It's getting colder by the minute.\n\n If we don't do something soon,\n\n we'll all freeze to death.\n\n Hans comes in, putting on his most distraught face.\n\n SPANISH DIGNITARY\n\n Prince Hans.\n\n HANS\n\n Princess Anna is...dead.\n\n VARIOUS DIGNITARIES\n\n What...? No.... Mon dieu.\n\n Hans stumbles, weak with grief. The men help him to a chair.\n\n DUKE\n\n What happened to her?\n\n HANS\n\n She was killed by Queen Elsa.\n\n DUKE\n\n Her own sister.\n\n HANS\n\n (really putting it on)\n\n At least we got to say our marriage\n\n vows...before she died in my arms.\n\n He bows his head in a brilliant display of teary grief.\n\n DUKE\n\n There can be no doubt now; Queen\n\n Elsa is a monster and we are all in\n\n grave danger.\n\n SPANISH DIGNITARY\n\n Prince Hans, Arendelle looks to\n\n you.\n\n Hans nods; he knows what he's being asked to do, and he'll do\n\n it with the perfect amount of authority and gravitas.\n\n 98\n\nFROZEN - J. Lee\n\n HANS\n\n With a heavy heart, I charge Queen\n\n Elsa of Arendelle with treason and\n\n sentence her to death.\n\n INT. ELSA'S DUNGEON -- DAY\n\n The cell ices over. Elsa looks out at the storm that is\n\n devastating Arendelle, then hears the guards approaching.\n\n GUARD (O.S.)\n\n She's dangerous. Move quickly and\n\n with resolve.\n\n Elsa pulls at her shackles. They crack. Just as the door\n\n busts open, the weight of the ice crumbles the walls. The men\n\n duck out of the way.\n\n Hans pushes his way into the room...sees...\n\n The back wall is blown open. Broken shackles rest on the\n\n floor. Elsa is gone.\n\n CUT TO:\n\n EXT. MOUNTAIN SLOPE -- DAY\n\n Kristoff heads into the mountains. Sven lags behind, not\n\n wanting to follow. He looks back at the kingdom, then shakes\n\n his head. Enough.\n\n He runs past Kristoff. Stops and turns to face him. He snorts\n\n and grunts.\n\n KRISTOFF\n\n What is it, buddy?\n\n Sven nudges Kristoff with his antlers.\n\n KRISTOFF (CONT'D)\n\n Hey, watch it. What's wrong with\n\n you?\n\n Sven snorts with more conviction, moos, brays.\n\n KRISTOFF (CONT'D)\n\n (avoiding)\n\n ...I don't understand you when you\n\n talk like that.\n\n 99\n\nFROZEN - J. Lee\n\n Kristoff tries to walk on ahead, but Sven uses his antlers to\n\n lift Kristoff off the ground.\n\n KRISTOFF (CONT'D)\n\n Ah! Stop it! Put me down!\n\n Sven drops him hard then "yells" at him once more.\n\n KRISTOFF (CONT'D)\n\n No, Sven! We're not going back!\n\n Sven shakes his head, angrily.\n\n KRISTOFF (CONT'D)\n\n She's with her true love.\n\n Sven makes an "of-course-she-isn't" face. Kristoff gets it;\n\n he's made his point.\n\n Just then the wind picks up. Kristoff looks back at the\n\n kingdom. Sees a violent winter storm swirling over the\n\n castle. Sharp ice claws its way up the castle, encasing it.\n\n KRISTOFF (CONT'D)\n\n Anna.\n\n Without hesitating, he dashes back down the mountain. Sven\n\n runs after him, catches up. Kristoff grabs Sven's harness and\n\n jumps onto his back.\n\n CUT TO:\n\n INT. LIBRARY -- NIGHT\n\n Anna shivers by the door. She looks up to see ice overtaking\n\n the ceiling.\n\n The door handle suddenly jiggles. Stops. Jiggles again.\n\n ANNA\n\n (barely a whisper)\n\n Help.\n\n CLICK. The door swings open. We see a carrot in the lock and\n\n hear a giggle of victory. Olaf takes the carrot, puts it back\n\n on his face. Then he sees Anna lying there.\n\n OLAF\n\n Anna. Oh no.\n\n He runs to the fireplace. Throws in some fresh wood,\n\n including one of his own arms, which he quickly rescues,\n\n before striking a match and relighting the fire.\n\n 100\n\nFROZEN - J. Lee\n\n ANNA\n\n Olaf? Olaf. Get away from there.\n\n OLAF\n\n Whoa! So this is heat....\n\n (considering)\n\n I love it.\n\n He reaches a twig finger toward the flames. It catches on\n\n fire.\n\n OLAF (CONT'D)\n\n Ooh! But don't touch it!\n\n He shakes the flame out, as he rushes over to help Anna to\n\n the fire.\n\n OLAF (CONT'D)\n\n So, where's Hans? What happened to\n\n your kiss?\n\n ANNA\n\n I was wrong about him. It wasn't\n\n true love.\n\n OLAF\n\n (confused innocence)\n\n Huh. But we ran all the way here?\n\n ANNA\n\n Please Olaf, you can't stay here;\n\n you'll melt.\n\n OLAF\n\n I am not leaving here until we find\n\n some other act of true love to save\n\n you.\n\n He sits down behind her, stubbornly. Leans his back against\n\n hers and thinks.\n\n OLAF (CONT'D)\n\n ...Do you happen to have any ideas?\n\n ANNA\n\n I don't even know what love is.\n\n OLAF\n\n (confident)\n\n That's okay, I do....\n\n Olaf hops back up and puts a soothing hand on her shoulder.\n\n 101\n\nFROZEN - J. Lee\n\n OLAF (CONT'D)\n\n Love is...putting someone else's\n\n needs before yours, like, you know,\n\n how Kristoff brought you back here\n\n to Hans and left you forever.\n\n ANNA\n\n ...Kristoff loves me?\n\n OLAF\n\n Wow, you really don't know anything\n\n about love, do you?\n\n His face starts to melt.\n\n ANNA\n\n Olaf, you're melting.\n\n OLAF\n\n (sweet and reassuring)\n\n Some people are worth melting for.\n\n But then...his face REALLY melts. He panics, pushes the snow\n\n back in place.\n\n OLAF (CONT'D)\n\n Just maybe not right this second.\n\n Suddenly, the window blows open, cold wind sweeps in.\n\n OLAF (CONT'D)\n\n Don't worry, I've got it!\n\n Olaf flitters to the window. He pulls one panel of it shut\n\n but struggles with the second panel.\n\n OLAF (CONT'D)\n\n (determined)\n\n We're going to get through--\n\n (distracted)\n\n Oh, wait. Hang on. I'm getting\n\n something.\n\n He breaks an icicle off the window, uses it as a telescope\n\n and sees...\n\n Kristoff and Sven running back down the mountain.\n\n OLAF (CONT'D)\n\n It's Kristoff and Sven! They're\n\n coming back this way.\n\n ANNA\n\n ...They-they are?\n\n 102\n\nFROZEN - J. Lee\n\n OLAF\n\n Wow, he's really moving fast.\n\n Huh.... I guess I was wrong. I\n\n guess Kristoff doesn't love you\n\n enough to leave you behind.\n\n Anna tries to get to her feet.\n\n ANNA\n\n Help me up, Olaf. Please.\n\n He hurries over, tumbling over the couch, knocking over the\n\n chess set and water jugs.\n\n OLAF\n\n No, no, no, no, no. You need to\n\n stay by the fire and keep warm.\n\n ANNA\n\n I need to get to Kristoff.\n\n OLAF\n\n (clueless)\n\n Why...?\n\n (realizing)\n\n Oh, oh, oh, I know why.\n\n He hops around in an excited display of hope.\n\n OLAF (CONT'D)\n\n There's your act of true love,\n\n right there, riding across the\n\n fjords like a valiant, pungent\n\n reindeer king! Come on!\n\n The walls crack under the ice pressure.\n\n OLAF (CONT'D)\n\n Look out!\n\n They rush out the room just as the ceiling collapses.\n\n INT. CASTLE HALLWAY -- DAY\n\n Anna and Olaf struggle down the hall. Ice spikes grow and\n\n block their path.\n\n OLAF\n\n We're trapped.\n\n Anna looks around desperately for a way out.\n\n 103\n\nFROZEN - J. Lee\n\n EXT. FJORD -- DAY\n\n Elsa runs, but is nearly blinded by the snow and wind.\n\n EXT. CASTLE -- DAY\n\n Anna and Olaf bust open a window. The storm is so strong it\n\n sweeps the window panes away.\n\n OLAF\n\n Slide, Anna.\n\n It's a long, snowy way down. But what choice do they have?\n\n They slide down the iced-covered building.\n\n Anna arrives at the bottom, weak but uninjured. Olaf gathers\n\n snow along the way. He arrives at the bottom as a giant\n\n snowball.\n\n OLAF (CONT'D)\n\n We made it!\n\n He shakes off the extra snow as Anna struggles to her feet.\n\n EXT. FJORD -- DAY\n\n Kristoff and Sven bound off the mountain and sprint across\n\n the frozen fjord waters and right into the heart of the\n\n storm. Its white-out wind pushes them back. But they fight\n\n through.\n\n KRISTOFF\n\n Come on, buddy, faster.\n\n CUT TO:\n\n Anna and Olaf reach the shore of the fjords.\n\n ANNA\n\n Kristoff!\n\n The wind lifts Olaf up and pulls him apart. He goes swirling\n\n off into the storm.\n\n OLAF\n\n Keep going, Anna!\n\n Anna struggles on.\n\n 104\n\nFROZEN - J. Lee\n\n ANNA\n\n Kristoff!\n\n PAN TO:\n\n Kristoff rides Sven past cracking, frozen ships. Sven\n\n struggles over the uneven surface.\n\n KRISTOFF\n\n Come on! Come on!\n\n Suddenly, a mangled ship, risen by ice, capsizes over them.\n\n They give it all they've got as debris falls all around them\n\n and the mast shatters. They make it past just as the entire\n\n ship slams down and cracks the thick ice beneath their feet.\n\n The ice opens up. Sven bravely jumps over a gap. But it's too\n\n wide. He bucks Kristoff to safety, but lands in the freezing\n\n water and disappears below.\n\n KRISTOFF (CONT'D)\n\n Sven? Sven!\n\n At first there's nothing but the wind and the tumbling icy\n\n water. But suddenly, Sven surfaces and claws his way to a\n\n floating ice chunk. He calls out, signalling for Kristoff to\n\n go on.\n\n KRISTOFF (CONT'D)\n\n Good boy.\n\n CUT TO:\n\n Anna moves blindly across the fjord. Anna's hands frost over\n\n an icy blue. She stumbles on, determined. But she's running\n\n out of time.\n\n She clutches her chest. The color in her eyes fades, the\n\n inevitable is coming.\n\n CUT TO:\n\n Kristoff, lost in the white-out, doesn't know which way to\n\n turn. But then he hears a faint--\n\n ANNA (O.S.)\n\n Kristoff.\n\n KRISTOFF\n\n Anna...? Anna!\n\n WHITE OUT TO:\n\n 105\n\nFROZEN - J. Lee\n\n Elsa struggles through her own storm, but the fear is\n\n consuming her. A dark shadow approaches. It's Hans.\n\n HANS\n\n Elsa. You can't run from this!\n\n Elsa backs away from him.\n\n ELSA\n\n ...Just take care of my sister.\n\n HANS\n\n Your sister? She returned from the\n\n mountain weak and cold. She said\n\n you froze her heart.\n\n ELSA\n\n What? No.\n\n HANS\n\n I tried to save her, but it was too\n\n late. Her skin was ice. Her hair\n\n turned white...\n\n Elsa's face sinks as she realizes what she has done.\n\n HANS (CONT'D)\n\n Your sister is dead... because of\n\n you.\n\n Elsa drops to her knees, emotionally broken. And with that,\n\n the swirling storm suddenly stops. The snow freezes mid-air,\n\n hangs suspended, trapped in grief.\n\n Citizens and dignitaries rush to the wall's edge and look out\n\n to see...\n\n Anna, barely able to move but now able to see across the\n\n fjords to...\n\n ANNA\n\n (a whisper)\n\n Kristoff.\n\n KRISTOFF\n\n Anna.\n\n Anna pushes on towards Kristoff. He runs top speed towards\n\n her. There's still a lot of fjord to cross, but Kristoff is\n\n giving it all he's got. He's going to make it.\n\n But then, Anna hears the sound of a sword being drawn from\n\n its scabbard. She turns and sees Hans, behind Elsa, as he\n\n raises his sword over his head.\n\n 106\n\nFROZEN - J. Lee\n\n ANNA\n\n Elsa.\n\n Anna looks back at Kristoff as he runs for her. She gives him\n\n a longing look, but then turns away from him and then...\n\n Using all of her remaining strength, as Hans brings his sword\n\n down, Anna throws herself in front of Elsa.\n\n ANNA (CONT'D)\n\n No!\n\n In that instant, Anna freezes to solid ice. The sword hits\n\n her instead of Elsa. The sword shatters completely. The force\n\n of it sends Hans flying back and knocks him out.\n\n ELSA\n\n Anna!\n\n Elsa rushes to Anna and touches her sister's frozen face.\n\n ELSA (CONT'D)\n\n Oh, Anna...no...no, please no.\n\n Olaf walks up and sees Anna, frozen.\n\n OLAF\n\n (confused, sad)\n\n Anna?\n\n Elsa hugs Anna and cries.\n\n Kristoff watches in shocked despair. Sven steps up to his\n\n side.\n\n Citizens and dignitaries on the castle walls bow their heads.\n\n All of Arendelle is joined in somber silence.\n\n But then, Anna warms. She begins to thaw.\n\n Olaf looks up and gasps. Kristoff and Sven notice, light up.\n\n Anna bends her arm and embraces Elsa.\n\n ELSA\n\n Wha-? Anna?\n\n Anna opens her eyes. She smiles at Elsa, relieved.\n\n ANNA\n\n Oh, Elsa.\n\n They embrace.\n\n 107\n\nFROZEN - J. Lee\n\n ELSA\n\n ...You sacrificed yourself for me?\n\n ANNA\n\n (weak)\n\n ...I love you.\n\n Olaf realizes what's happened. He's so excited about it, he\n\n lifts his head right off his body and exclaims--\n\n OLAF\n\n An act of true love will thaw a\n\n frozen heart.\n\n ELSA\n\n (processing)\n\n Love...will thaw...\n\n (realizing)\n\n Love.... Of course.\n\n Elsa looks at Anna with confidence.\n\n ANNA\n\n Elsa?\n\n ELSA\n\n Love.\n\n Elsa lifts her arms, and the ground shakes and cracks. The\n\n ice and snow breaks away and rises high into the air.\n\n Beneath their feet the bow of a ship thaws.\n\n The entire fjord melts and other boats right themselves.\n\n The villagers come out to see the warmth returning.\n\n In one final wave, Elsa draws all of the snow into a giant\n\n snowflake in the sky, then waves it away, leaving only a warm\n\n summer day.\n\n ANNA\n\n I knew you could do it.\n\n OLAF\n\n (melting, good-naturedly)\n\n Hands down, this is the best day of\n\n my life...and quite possibly the\n\n last.\n\n ELSA\n\n Oh, Olaf. Hang on, little guy.\n\n 108\n\nFROZEN - J. Lee\n\n Elsa waves her hand and surrounds Olaf with a swirl of cold\n\n air. He refreezes. Above his head she leaves a little,\n\n perpetually-snowing storm cloud. Olaf loves it.\n\n OLAF\n\n Hey, my own personal flurry.\n\n Kristoff sees Hans trying to get to his feet. He marches\n\n toward him, prepared for a fight. But Anna puts up a hand and\n\n stops him.\n\n ANNA\n\n Uh. Uh. Uh.\n\n She'll handle this. She goes over to Hans.\n\n HANS\n\n (confused)\n\n Anna? But she froze your heart.\n\n ANNA\n\n The only frozen heart around here\n\n is yours.\n\n She turns away from him, proud of her words. But not yet\n\n satisfied, she turns back and punches him right in the face.\n\n HANS\n\n Ah! Whoa, whoa, whoa!\n\n He falls overboard.\n\n Elsa comes over to Anna and hugs her. Over her shoulder,\n\n Kristoff meets Anna's eyes. She smiles brighter, happy.\n\n DISSOLVE TO:\n\n EXT. ARENDELLE -- DAY\n\n It's a beautiful summer day. The mighty ships have been\n\n repaired and are sailing away.\n\n On one of the ships, HANS is thrown into a brig.\n\n FRENCH DIGNITARY\n\n (to Kai)\n\n I will return this scoundrel to his\n\n country. We shall see what his\n\n twelve big brothers think of his\n\n behavior.\n\n KAI\n\n Arendelle thanks you, my Lord.\n\n 109\n\nFROZEN - J. Lee\n\n Down on the dock, Arendelle guards lead the Duke and his two\n\n thugs to their ship.\n\n DUKE\n\n This is unacceptable. I am\n\n innocent. I'm a victim of fear.\n\n I've been traumatized.\n\n (bad acting)\n\n Ow! My neck hurts. Is there a\n\n doctor I could...No? And I demand\n\n to see the Queen!\n\n Kai steps down from the gangplank to the dock.\n\n KAI\n\n I have a message from the Queen.\n\n (reading a scroll)\n\n Arendelle will henceforth and\n\n forever no longer do business of\n\n any sort with Weaseltown.\n\n DUKE\n\n Weselton. It's Weselton!\n\n The guards usher him and his thugs onto their ship.\n\n EXT. VILLAGE SQUARE -- DAY\n\n Anna runs through the crowd, pulling a blindfolded Kristoff\n\n along behind her. She's so excited she can't stand it.\n\n ANNA\n\n Come on. Come on. Come on. Come on!\n\n She runs him right into a pole.\n\n KRISTOFF\n\n Pole.\n\n ANNA\n\n Oops. Sorry.\n\n EXT. ARENDELLE DOCKS -- DAY\n\n Anna skips to the perfect spot and stops.\n\n ANNA\n\n (stopping)\n\n Okay. Okay. Here we are.\n\n 110\n\nFROZEN - J. Lee\n\n She takes off the blindfold. Kristoff opens his eyes. Before\n\n him sits the most beautiful, suped-up sled. Sven poses in\n\n front of it -- Vanna White-style.\n\n ANNA (CONT'D)\n\n I owe you a sled.\n\n KRISTOFF\n\n (blown away)\n\n Are you serious?\n\n ANNA\n\n Yes. And it's the latest model.\n\n KRISTOFF\n\n No. I can't accept this...\n\n ANNA\n\n You have to. No returns. No\n\n exchanges. Queen's orders. She's\n\n named you the official Arendelle\n\n Ice Master and Deliverer.\n\n Sven shows off the Ice-Master-and-Deliverer medal like he's\n\n king of the bucks.\n\n KRISTOFF\n\n What? That's not a thing.\n\n But he can't help but admire her enthusiasm.\n\n ANNA\n\n Sure it is. And it even has a cup\n\n holder.... Do you like it?\n\n KRISTOFF\n\n Like it?\n\n He sweeps her up high overhead and spins her around.\n\n KRISTOFF (CONT'D)\n\n I love it.... I could kiss you!\n\n He drops her, suddenly embarrassed.\n\n KRISTOFF (CONT'D)\n\n ...I could. I mean I'd like to.\n\n I'd... may I? We me....I mean, may\n\n we? Wait, what?\n\n She gives him a quick kiss on the cheek.\n\n ANNA\n\n We may.\n\n 111\n\nFROZEN - J. Lee\n\n He smiles and goes for it. It's a true love's kiss, alright.\n\n We move past them to find Olaf enjoying the summer.\n\n With his snow cloud safely overhead, he's free to smell the\n\n flowers, which he does. Then sneezes his carrot nose off.\n\n Sven catches it between his teeth. Olaf gasps as Sven sucks\n\n the whole carrot into his mouth. It's gone.\n\n Olaf's face sinks in sadness. But not to fear, Sven spits the\n\n carrot back out and jams it into Olaf's face where it\n\n belongs. It's completely covered in reindeer spit, but Olaf\n\n doesn't seem to mind. He hugs Sven happily.\n\n CUT TO:\n\n EXT. CASTLE COURTYARD -- DAY\n\n The gates to the castle are wide open. In the courtyard,\n\n stands Elsa.\n\n ELSA\n\n Are you ready?\n\n Villagers cheer. Elsa stops and creates an ice rink. The\n\n people, skates at the ready, hope onto it and twirl about.\n\n Elsa then freezes the fountain in a beautiful design and adds\n\n some snow flurries for atmosphere.\n\n Anna comes slipping in. Elsa catches her.\n\n ANNA\n\n I like the open gates.\n\n ELSA\n\n We are never closing them again.\n\n Elsa then waves her hand and magical ice skates (literally\n\n made of ice) form on Anna's boots.\n\n ANNA\n\n What? Oh, Elsa, they're beautiful,\n\n but you know I don't ska--\n\n Elsa grabs Anna's hands and pulls her along on the ice. Anna\n\n slips and slides, but laughs in delight.\n\n Sven goes slipping past. Kristoff runs after him.\n\n KRISTOFF\n\n Look out. Reindeer coming through!\n\n 112\n\nFROZEN - J. Lee\n\n Olaf skates and helps Elsa coach Anna.\n\n OLAF\n\n That's it. Glide and pivot and\n\n glide and pivot.\n\n We pull away slowly, into the sky. We arrive at a bird's-eye\n\n view to see that where the castle had crumbled has been\n\n repaired with a ice.\n\n All is right in Arendelle.\n\n FINAL FADE OUT.\n\n THE END\n\n |
| 16 | script_strangersonatrain.txt | 9e6bfae4-7c2 | Script | STRANGERS ON A TRAIN\n\n by\n\n Raymond Chandler and Czenzi Ormonde\n\nFINAL DRAFT\n\nOctober 18, 1950\n\nConverted to PDF by SCREENTALK\n\n FOR EDUCATIONAL PURPOSES ONLY\n\nwww.screentalk.org\n\nFADE IN:\n\nEXT. UNION STATION, WASHINGTON, D.C. DAY\n\nLONG SHOT THE CAPITOL DOME IN THE B.G. AND THE AUTOMOBILE\n\nENTRANCE TO THE STATION IN THE F.G. LOW CAMERA\n\nActivity of cars and taxis arriving and discharging passengers\n\nwith luggage, busy redcaps, etcetera.\n\nWe FOCUS on a taxi pulling up and stopping, The driver hands\n\nout modest looking luggage, including a bunch of tennis\n\nrackets in cases to a redcap. CAMERA PANS DOWN as the\n\npassenger gets out of the taxi so that we see only his shoes\n\nand the lower part of his trousers. He is wearing dark\n\ncolored brogues and a conservative suit apparently. The\n\nfeet move toward, the entrance to the station and out of\n\nscene. Immediately a chauffeur-driven limousine drives up\n\nand an expensive place of airplane luggage is handed out of\n\nthis, and the passenger alighting from the back is seen to\n\nbe wearing black and white sport shoes which, as before, are\n\nall we see of him. The sport shoes start off in the wake of\n\nthe brogues.\n\nINT. STATION LOBBY\n\nCAMERA FOLLOWS the sport shoes and the brogues across the\n\nlobby into a passenger tunnel. There is the usual activity\n\nof passengers walking to and from, a loud-speaker announcing\n\ntrains, etc.\n\nEXT. PASSENGER TUNNEL\n\nAs the brogues and the sport shoes emerge to the train\n\nplatform, CAMERA PANS them over to the steps of the train.\n\nINT. TRAIN\n\nThe brogues and the sport shoes pass separately down the\n\naisle, the sport shoes turning in at a compartment door and\n\nthe brogues continuing toward the parlor car.\n\n DISSOLVE TO:\n\nINT. PARLOR CAR (PROCESS)\n\nThe brogues come to rest before a chair as the owner sits\n\ndown. A moment later the sport shoes come to rest. before\n\nin adjoining chair.\n\n Converted to PDF by www.screentalk.org 2.\n\nThe legs belonging to the sport shoes stretch out, and one\n\nof the shoes touches one of the brogues.\n\n MAN'S VOICE (over scene)\n\n Oh, excuse Me!\n\nCAMERA PULLS BACK AND UP to SHOW two young men seated in two\n\nparlor car chairs. BRUN0 ANTHONY, the wearer of the sport\n\nshoes, is about twenty-five. He wears his expensive clothes\n\nwith the tweedy nonchalance of a young man who has always\n\nhad the best. The wearer of the brogues is a fine looking\n\nbut, at the moment, a somewhat troubled young man. This is\n\nGUY HAINES. He, too, is in his middle twenties and is well\n\ndressed because he can now afford to be. He nods politely,\n\nacknowledging Bruno's apology, then turns away with the\n\ngesture implying he wants privacy.\n\n BRUNO\n\n (smiling with sudden\n\n recognition)\n\n I beg your pardon, but aren't you\n\n Guy Haines.\n\nGuy nods with a polite half smile. Being a well known\n\ntournament tennis player, he has had this sort of experience\n\nbefore.\n\n BRUNO\n\n (snapping his finger)\n\n Sure! I saw you blast Faraday right\n\n off the court in South Orange last\n\n season. What a backhand! Made the\n\n semi-finals, didn't you?\n\nGuy acknowledges this with a modest nod and turns to his\n\nmagazine rolled up in is fist.\n\n BRUNO\n\n (with open admiration)\n\n I certainly admire people who do\n\n things.\n\n (smiling and\n\n introducing himself)\n\n I'm Bruno Anthony. Bruno. See Guy\n\n looks up. Bruno indicates his gold\n\n tie pin which bears his name in cut-\n\n out letters. Guy looks at it with\n\n the faintest expression of disdain.\n\n I suppose you think it's corny. But\n\n my mother gave it to me so of course\n\n I wear it to please her.\n\n Converted to PDF by www.screentalk.org 3.\n\n GUY\n\n (patiently)(a faint\n\n smile)\n\n How do you do.\n\n BRUNO\n\n (with an apologetic\n\n grin)\n\n I don't usually talk so much. Go\n\n Ahead and read.\n\n GUY\n\n (wryly)\n\n Thanks.\n\nGuy tries to read but is uneasily aware of Bruno's open\n\nappraisal.\n\n BRUNO\n\n It must be pretty exciting to be so\n\n important.\n\n GUY\n\n (fidgeting slightly)\n\n A tennis player isn't so important.\n\n BRUNO\n\n People who do things are important.\n\n I never seem to do anything.\n\nNot knowing how to answer this, Guy looks a little\n\nembarrassed.\n\n BRUNO\n\n (still insistent on\n\n being friendly)\n\n I suppose you're going to Southampton --\n\n for the doubles.\n\n GUY\n\n (politely)\n\n You are a tennis fan.\n\nBruno is inordinately pleased by this small tribute.\n\n BRUNO\n\n Wish I could see you play. But I've\n\n got to be back in Washington tomorrow.\n\n I live in Arlington, you know.\n\nHe has taken out a cigarette case. Holds it out to Guy.\n\n Converted to PDF by www.screentalk.org 4.\n\n BRUNO\n\n Cigarette?\n\n GUY\n\n Not now, thanks. I don't smoke much.\n\n BRUNO\n\n I smoke too much.\n\nHe fumbles for a match. Guy brings out a lighter and hands\n\nit to Bruno.\n\n BRUNO\n\n Thanks.\n\n (he stares at the\n\n lighter, impressed)\n\n Elegant.\n\nCLOSE SHOT OF THE LIGHTER\n\nShowing that it has the insignia of crossed rackets embossed\n\non it, and underneath is engraved the inscription: "To G\n\nfrom A".\n\n BRUNO'S VOICE\n\n (reading)\n\n To G from A. Bet I can guess who A\n\n is.\n\nWIDER SHOT\n\nGuy reacts sharply.\n\n GUY\n\n (coldly)\n\n Yes?\n\n BRUNO\n\n Anne Burton. Sometimes I turn the\n\n sport page and look at the society\n\n news. And the pictures. She's very\n\n beautiful, Senator Burton's daughter.\n\n GUY\n\n You're quite a reader, Mr. Anthony.\n\n BRUNO\n\n Yes, I am. Ask me anything, from\n\n today's stock reports to Li'l Abner,\n\n and I got the answer.\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 5.\n\n BRUNO (CONT'D)\n\n Even news about people I don't know.\n\n Like who'd like to marry whom when\n\n his wife gets her divorce.\n\n GUY\n\n (sharply)\n\n Perhaps you read too much.\n\n BRUNO\n\n (contritely)\n\n There I go again. Too friendly. I\n\n meet someone I' like and open my yap\n\n too wide. I'm sorry...\n\nAt the appeal on Bruno's face, Guy slowly relents.\n\n GUY\n\n That's all right. Forget it. I\n\n guess I'm pretty jumpy.\n\nBruno smiles with and signals a waiter.\n\n BRUNO\n\n There's a new cure for that.\n\n (to waiter)\n\n Scotch and plain water. A pair.\n\n Double.\n\n (to Guy with a chuckle)\n\n Only kind of doubles I play.\n\n GUY\n\n You'll have to drink both of them.\n\n BRUNO\n\n (grinning)\n\n And I can do it.\n\n (moving in)\n\n When's the wedding?\n\n GUY\n\n What?\n\n BRUNO\n\n The wedding. You and Anne Burton.\n\n (a gesture of\n\n explanation)\n\n It was in the papers.\n\n GUY\n\n It shouldn't have been. Unless\n\n they've legalized bigamy overnight.\n\n Converted to PDF by www.screentalk.org 6.\n\n BRUNO\n\n I have a theory about that. I'd\n\n like to tell you about it some time.\n\n But right now I suppose divorce Is\n\n still the simplest operation.\n\nThe waiter has brought the drinks. Bruno slips the lighter\n\ninto hip pocket to free his hands for the bills which he\n\ngives to the waiter, waving away the change. He offers a\n\nglass to Guy. Guy takes it.\n\n GUY\n\n (as if he needs it)\n\n I guess I will.\n\n BRUNO\n\n (happily)\n\n This is wonderful -- having your\n\n company all the way to New York.\n\n GUY\n\n (forced to explain)\n\n As a matter of fact, I'm not going\n\n direct. I'm stopping off. At\n\n Metcalf.\n\n BRUNO\n\n Metcalf? What would anybody want to\n\n go there for?\n\n GUY\n\n It's my home town.\n\n BRUNO\n\n Oh, I get it! A little talk with\n\n your wife to about the divorce! I\n\n suppose she was the girl next door.\n\n Held her hand in high school and\n\n before you knew it -- hooked!\n\n (proud of his\n\n perspicacity)\n\n Am I right?\n\n GUY\n\n (laconically)\n\n Close enough.\n\n BRUNO\n\n (raises his glass)\n\n Well, here's luck, Guy. Drink up --\n\n then we'll have some lunch sent to\n\n my compartment.\n\n Converted to PDF by www.screentalk.org 7.\n\n GUY\n\n Thanks very much. But I think I'll\n\n go to the dining car.\n\n (he hails a waiter\n\n who is passing through\n\n with a food-laden\n\n tray)\n\n Do you know if there are any vacant\n\n seats in the dining car now?\n\n WAITER\n\n Not for about twenty minutes I'm\n\n afraid, Sir.\n\n BRUNO\n\n (pleased)\n\n See? You'll have to lunch with me.\n\n (motions the waiter\n\n back)\n\n Say, waiter, bring me some lamb chops\n\n and French fries and chocolate ice\n\n cream, Compartment D, Car 121.\n\n (turns to Guy)\n\n What'll you have, Guy?\n\n GUY\n\n Thanks just the same, but I really\n\n don't think --\n\n BRUNO\n\n Oh, go on and order.\n\nThe waiter is hovering impatiently. Guy gives in out of\n\nembarrassment.\n\n GUY\n\n Well, I'll Just have a hamburger and\n\n a cup of coffee.\n\n BRUNO\n\n (delighted, lifts his\n\n glass in another\n\n toast)\n\n To the next Mrs. Haines.\n\nGuy nods curtly.\n\n DISSOLVE TO:\n\n Converted to PDF by www.screentalk.org 8.\n\nINT. BRUNO'S COMPARTMENT ON TRAIN (PROCESS)\n\nBruno and Guy are finishing lunch. Bruno has been drinking\n\nand his eyes are bright and feverish. An almost empty liquor\n\nbottle is near a couple of detective novels covered with\n\ngaudily Illustrated dust jackets. Bruno has in unlighted\n\ncigarette in his mouth. Guy's lighter is on the table.\n\nBruno snaps it a couple of times, as though fascinated, lights\n\nhis cigarette and puts the lighter on the table again.\n\n BRUNO\n\n Sure, I went to college. Three of\n\n them. Every time they kicked me out\n\n my father threw me back in.\n\n (bitterly)\n\n He finally gave up. He thinks I'm\n\n awfully small fry, not worth the\n\n bait.\n\n (wistfully)\n\n You my friend, Guy?\n\n GUY\n\n Sure. I'm your friend, Bruno.\n\n BRUNO\n\n (a little woozy)\n\n No, you're not, nobody thinks I'm\n\n anything special. Only my mother.\n\n (empties the bottle\n\n into his glass)\n\n My father hates me.\n\nGuy smiles this off as nonsense.\n\n GUY\n\n You must be imagining things.\n\n BRUNO\n\n (hitting the bottom\n\n of the bottle for\n\n the last drop)\n\n And I hate him. He thinks I ought\n\n to catch the eight-five bus every\n\n morning, punch a timeclock and work\n\n my way up selling paint or something.\n\n Him -- with all his money!\n\n GUY\n\n (amused by Bruno)\n\n Well, what do you want to do?\n\n BRUNO\n\n You mean before or after I kill him?\n\nConverted to PDF by www.screentalk.org 9.\n\n GUY\n\n (chuckling)\n\n Before, of course.\n\n BRUNO\n\n (leaning forward\n\n eagerly)\n\n I want to do everything. I got a\n\n theory you're supposed to do\n\n everything before you die. Have you\n\n ever driven a car, blindfolded, at a\n\n hundred and fifty miles an hour?\n\n GUY\n\n Not lately.\n\n BRUNO\n\n I did. I flew in a jet plans too.\n\n (his hand traces a\n\n swift streak through\n\n the air, and he adds\n\n sound effects)\n\n Zzzzzzzp! Man, that's a thrill!\n\n Almost blow the sawdust out of my\n\n head. I'm going to make a reservation\n\n on the first rocket to the moon...\n\n GUY\n\n (amused and curious)\n\n What are you trying prove?\n\n BRUNO\n\n I'm not like you, Guy. You're lucky.\n\n You're smart. Marrying the boss's\n\n daughter is a nice short cut to a\n\n career, isn't it?\n\n GUY\n\n (quickly)\n\n Marrying the senator's daughter has\n\n nothing to do with it. Can't a fellow\n\n look past a tennis not without being\n\n a goldbricker?\n\n BRUNO\n\n Take it easy, boy. I'm your friend,\n\n remember? I'd do anything for you.\n\n GUY\n\n (humoring Bruno)\n\n Sure, Bruno, sure.\n\n (MORE)\n\nConverted to PDF by www.screentalk.org 10.\n\n GUY (CONT'D)\n\n (glancing at his watch)\n\n We'll be pulling in soon. I've got\n\n to change trains.\n\n BRUNO\n\n What'd you say her name was -- your\n\n wife's?\n\n GUY\n\n Miriam.\n\n BRUNO\n\n That's it. Miriam Joyce Haines.\n\n Played around a lot, I suppose?\n\n GUY\n\n Let's not talk about it any more.\n\n BRUNO\n\n (almost hopefully)\n\n Maybe she'll make more trouble for\n\n you.\n\n GUY\n\n I don't think so.\n\n BRUNO\n\n You mean you got enough on her to\n\n get your divorce no matter what?\n\n GUY\n\n Let's change subject, Bruno, can't\n\n we?\n\n BRUNO\n\n Okay, Guy. Want me to tell you one\n\n of my ideas for murdering my father?\n\n GUY\n\n (indicating the\n\n detective novels)\n\n You've been reading too many of these.\n\n BRUNO\n\n (going right on)\n\n You want to hear about the busted\n\n light socket in the bathroom, or the\n\n carbon monoxide in the garage?\n\n GUY\n\n No. I may be old fashioned, but I\n\n thought murder was against the law.\n\nConverted to PDF by www.screentalk.org 11.\n\n BRUNO\n\n But not against the law of nature.\n\n My theory is that everybody is a\n\n potential murderer. Didn't you ever\n\n want to kill somebody? Say one of\n\n those useless fellows Miriam was\n\n running around with?\n\n GUY\n\n You can't go around killing people\n\n just because you think they're\n\n useless.\n\n BRUNO\n\n Oh, what's a life or two? Some people\n\n are bitter off dead, Guy. Take your --\n\n wife and my father, for instance.\n\n It reminds me of a wonderful idea\n\n had once. I used to put myself to\n\n sleep at night -- figuring it out.\n\n Now, let's say you want to get rid\n\n of your wife.\n\n GUY\n\n Why?\n\n BRUNO\n\n Let's say she refuses to give you a\n\n divorce --\n\n (raises a finger and\n\n stops Guy's protest)\n\n Let's say. You'd be afraid to kill\n\n her because you'd get caught. And\n\n what would trip you up? Motive.\n\n Now here's the plan...\n\n GUY\n\n I'm afraid I haven't time to listen.\n\n BRUNO\n\n (ignoring the remark)\n\n It's so simple, too. A couple of\n\n fellows meet accidentally, like you\n\n and me. No connection between them\n\n at all. Never saw each other before.\n\n Each of them has somebody he'd like\n\n to get rid of, but he can't murder\n\n the person he wants to get rid of.\n\n He'll get caught. So they swap\n\n murders.\n\n GUY\n\n Swap murders?\n\n Converted to PDF by www.screentalk.org 12.\n\n BRUNO\n\n Each fellow does the other fellow's\n\n murder. Then there is nothing to\n\n connect them. The one who had the\n\n motive isn't there. Each fellow\n\n murders a total stranger. Like you\n\n do my murder and I do yours.\n\n GUY\n\n (with relief)\n\n We're coming into my station.\n\n BRUNO\n\n For example, your wife, my father.\n\n Criss-cross.\n\n GUY\n\n (sharply)\n\n What?\n\n BRUNO\n\n (with a smile)\n\n We do talk the same language -- don't\n\n we, Guy?\n\n GUY\n\n (preparing to leave)\n\n Sure, we talk the same language.\n\n Thanks for the lunch.\n\n BRUNO\n\n (beaming)\n\n I'm glad you enjoyed it. I thought\n\n the lamb chops were a little overdone\n\n myself.\n\nHe holds out his hand. Guy is in a hurry but he shakes hands.\n\n GUY\n\n Nice meeting you, Bruno.\n\n BRUNO\n\n (detaining him at the\n\n door)\n\n You think my theory is okay, Guy?\n\n You like it?\n\n GUY\n\n Sure, sure, Bruno. They're all okay.\n\n (he salutes a quick\n\n goodbye and hurries\n\n away)\n\n Converted to PDF by www.screentalk.org 13.\n\nLeft alone, Bruno picks up Guy's lighter from the table,\n\nstarts to call Guy back to hand It to him.Then he looks closer\n\nat the insignia of crossed tennis rackets.\n\n BRUNO\n\n (smiling)\n\n Criss-cross.\n\n DISSOLVE TO:\n\nA WIDE VIEW OF THE TOWN OF METCALF\n\nMETCALF RAILROAD STATION\n\nas the train comes in.\n\nTHE TRAIN STATION PLATFORM MED. SHOT\n\nAs Guy gets off the with his suitcase and tennis rackets. A\n\nbaggage man with baggage truck is passing.\n\n GUY\n\n Hi, Bill.\n\n BAGGAGE MAN\n\n (smiling)\n\n Guy Haines! Good to too you, boy.\n\n You be sure to win at Southampton\n\n tomorrow, hear me? I've got two\n\n dollars on your nose.\n\n GUY\n\n (indicating his\n\n suitcase and rackets)\n\n Then park these in a lucky spot for\n\n a few hours, will you?\n\n BAGGAGE MAN\n\n Sure thing.\n\nHe loads them onto a truck.\n\n DISSOLVE TO:\n\nINT. METCALF STREET LONG SHOT\n\nGuy is walking up the main street.\n\n Converted to PDF by www.screentalk.org 14.\n\nEXT. MUSIC SHOP\n\nTypical music shop of a small town, with plate glass windows\n\nand displays of radios, records, sheet music, etc. Activity\n\nof a couple of customers and salespeople inside. Guy comes\n\nalong the street and goes into the shop.\n\nINT. MUSIC SHOP\n\nAs Guy enters. There are the usual counters and shelves,\n\npianos and radios on display, and the sound of a piano being\n\ntuned in the back of the store. MIRIAM is finishing with a\n\ncustomer at a counter. MR. HARGREAVES, the manager, is busy\n\nat the shelves. Another girl clerk is serving a customer.\n\nIn one of the glass cubicles where records are tried out, a\n\ncustomer is playing symphonic music; in a second glass cubicle\n\nanother customer is listening to a record of popular music.\n\nA third cubicle is empty. Activity of the street is seen\n\nthrough the plate glass front.\n\nGuy walks straight to Miriam, just as she is finishing with\n\nher woman customer, handing over a small package.\n\n MIRIAM\n\n (taking money from\n\n customer)\n\n Even change. Thank you, Madam.\n\n (she looks up at Guy\n\n as the woman moves\n\n off)\n\n Well -- hello, Guy.\n\n GUY\n\n You're looking well, Miriam.\n\nMiriam's face is pretty because it is still young. She is\n\nself-centered and inclined to be vindictive. She wears\n\nharlequin glasses with myopic lenses which tend to make her\n\neyes look small.\n\n MIRIAM\n\n So are you. You've got a nice tan,\n\n playing tennis with all your rich\n\n friends.\n\n GUY\n\n (ignoring the remark)\n\n What time do we meet your lawyer?\n\n MIRIAM\n\n (sly little smile)\n\n What's your hurry?\n\n Converted to PDF by www.screentalk.org 15.\n\n GUY\n\n My hurry? That's funny, coming from\n\n you! You're the one who's in a hurry,\n\n aren't you?\n\n MIRIAM\n\n (coyly)\n\n When you wouldn't give me the divorce\n\n right away, I sort of hoped it was\n\n because you were a little bit jealous.\n\n GUY\n\n (biting)\n\n I got over being jealous, a long\n\n time ago Miriam.\n\nMiriam's eyes slide toward the other girl clerk who has moved\n\ncloser, within listening range.\n\n MIRIAM\n\n (indicating empty\n\n glass cubicle)\n\n Let's talk in there.\n\nGuy follows Miriam across to the empty room. Miriam has\n\nbrought her purse along.\n\nThey enter.\n\nINT. CUBICLE\n\nOnce inside, the sounds of the music playing from other parts\n\nof the shop are heard but very faintly. The piano tuning\n\nstill goes on, but less stridently. Miriam and Guy are cooped\n\ntogether in the close quarters.\n\n MIRIAM\n\n (intimately)\n\n Now this is cosier. Sort of like\n\n old times, isn't it, Guy?\n\n GUY\n\n (coldly)\n\n Oh, skip it, Miriam. It's pretty\n\n late to start flirting with a\n\n discarded husband. Especially when\n\n you're going to have another man's\n\n baby.\n\n MIRIAM\n\n Do you know, I think you're handsomer\n\n than ever?\n\nConverted to PDF by www.screentalk.org 16.\n\n GUY\n\n Let's see your lawyer and get this\n\n over with.\n\n MIRIAM\n\n Did you bring the money, Guy? Lawyers\n\n are expensive.\n\n GUY\n\n (taking money from\n\n his wallet)\n\n Here it is.\n\n MIRIAM\n\n (taking the money\n\n greedily)\n\n If I'd known what all that tennis\n\n nonsense of yours was going to lead\n\n to, I wouldn't have run out on you.\n\n GUY\n\n What are you trying to say, Miriam?\n\n Come out with it.\n\n MIRIAM\n\n (tucking the bills\n\n away)\n\n I'm not getting a divorce.\n\n GUY\n\n (tense and angry)\n\n Why, you little doublecrosser. I\n\n didn't want this divorce, you did.\n\n That's what you've been harping about\n\n for the past year.\n\n MIRIAM\n\n It's a woman's privilege to change\n\n her mind... Now I can shop for some\n\n pretty clothes. I wouldn't want you\n\n to be ashamed of me in Washington\n\n when we go to all those dinners and\n\n swanky parties.\n\n GUY\n\n And what do you mean by that?\n\n MIRIAM\n\n (Coyly)\n\n Don't look so mad, Guy. You always\n\n smile when your picture is being\n\n taken for the papers.\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 17.\n\n MIRIAM (CONT'D)\n\n Especially when you have Anne Burton\n\n hanging on your arm.\n\n GUY\n\n Let's not talk about Anne Burton.\n\n MIRIAM\n\n So, it's really serious between you\n\n two? Well, you can throw your dreams\n\n about her into the ashcan. Guy, I'm\n\n coming to Washington.\n\n GUY\n\n What for?\n\n MIRIAM\n\n To have my baby and be with you.\n\n GUY\n\n Why me? It's not my baby.\n\n MIRIAM\n\n But people don't know that, Guy, do\n\n they? It would make a pretty story,\n\n wouldn't it -- the senator's daughter\n\n involved with a married man who's\n\n about to become a father.\n\n GUY\n\n (furiously)\n\n You black conniving little liar!\n\nA few people in the shop look around as Guy's voice rises\n\nabove the sound of the record playing.\n\n MIRIAM\n\n Keep your voice down.\n\n GUY\n\n What happened? Did he run out on\n\n you?\n\n MIRIAM\n\n No man runs out on me. Not even\n\n you.\n\n GUY\n\n You're a liar and a cheat, Miriam.\n\n You've wanted to get rid of me long\n\n enough and now I'll go you one better --\n\n I never want to see or hear of you\n\n again.\n\n Converted to PDF by www.screentalk.org 18.\n\n MIRIAM\n\n (demurely)\n\n I could be very pathetic as the\n\n deserted little mother in a courtroom,\n\n Guy. Think it over. Who would\n\n believe you?\n\nGuy seizes her angrily and in so doing, knocks the tone arm\n\nacross the record with a loud screech. From outside we can\n\nsee heads turn. Mr. Hargreaves, the manager, is very\n\ndisturbed.\n\nMED. SHOT THROUGH GLASS PARTITION FROM HARGREAVES' VIEWPOINT\n\nWe see Guy gripping Miriam's arms and apparently addressing\n\nher in a threatening manner, although we do not hear his\n\nwords. The smile has faded from Miriam's face and something\n\nlike cringing fear has taken its place. She is drawn and\n\ntense and seems to cower beneath Guy's rage.\n\nMr. Hargreaves moves forward and opens Guy's tirade.\n\n GUY\n\n ...That's what should happen to people\n\n like you. And if I...\n\n HARGREAVES\n\n (interrupts)\n\n Break it up, folks. This isn't the\n\n place for a family quarrel.\n\n GUY\n\n (his eyes blazing)\n\n Sorry. I'm leaving.\n\nHe starts to exit from the booth. Miriam grabs his arm and\n\nscreams at him:\n\n MIRIAM\n\n (yelling like a\n\n fishwife)\n\n You heard what I said, Guy Haines.\n\n You can't throw me away like an old\n\n shoe. I'm coming to Washington to\n\n have my baby. Tell that to the\n\n senate!\n\nGuy strides out of the store, the manager and a few customers\n\nturning around in surprise.\n\n Converted to PDF by www.screentalk.org 19.\n\nThe two customers in other booths, seeing the quarrel, open\n\ntheir doors simultaneously and Miriam's tirade is climaxed\n\nby a cacophony of noise, a big symphony, loud hot music, and\n\nthe apparently unaware piano tuner.\n\nEXT. MAIN STREET METCALF SHOOTING TOWARDS STATION\n\nGuy is striding along angrily. He comes to the same\n\nintersection and the same cop. The officer makes a friendly\n\ngesture, is if he'd like to talk awhile, but Guy strides\n\npast him without noticing.\n\nEXT. METCALF STATION (PROCESS)\n\nGuy comes into the scene, crosses to a row of public telephone\n\nbooths, enters one. Inside the telephone booth, he dumps\n\nsome loose change on the shelf, sticks a nickel in the\n\ntelephone, speaks into it.\n\n GUY\n\n Long distance.\n\n (a pause)\n\n I want Washington, D. C. The number\n\n is Republic 0800. Person to person.\n\n Miss Anne Burton.\n\nAnother pause, very long. Guy is very restless. He digs a\n\ncigarette out of his pocket and sticks it in his mouth, then\n\nlooks through his pockets for his lighter, doesn't find it.\n\nHe looks puzzled, but about that time the operator speaks to\n\nhim.\n\n GUY\n\n (continuing)\n\n Right.\n\nGuy picks coins up off the shelf and drops them into the\n\ntelephone, then waits. He shifts the receiver and fumbles\n\nin his other jacket pocket, then turns to the phone.\n\n GUY\n\n (tautly, into phone)\n\n Anne, -- Anne darling. Yes, I'm in\n\n Metcalf --\n\n (gets a grip on himself)\n\n No, everything didn't go smoothly.\n\n She doesn't want a divorce, not\n\n now....\n\n Converted to PDF by www.screentalk.org 20.\n\nINT. BURTON LIVING ROOM\n\nANNE BURTON is a beautiful, high-spirited and well-bred young\n\nwoman. The smile on her face his faded to anxiety as she\n\nlistens over the telephone which is on the desk.\n\n ANNE\n\n (after a pause then\n\n with unpleasant\n\n realization)\n\n Another man's child! But she can't\n\n do that to you, Guy -- it's\n\n unbelievable -- it's, it's evil!\n\n (she listens, then\n\n calmly)\n\n Yes, I know how you must feel.\n\n (pause)\n\n But you sound so savage.\n\nBACK TO GUY IN TELEPHONE BOOTH\n\n GUY\n\n (furiously)\n\n Sure I sound savage. I feel savage.\n\n I'd like to break her neck!\n\n (a pause, then raising\n\n his voice)\n\n I said I'd like to break her foul,\n\n poisonous, useless little neck!\n\n (the connection is\n\n bad and he strains\n\n to hear)\n\n What's that?\n\nMeantime the noise of a through train has been HEARD, and\n\nthe horn on a streamliner locomotive. It has come up very\n\nfast, it is now almost to the station. Guy rises his voice\n\nand yells into the telephone. His voice fights the roar of\n\nthe train:\n\n GUY\n\n I SAID I COULD STRANGLE HER!\n\nThe expression on his face is frenzied and suggesting that\n\nhe means exactly what he is saying.\n\n DISSOLVE TO:\n\n Converted to PDF by www.screentalk.org 21.\n\nINT. ANTHONY LIVING ROOM\n\nThe scene opens on a CLOSEUP OF A MAN'S HANDS. One of them\n\nis semi-flexed and turning slowly, The other is receiving\n\nthe final touches of a manicure.\n\nCAMERA PULLS BACK to reveal that these are Bruno's hands,\n\nand that, he is studying them moodily, CAMERA PULLS BACK\n\nFARTHER to reveal his mother, MRS. ANTHONY, sitting opposite\n\nhim at a little table in the Anthony living room. She is\n\nworking with scissors, file and nail buffer. Mrs. Anthony\n\nis a gentle, once pretty woman, whose pastel exterior harbors\n\na tigress-like determination to protect her son, Bruno is in\n\nhis robe and is unshaven.\n\nThere is evidence of long established wealth in the heavy\n\ndark appointments of this room.\n\n MRS. ANTHONY\n\n Since you insisted on a manicure,\n\n dear, I do wish you'd keep your hands\n\n quiet. You're so restless lately.\n\n BRUNO\n\n (almost dreamily as\n\n he admires the free\n\n hand)\n\n I like them to look just right.\n\nMrs. Anthony looks up, notices his moody expression.\n\n MRS. ANTHONY\n\n Did I file them too short?\n\n BRUNO\n\n No, Ma. They look fine. Thanks.\n\n MRS. ANTHONY\n\n Then what's the matter?\n\n BRUNO\n\n I'm all right, Ma. Don't worry about\n\n me.\n\n MRS. ANTHONY\n\n You look so Pale, dear. Are you out\n\n of vitamins?\n\n BRUNO\n\n I bought a bottle of them yesterday.\n\n A whole fifth.\n\n Converted to PDF by www.screentalk.org 22.\n\n MRS. ANTHONY\n\n (anxiously)\n\n But you have that 'look'. I can\n\n always tell. You haven't got into\n\n any more mischief, Bruno?\n\nHe denies this with a slow, solemn shake of his head.\n\n MRS. ANTHONY\n\n I do hope you've forgotten about\n\n that silly little plan of yours?\n\n BRUNO\n\n (sharply)\n\n Which one?\n\n MRS. ANTHONY\n\n (smiling)\n\n About blowing up the White House?\n\n BRUNO\n\n (his eyes dancing)\n\n I was only kidding, Ma. Besides,\n\n what would the president say?\n\n MRS. ANTHONY\n\n (laughing gaily)\n\n You're a naughty boy, Bruno. But\n\n you can always make me laugh.\n\n (she rises)\n\n Now get shaved, dear, before your\n\n father gets home.\n\nBruno's fist crashes down on the little table, upsetting it,\n\nas he gets to his feet.\n\n BRUNO\n\n I'm sick and tired of bowing and\n\n scraping to the king.\n\n MRS. ANTHONY\n\n (placating him)\n\n Now, now, Let's not lose control.\n\n Come see my painting, dear --\n\n (she leads him toward\n\n an easel)\n\n I do wish you'd take up painting.\n\n It's such a soothing pastime.\n\nThey look at the painting.\n\n Converted to PDF by www.screentalk.org 23.\n\nINSERT\n\nThe painting is a horrible mess. Out of the violence of the\n\npattern a man's face can be discerned, wild-eyed and\n\ndistorted. We hear laughter from Bruno.\n\nBACK TO SCENE\n\nBruno's roar of laughter puzzles Mrs. Anthony, but she is\n\npleased to hear his good humor. He puts an arm around her.\n\n BRUNO\n\n You're wonderful, Ma! It's the old\n\n boy, all right. That's father!\n\n MRS. ANTHONY\n\n (bewildered)\n\n It is? I was trying to paint Saint\n\n Francis.\n\nAt this moment there is the sound of the front door opening.\n\nThen immediately the telephone bell rings in the hall. Bruno\n\nis instantly alert, as if he had been expecting a call. He\n\ngoes toward the door to the hall, as the butler enters.\n\n BUTLER\n\n (to Bruno)\n\n They are ready with your call to\n\n Southampton, Sir.\n\nBruno's father MR. ANTHONY, purposefully enters the living\n\nroom. He an impeccably dressed business man with an\n\nuncompromising eye. His entrance momentarily blocks Bruno's\n\nexit.\n\n MRS. ANTHONY\n\n (to her husband)\n\n How nice that you're early, Charles.\n\n I'll tell cook....\n\nBruno now exits into the hall, passing his father without\n\nspeaking.\n\n MR. ANTHONY\n\n Just a minute, Eunice.\n\n (calls after Bruno)\n\n Bruno! Come here! I want to talk\n\n to you and your mother.\n\n Converted to PDF by www.screentalk.org 24.\n\nINT. HALL CLOSE SHOT BRUNO\n\nas he approaches the telephone.\n\n BRUNO\n\n (calls back to his\n\n father)\n\n Sorry father. Long distance.\n\n (he picks up the\n\n telephone)\n\n Hello...\n\nCAMERA MOVES IN TO A BIG HEAD CLOSEUP OF BRUNO at the\n\ntelephone as the Voices of his mother and father can be heard\n\nfrom the other room.\n\n MR. ANTHONY'S VOICE\n\n Now it's hit and run driving! And\n\n you knew about it all the time!\n\n BRUNO\n\n (eagerly into phone)\n\n Guy?\n\n (pause)\n\n Bruno, Bruno Anthony.\n\n MR. ANTHONY'S VOICE\n\n You're going to protect him once too\n\n often. After all we do have a\n\n responsibility to society.\n\nBruno gives a look in his father's direction, before he speaks\n\ninto the telephone in a low voice.\n\n BRUNO\n\n I just wanted to ask how you made\n\n out with Miriam.\n\nINT. LOCKER ROOM OF TENNIS CLUB CLOSE SHOT GUY AT TELEPHONE\n\n GUY\n\n (puzzled)\n\n What?\n\n (listens)\n\n Metcalf? Who'd you say you were?\n\n Converted to PDF by www.screentalk.org 25.\n\nCLOSEUP BRUNO\n\n BRUNO\n\n (sotto voce)\n\n Bruno, Guy. Bruno Anthony. Don't\n\n you remember? On the train.\n\nThe voices of Mr. and Mrs. Anthony can still be heard in\n\ndispute as Bruno listens at phone:\n\n MRS. ANTHONY\n\n I never permit it!\n\nBruno gives a significant look in direction of the living\n\nroom as he speaks into the phone.\n\n BRUNO\n\n (softly)\n\n Are you getting your divorce?\n\n MR. ANTHONY'S VOICE\n\n I tell you he should be sent somewhere\n\n for treatment before it's too late.\n\n BRUNO\n\n (into phone, with\n\n satisfaction)\n\n So she double-crossed you! Are you\n\n going to see her again?\n\nThe phone clicks in Bruno's ear. He looks hurt for an\n\ninstant, then replaces the receiver. Bruno listens to his\n\nfather off scene and his expression becomes more enigmatic.\n\n MR. ANTHONY'S VOICE\n\n I tell you, Eunice, I'm going to\n\n have that boy put away if it's the\n\n last thing I do!\n\nBruno looks off in direction of his farther's voice with an\n\nexpression which says, "Crow while you can, you haven't\n\nmuch time." He reaches into his pocket, brings out Guy's\n\ncigarette lighter and as he flicks it on and off.\n\n DISSOLVE TO:\n\nEXT. METCALF STATION LONG SHOT DAY\n\nThis is the same shot we saw when Guy arrived in Metcalf.\n\nWe see the station and one of the main streets beyond the\n\nstation.\n\n Converted to PDF by www.screentalk.org 26.\n\nLONG SHOT A NEARER VIEW\n\nWe see the train come around the curve. Again this is just\n\nthe same angle that we used for Guy. It comes to a stop in\n\nthe foreground and we see Bruno alight onto the platform.\n\nHe looks about him for a moment and then strolls away in the\n\ndirection of the town. He approaches the row of telephone\n\nbooths.\n\nEXT. STATION CLOSE SHOT\n\nWe see Bruno enter the small booth and start to glance through\n\nthe telephone directory.\n\nINSERT TELEPHONE DIRECTORY\n\nBruno's finger runs down the names until it stops at:\n\n Joyce, Miriam Haines. 2420 Metcalf Avenue.\n\nA RESIDENTIAL STREET IN METCALF LONG SHOT\n\nIt is now much later. It is beginning to get dark, and the\n\nstreet lights are on. In the far distance we see a local\n\nbus approaching.\n\nMED. SHOT\n\nSHOOTING DOWN onto a small seat by a bus stop, we see Bruno\n\nwith an open newspaper in front of him. It is held up as he\n\nreads it.\n\nCLOSEUP\n\nBruno is glancing over the top of the paper.\n\nLONG SHOT\n\nFrom his viewpoint we see a typical frame house. The upper\n\nwindows are lit as are the lower ones as well. A woman is\n\nsitting in a rocker on the front porch. This is MRS. JOYCE,\n\nMiriam's mother. She has white hair. A woman comes along\n\nthe street and pauses as she gets to Mrs. Joyce.\n\n Converted to PDF by www.screentalk.org 27.\n\n WOMAN\n\n (calls out as she\n\n passes)\n\n Hello Mrs. Joyce. Warm, ain't it?\n\n MRS. JOYCE\n\n That it is.\n\n WOMAN\n\n I've been reading where your son-in-\n\n law's been coming right along at\n\n tennis.\n\n MRS. JOYCE\n\n (sourly)\n\n We don't have any interest in tennis\n\n any more.\n\nThe neighbor passes on.\n\nCLOSE UP\n\nBruno, still glancing over the top of his paper.\n\nLONG SHOT\n\nAgain from Bruno's viewpoint, we see Miriam's house. At\n\nthis moment the front door swings open, emitting a long streak\n\not bright light. We see the silhouette of a woman emerge,\n\nfollowed by two other men. They're laughing and joking.\n\nSuddenly they look up the street. At this very moment the\n\nbus pulls up in front of Bruno's view, cutting off the sight\n\nof his quarry. The bus comes to a stop.\n\nCLOSE SHOT\n\nBruno rises in alarm and moves around toward the end of the\n\nbus so that he shall not lose sight of the girl coming out\n\nof the house.\n\nSEMI-LONG SHOT\n\nFrom his viewpoint, the girl, whom we now see is Miriam, is\n\nrunning followed by the two young men. They are calling for\n\nthe bus not to go - shouting, "Hi - stop!" Mrs. Joyce calls\n\nfrom the porch:\n\n MRS. JOYCE\n\n Don't you stay out too late, Miriam.\n\n Converted to PDF by www.screentalk.org 28.\n\n MIRIAM\n\n (calling back)\n\n Goodnight, Mother. See you later.\n\nCLOSE UP\n\nBruno watches Miriam.\n\nMED. SHOT\n\nMiriam comes nearer and nearer to Bruno. With her two\n\ncompanions she brushes past him and jumps onto the bus. THE\n\nCAMERA PANS BRUNO AFTER THEM.\n\nEXT. AMUSEMENT PARK LONG SHOT\n\nWe see the bus pull up outside the Amusement Park, and the\n\nvarious passengers alight. These include Miriam nd her\n\ncompanions, and Bruno.\n\nLONG SHOT NEARER VIEW OF THE AMUSEMENT PARK\n\nWe see the usual midway with its various concessions on each\n\nside: in the distance the Ferris wheel, Merry-go-rounds,\n\netc., and beyond that a lake. In the foreground we see people\n\nfilling in and out.\n\n DISSOLVE TO:\n\nMED. LONG SHOT A GROUP BY A FROZEN CUSTARD STAND\n\nThis group comprises Miriam and her two boy-friends. They\n\nlick their way out of the crowd and debate between themselves\n\nwhere to go next.\n\nCLOSE SHOT\n\nMiriam's eye catches the attention of something off screen.\n\nSEMI-LONG SHOT\n\nFrom her viewpoint we see Bruno standing and casually watching\n\nher. Other people pass around and in front of him, so that\n\nhe is the only immobile figure.\n\n Converted to PDF by www.screentalk.org 29.\n\nSEMI-CLOSEUP\n\nMiriam, with a kind of coy consciousness, turns away with\n\nthe others and they go on to some other concession.\n\nMED. SHOT\n\nAs Bruno starts to advance in the direction of Miriam he is\n\nmomentarily held up by a small boy in cowboy uniform carrying\n\na gun and a balloon. The small boy points the gun at Bruno.\n\nSEMI-CLOSE UP\n\nThe small boy pointing the gun fires it twice with a couple\n\nof 'bangs!' He then starts to move off.\n\nSEMI-CLOSE UP\n\nBruno moves on past the boy. He casually touches the balloon\n\nwith his cigarette end -- it goes off with a 'pop'.\n\nCLOSE UP\n\nThe small boy turns and looks with dismay at his pricked\n\nballoon, wondering what happened.\n\nSEMI-CLOSE UP\n\nBruno moves on, pleased with himself, returning his attention\n\nto Miriam who is somewhere ahead of him.\n\nMEDIUM SHOT\n\nMiriam and her two boy-friends by the sledge hammer concession\n\nwhere the aim is to swing the hammer hard enough down onto\n\nits target to ring the bell and register the 100 mark. Miriam\n\nis in the foreground of the shot. The first boy steps up to\n\ntry his hand. As he swings, Miriam turns and glances about\n\nher, obviously looking for Bruno.\n\nLONG SHOT FROM MIRIAM VIEWPOINT\n\nThe crowds milling, but no sign of Bruno.\n\n Converted to PDF by www.screentalk.org 30.\n\nMEDIUM SHOT\n\nThe first boy having failed to ring the bell, the second\n\nstops up and slams the hammer down.\n\nCLOSE SHOT\n\nThe register shooting up only to the hallway mark.\n\nCLOSE SHOT MIRIAM\n\nShe looks a little disdainful and again glances around for\n\nBruno. Looking first to her left where she sees nothing,\n\nshe then looks to her right, and as she does THE CAMERA PANS\n\nto show Bruno standing right it her shoulder. Miriam gives\n\na little start. Bruno smiles at her. With a smirk he walks\n\nover and after paying his fee, goes to take up the hammer.\n\nCLOSE UP MIRIAM\n\nShe watches Bruno.\n\nCLOSE SHOT\n\nBruno looks down at his hands.\n\nINSERT\n\nBruno's two strong hands - as he holds them palms tilted\n\nupward and fingers curled in.\n\nCLOSE UP\n\nBruno, as he smiles faintly, glancing across at Miriam.\n\nCLOSE UP MIRIAM\n\nShe gives a faint smile in return.\n\nCLOSE SHOT\n\nWith a studied movement, Bruno picks up the handle of the\n\nhammer and swings.\n\n Converted to PDF by www.screentalk.org 31.\n\nCLOSE SHOT\n\nThe register shoots up to the 100 mark and rings the bell.\n\nMEDIUM SHOT\n\nBruno drops the hammer and glances around at Miriam again.\n\nHer two boy-friends are calling for her from a little\n\ndistance.\n\n BOY'S VOICE\n\n Come On, Miriam. Come On!\n\nCLOSE SHOT MIRIAM\n\nShe turns away and is lost in the crowd.\n\nMEDIUM SHOT OVER BRUNO'S SHOULDER AT MERRY-GO ROUND IN\n\nBACKGROUND\n\nBruno turns to follow Miriam, his manner casual. As he takes\n\na few steps, WE PAN ACROSS with him until, over his shoulder,\n\nwe see a merry-go-round in the background. Miriam and the\n\ntwo boys are aboard and climbing onto horses. As Bruno goes\n\ntoward the merry-go-round, the CAMERA MOVES UP A LITTLE with\n\nhim. The merry-go-round starts to move slowly round as Bruno\n\nhops on.\n\nMEDIUM SHOT ON MERRY-GO-ROUND\n\nBruno begins to look around for Miriam, who is apparently on\n\nthe other side of the merry-go-round. He starts to thread\n\nhis way through the horses which are beginning to move up\n\nand down. CAMERA FOLLOWING HIM. He passes one or two of\n\nthe oncoming heads before he reaches Miriam. She is on an\n\noutside mount which is high in the air when she sees Bruno\n\nfacing her. Her laughter dies for a moment and she smiles\n\nat him coyly. Bruno passes her and gets on the horse directly\n\nbehind her, Miriam glancing at him as her horse comes down.\n\nMEDIUM SHOT BRUNO ON HORSE\n\nWith horse's head in foreground, as it is coming toward us.\n\n Converted to PDF by www.screentalk.org 32.\n\nSIDE VIEW MIRIAM\n\nMiriam on her horse, moving from left to right. Miriam,\n\nholding the reins, glances back with a gay laugh.\n\nSIDE VIEW BRUNO\n\nBruno on his horse, as though he is chasing Miriam. He is a\n\nlittle more open now in his laughter.\n\nGROUP SHOT MIRIAM AND TWO BOYS\n\nMiriam and her boy friends begin to sing the song being played\n\non the calliope.\n\nCLOSE UP MIRIAM\n\nAs she starts to sing, she glances back.\n\nCLOSE UP BRUNO\n\nHe is starting to join in the singing.\n\nMEDIUM SHOT\n\nThe horses of the merry-go-round are filling the screen as\n\nthey whizz by, and again we get the picture of Bruno chasing\n\nMiriam as they rush past the CAMERA, the music and tempo at\n\na high speed.\n\n LAP DISSOLVE TO:\n\nEXTERIOR OF BOAT LANDING ON SHORE OF ARTIFICIAL LAKE\n\nAcross the water may be seen a small wooded island. Between\n\nthis and the boat landing there is an artificially constructed\n\n"Tunnel of Love".\n\nWe see Miriam and her companions approach the boat concession\n\nand CAMERA FOLLOWS THEM onto the little landing stage. CAMERA\n\nMOVES UP SLOWLY over the boy's shoulders until we get MIRIAM\n\nIN CLOSE UP. She glances back. Her expression changes to a\n\ncoy smile of satisfaction as she sees:\n\n Converted to PDF by www.screentalk.org 33.\n\nMEDIUM SHOT (FROM MIRIAM'S VIEWPOINT)\n\nBruno is approaching the pay box.\n\nMEDIUM SHOT\n\nMiriam and her companions are escorted to a small boat with\n\nelectric motor. Once they are seated the boat chugs away\n\nfrom the landing stage and off into the darkness.\n\nBruno steps into the foreground and gets into the next boat\n\nwhich floats alongside. He, too, moves away into the\n\ndarkness.\n\nENTRANCE TO THE TUNNEL\n\nAs Miriam's boat passes through, she gives another little\n\nglance over shoulder before her boat disappears into the\n\ndarkness of the tunnel.\n\nAfter a brief moment Bruno's boat comes into the picture,\n\nand it, too, goes into the tunnel.\n\nINSIDE THE TUNNEL\n\nWe see the silhouettes of the occupants of Miriam's boat on\n\nthe wall of the tunnel, lit dimly from the light coming from\n\nthe tunnel exit.\n\nThe silhouette of Bruno in his boat, lit by the tunnel\n\nentrance, gradually approaches the other three. When the\n\nsilhouettes are almost touching, we --\n\n CUT TO:\n\nEXIT OF THE TUNNEL\n\nIt is empty. There is a sudden piercing scream from inside,\n\nfollowed after a second or two by protestations and giggling\n\nas Miriam's boat emerges into the light. She is pushing one\n\nof the boys away from her.\n\n MIRIAM\n\n (squealing)\n\n George, stop it, I tell you!\n\nTheir boat moves out of the picture, toward the island.\n\nPresently Bruno's boat comes smilingly following and he,\n\ntoo, moves on out of the picture.\n\n Converted to PDF by www.screentalk.org 34.\n\nMEDIUM SHOT ISLAND\n\nThe group of Miriam and her companions are scrambling out of\n\ntheir boat and moving onto the island, one of the boys trying\n\nthe boat on the shore. They disappear into the Woods of the\n\nisland.\n\nAgain Bruno's boat comes into the picture. He steps out,\n\nlift the prow of the boat a little onto the shore.\n\nLONG SHOT ISLAND\n\nWe see the amusement park lighted beyond the lake.\n\nSilhouetted in the foreground, the trees and foliage of the\n\nisland. Nearby we see the silhouetted figures of Miriam and\n\nher companions move across the scene, right to left. Miriam\n\nis pushing George away from her.\n\n MIRIAM\n\n (protesting\n\n perfunctorily)\n\n George, no!\n\nShe backs away from him and the boys go on picture. Miriam\n\ngoes in another direction, around, the bushes. George\n\nobviously misses her, for we hear his voice call out:\n\n GEORGE'S VOICE\n\n Miriam!\n\nMiriam backs out of the bushes until the back of her head is\n\nin CLOSEUP in the foreground of the shot. Suddenly she hears\n\nsteps in back of her and turns her head toward CAMERA. Her\n\nface changes as she recognizes someone offscene.\n\n MIRIAM\n\n Oh!\n\nShe gives a coy smile of recognition. CAMERA PULLS BACK to\n\nreveal the mad and shoulders of Bruno between Miriam and the\n\ncamera. His hand holds Guy's lighter which he flicks on as\n\nhe raises it above Miriam's face. 0f Bruno, we see only the\n\nback of his head and shoulders.\n\n BRUNO\n\n Is your name Miriam?\n\n MIRIAM\n\n (with surprise)\n\n Why yes. How did you --\n\n Converted to PDF by www.screentalk.org 35.\n\nWe see Bruno's gloved hands dart quickly to Miriam's throat.\n\nThe lighter falls down out of picture, and as Bruno's hands\n\ngrip her throat, his head moves slightly to blot out Miriam's\n\nface. His head moves a bit farther until Miriam's face is\n\nnearly uncovered at the other side of the screen, and we see\n\nher glasses fall off.\n\nCLOSE SHOT\n\nMiriam's glasses hit the ground. The shadows of their\n\nstruggling figures over the shot.\n\nCLOSE UP\n\nThe screen is filled with one of the lenses of the glasses.\n\nThey are of the diminishing type. Against the moonlit sky\n\nwe see reflected, the elongated struggling figures, as though\n\nwe were shooting up at them. Suddenly one of the figures\n\nfalls forward.\n\nCLOSE UP\n\nMiriam's head drops into the picture by the glasses.\n\nBruno's hand comes into the picture and picks up the glasses.\n\nOne of the lenses has been broken by Miriam's fall.\n\nAs we see Bruno's sport shoes move away, the CAMERA MOVES\n\nPAST MIRIAM'S HEAD until it comes to Guy's lighter pressed\n\ninto the earth.\n\nCLOSE UP BRUNO\n\nBruno glances back over his shoulder. He looks down and\n\ngoes back one or two steps.\n\nCLOSE UP BRUNO'S HAND\n\nBruno's hands retrieve the lighter from the ground.\n\nLONG SHOT ISLAND\n\nWe see a full view of the island again, with the amusement\n\npark beyond. The faint noise of the calliope continues in\n\nthe distance. Bruno has been lost to view.\n\n Converted to PDF by www.screentalk.org 36.\n\nMiriam's companions are still searching for her. We hear\n\ntheir faint voices in the distance.\n\n VOICES\n\n Miriam! Miriam! Where are you?\n\nMEDIUM SHOT\n\nBruno comes to the shore where his boat is moored. He gets\n\nin and is quickly chugging away. He moves calmly, matter-of-\n\nfact and not furtively.\n\nLONG SHOT LAKE\n\nBruno's boat throbbing its way across toward the landing\n\nstage.\n\nMEDIUM SHOT LANDING STAGE\n\nThere are two boats unloading. Bruno's boat is approaching.\n\nWe hear a loud call from the island. Someone has found\n\nMiriam.\n\n VOICES\n\n Hey, here she is! What's the matter\n\n with her? Has she fainted?\n\nMore shouts from the island cause the people at the landing\n\nstage to look back. The boatman's attention is also\n\nattracted. Suddenly, as Bruno is getting out of boat, there\n\nis a loud scream from the island.\n\n VOICE\n\n (crying out)\n\n She is dead!\n\n OTHER VOICE\n\n (from island)\n\n Help! Help!\n\nBruno by this time has stopped onto the landing stage, and\n\nin company with the other people, is looking back as if to\n\nsee what's wrong on the island. Then he moves away, starting\n\noff of the landing stage. The boatman turns and glances at\n\nBruno, but quickly returns his attention to the disturbance\n\nacross on the island. He hurries forward and with a couple\n\nof men passengers jumps into one of the boats. He calls to\n\nhis assistant as he gets into the boat.\n\n Converted to PDF by www.screentalk.org 37.\n\n BOATMAN\n\n Got a cop!\n\nThe assistant runs off out of the pictures\n\nMEDIUM SHOT BRUNO\n\nAs Bruno calmly threads his way along the midway, we hear\n\nabove the noise of the various concessions, a shrill police\n\nwhistle in the distance. Presently a couple of policemen\n\ncomes running from direction of the main entrance and past\n\nBruno. He glances at them over his shoulder, then strolls\n\non toward the main entrance to the park.\n\nENTRANCE TO AMUSEMENT PARK EXTERIOR\n\nAs Bruno comes out through the turnstile, he stands for a\n\nmoment on the street. At this moment a man hesitates at the\n\ncurbstone. He is blind and tapping the sidewalk with his\n\nwhite cane. He takes one step into the roadway, then\n\nhesitates. Bruno steps forward and takes the blind man's\n\narm. CAMERA PULLS BACK as Bruno escorts the blind man across\n\nthe road. With a sweeping gesture he holds back a couple of\n\ncars to lot them pass.\n\nOnce on the other side of the road, the blind man utters his\n\nthanks.\n\n BLIND MAN\n\n Thanks.\n\nHe goes off.\n\nBruno looks back toward the park, then glances down at his\n\nwristwatch.\n\nINSERT BRUNO'S WRISTWATCH\n\nThe time is 9.30.\n\n LAP DISSOLVE TO:\n\nINT. OBSERVATION CAR OF A TRAIN NIGHT\n\nThrough the rear window we see the tracks rushing away from\n\nus. Seated in the foreground are Guy Haines and a rather\n\nprofessorial type opposite him, a bespectacled man around\n\nforty-five or fifty who is extremely drunk.\n\n Converted to PDF by www.screentalk.org 38.\n\nMEDIUM SHOT GUY\n\nHe is reading an evening newspaper.\n\nCLOSE SHOT\n\nThe feet opposite Guy stretch out and touch Guy's feet.\n\nCLOSEUP GUY\n\nHe lowers his paper and looks across.\n\nMED. SHOT\n\nThe drunk opposite Guy looks down at his feet and then up to\n\nGuy resentfully as though Guy had kicked him. He eyes Guy\n\nup and down, then suddenly, without warning, bursts into\n\nsong, to the tune of the Barber Shop Chord.\n\n COLLINS\n\n There was a man, now please take\n\n note. There was a man who had a\n\n goat. He loved that goat, Indeed he\n\n did. He loved that goat, just like\n\n a kid.\n\n (He stops singing\n\n abruptly and addresses\n\n Guy)\n\n What is your opinion?\n\n GUY\n\n (amused)\n\n You'll never make the Metropolitan.\n\n COLLINS\n\n (fuzzily -- pumping\n\n Guy's hand)\n\n Name's Collins. On sabbatical -\n\n Delaware Tech. Glad to meet you. I\n\n jus' gave a speech in New York. On\n\n integration. In the differential\n\n calculus a function is given and its\n\n differential is obtained. Understand?\n\n GUY\n\n (solemnly)\n\n Sure, I understand.\n\n Converted to PDF by www.screentalk.org 39.\n\n COLLINS\n\n (resentfully)\n\n Y'do?\n\nAgain he bursts into loud song.\n\n LAP DISSOLVE TO:\n\nLONG SHOT WASHINGTON EXTERIOR ABOUT 1 A.M. MOONLIGHT\n\nA solitary taxi is seen driving past the Capitol Building.\n\n LAP DISSOLVE TO:\n\nThe taxi comes to a side street and stops outside a small\n\napartment house.\n\nMED. SHOT\n\nGuy gets out of the taxi with his rackets and bag, pays the\n\ndriver and goes up the steps to the front door of his\n\napartment.\n\nCLOSE SHOT\n\nAs Guy is about to enter the front door and we see his name\n\nposted on a small card as one of the several tenants, he\n\nhears a soft call from across the street.\n\n VOICE\n\n (softly)\n\n Guy!\n\nGuy turns his head and looks across the street.\n\nMED. LONG SHOT (FROM GUY'S VIEWPOINT)\n\nWe see a small space between two houses across the street.\n\nOut of the darkness the voice repeats.\n\n VOICE\n\n Over here, Guy.\n\nMED. SHOT GUY\n\nHe turns, and with a slightly bewildered and wary expression,\n\ngoes out of the picture to cross the street.\n\n Converted to PDF by www.screentalk.org 40.\n\nMED. SHOT\n\nGuy reaches the other side of the street and still puzzled\n\nand cautious, approaches the dark alleyway.\n\nMED. SHOT\n\nAfter a moment a figure steps out of the darkness. It is\n\nBruno. He steps back into the darkness again as Guy comes\n\nup to him.\n\nTWO SHOT\n\nGuy frowning in puzzlement as he looks at Bruno.\n\n BRUNO\n\n (cheerfully)\n\n Hello, Guy.\n\n GUY\n\n (recognizes Bruno --\n\n not pleased)\n\n What are you doing here? At this\n\n time of night?\n\n BRUNO\n\n (a little sadly)\n\n You don't seem very pleased to see\n\n me, Guy.\n\nGuy stands without answering.\n\n BRUNO\n\n (pleased again)\n\n I brought you a little present.\n\n GUY\n\n What do you mean?\n\nBruno's hand comes out of his pocket and he hands Miriam's\n\nglasses to Guy.\n\nINSERT\n\nGuy's hands taking Miriam's glasses from Bruno. One of the\n\nlenses is broken.\n\n Converted to PDF by www.screentalk.org 41.\n\nTWO SHOT\n\nAs Guy takes the glasses he looks at Bruno in bewilderment.\n\n GUY\n\n What's this all about?\n\n BRUNO\n\n Recognize them?\n\nCLOSEUP GUY\n\nHe looks down at the glasses, mystified. He looks up again\n\nto Bruno.\n\nCLOSEUP BRUNO\n\n BRUNO\n\n It was very quick, Guy. She wasn't\n\n hurt in any way. It was all over in\n\n no time.\n\nCLOSEUP GUY\n\nHe is horrified. He looks swiftly down at the glasses in\n\nhis hand, then back to Bruno.\n\n BRUNO'S VOICE\n\n (bragging)\n\n I know you'd be surprised. Nothing\n\n for us to worry about. Nobody saw\n\n me, only Miriam.\n\nTWO SHOT\n\nGuy can hardly believe what he is hearing.\n\n BRUNO\n\n I was very careful. Even when I\n\n dropped your lighter there, I went\n\n right back to it up. If It'd been\n\n found, it would have ruined our whole\n\n scheme, wouldn't it?\n\n GUY\n\n Are you trying to tell me you've --\n\n Why, you maniac!\n\n Converted to PDF by www.screentalk.org 42.\n\n BRUNO\n\n (looks at Guy with\n\n astonishment)\n\n But, Guy, you wanted it! We planned\n\n it on the train together, remember?\n\nGuy suddenly starts to go. Bruno grabs his arm.\n\n BRUNO\n\n Where are you going?\n\n GUY\n\n Where do you think I'm going? I'm\n\n going to call the police, of course.\n\n BRUNO\n\n But you can't, Guy. We'd both be\n\n arrested for murder.\n\nGuy turns back slowly and faces him.\n\n GUY\n\n We'd both be arrested for murder?\n\n BRUNO\n\n You're is much in it as I am. We\n\n planned it together. Criss-cross.\n\n I do your murder --\n\n GUY\n\n (suddenly angry)\n\n You crazy fool! You think you can\n\n get away with that?\n\n BRUNO\n\n (a little hurt)\n\n Oh, come now, Guy. Why should I go\n\n to Metcalf and kill a total stranger,\n\n unless it was part of the plan and\n\n you were in on it? You're the one\n\n that benefits, Guy. You're a free\n\n man. I didn't even know the girl.\n\nGuy makes a move to leave, but Bruno holds on tight.\n\n GUY\n\n Let me go, Bruno. I had nothing to\n\n do with this and the police will\n\n believe me.\n\n Converted to PDF by www.screentalk.org 43.\n\n BRUNO\n\n (concerned)\n\n If you go to the police now, you'll\n\n just be turning yourself in as in\n\n accessory. You see, you have the\n\n motive.\n\nAt this moment both turn at a sound across the street.\n\nLONG SHOT (FROM THEIR VIEWPOINT)\n\nWe hear the sound of a telephone ringing in Guy's apartment.\n\nThe top of one of his windows is open.\n\n BRUNO\n\n What is it?\n\n GUY\n\n My telephone.\n\n BRUNO\n\n (amused)\n\n Someone has some news for you, Guy.\n\nGuy still stares across the street.\n\nLONG SHOT (FROM HIS VIEWPOINT)\n\nWe see a police car pull up outside Guy's apartment.\n\nTWO SHOT\n\nBruno pulls Guy back further into the shadows. Guy\n\ninstinctively flattens himself against the wall. He looks\n\nacross the street again.\n\nLONG SHOT (FROM HIS VIEWPOINT)\n\nWe see the two policemen go into his apartment building.\n\nTWO SHOT\n\nGuy is still flattened against the wall to keep out of light.\n\n BRUNO\n\n Tell them you know about it already,\n\n Guy.\n\n Converted to PDF by www.screentalk.org 44.\n\nCLOSEUP GUY\n\nHe looks across at the police, then down at himself with\n\nsome surprise and disgust, then over at Bruno, suddenly\n\nconscious he is behaving like a criminal and that Bruno is\n\nresponsible for his predicament.\n\n GUY\n\n (muttering)\n\n You've got me acting, like a criminal,\n\n you crazy fool!\n\nBruno for a moment looks menacingly at Guy.\n\n BRUNO\n\n Don't you call me that.\n\nBruno's flare of anger dies. They both look again across\n\nthe street.\n\nLONG SHOT (FROM THEIR VIEWPOINT)\n\nThe two policemen come out of the house, get into their car\n\nand drive off.\n\nGuy's telephone is still ringing.\n\nTWO SHOT\n\n BRUNO\n\n You must be tired, Guy. I know I\n\n am. I've sure had a strenuous\n\n evening.\n\nGuy looks at him, almost numb.\n\n BRUNO\n\n Now look, Guy, about my father. I\n\n have the plans made. Two plans. A\n\n plan of the grounds and a plan of\n\n the house. I have in old Luger I\n\n bought at a pawn shop in San\n\n Francisco. My father --\n\nGuy turns and starts to move in across the street.\n\nTWO SHOT\n\nBruno follows Guy and we FOLLOW them across the street.\n\nCAMERA ON THEIR BACKS. Guy strides ahead to the house.\n\n Converted to PDF by www.screentalk.org 45.\n\n BRUNO\n\n Wait a minute, Guy. To have to talk.\n\n We have to arrange things.\n\nGuy turns at the door to his apartment building.\n\n GUY\n\n (furiously)\n\n Get away before I give you what you\n\n gave Miriam.\n\n BRUNO\n\n (sadly)\n\n You're not yourself, Guy. You're\n\n tired. When you think things over,\n\n you'll see I'm right. Tomorrow --\n\nGuy opens his door, turns on Bruno.\n\n GUY\n\n (with finality)\n\n I don't know you. I never saw you\n\n before. I never want to see you\n\n again.\n\nHe goes in and slams the door in Bruno's face.\n\n BRUNO\n\n (to the closed door)\n\n But we have to --\n\nHe realizes there is no use in trying to talk to Guy any\n\nfurther. He turns and faces the CAMERA IN CLOSE UP as he\n\nmoves away, looking sad almost to the point of tears.\n\nINT. GUY'S APARTMENT\n\nGuy is standing at the telephone which is still ringing. He\n\nhas Miriam's glasses in his hand. He looks down at them for\n\na moment, then picks up the receiver. He hesitates, then\n\nspeaks into the phone.\n\n GUY\n\n (hoarsely, into phone)\n\n Yes?\n\n (Pause)\n\n Yes, Anne. I'm sorry, darling. I\n\n just got in.\n\n (pause)\n\n Of course I'm all right.\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 46.\n\n GUY (CONT'D)\n\n (forcing his voice to\n\n sound normal)\n\n But you sound upset. Is anything\n\n wrong?\n\n (Pause)\n\n All right. I'll come over. Right\n\n away.\n\nHe hangs up but keeps his hand on the telephone, deliberating.\n\nHe starts to dial, then suddenly hangs up and starts out.\n\n DISSOLVE TO:\n\nEXT. A RESIDENTIAL STREET, WASHINGTON LONG SHOT NIGHT\n\nA taxi drives up and stops in front of a handsome residence.\n\nIt is the Burton home. Guy gets out of the taxi and goes up\n\nthe steps.\n\nMED. SHOT OVER GUY'S SHOULDER\n\nHis figure tense, he rings the bell. After a moment's wait,\n\nthe door is opened from inside and Anne Burton stands in the\n\nlighted hallway. She looks at Guy with an anxious, taut\n\nexpression, searches his face hastily, then as he takes a\n\nstep inside she is suddenly in his arms. They embrace with\n\nwordless fervor.\n\n GUY\n\n (holding her close)\n\n Anne darling, you're trembling.\n\nAnne draws back and looks into his face as if searching for\n\nan answer to some question in her mind.\n\n ANNE\n\n Guy --\n\n (her fingers gently\n\n touch his face)\n\n I wonder if you know how much I love\n\n you.\n\nGuy takes her hand from hIs face, caresses it with his lips.\n\n GUY\n\n (forcing a smile)\n\n Brazen woman. I'm the one to say\n\n that.\n\n Converted to PDF by www.screentalk.org 47.\n\n ANNE\n\n (tensely)\n\n But I wanted you to know, before...\n\n (forcing herself to\n\n be calm)\n\n Before we go into the living room.\n\nFather wants to see you.\n\nCLOSEUP GUY\n\nHe looks apprehensively in direction of the living room,\n\nconscious of what the news is to be, but covering up.\n\nLONG SHOT LIVING ROOM FROM GUY'S VIEWPOINT\n\nSENATOR BURTON and BARBARA BURTON are seated near a desk on\n\nthe farthest side of the room. Senator Burton is a\n\ndistinguished fifty, a man with great pride in tradition,\n\nhis family and his career. Barbara, Anne's younger sister,\n\nis a lively seventeen who loves excitement, says exactly\n\nwhat she thinks and rarely thinks before she says it.\n\nSuperficially, in height and figure, she resembles Miriam.\n\nShe also weirs glasses. By her gestures we gather she is\n\nspeaking urgently, but softly, to her father, who lifts a\n\nweary hand to quiet her as she looks toward Guy in the\n\nhallway, Barbara keeps quiet and also looks toward Guy.\n\nThey both wait for him to enter.\n\nCLOSEUP GUY\n\nHe steels himself for the long walk across the hall and the\n\nliving room.\n\nCLOSEUP ANNE\n\nWatching Guy closely.\n\nMED. SHOT\n\nAs Guy starts to make the long trek across the living room,\n\nwith Anne behind him --\n\n GUY\n\n (stiffly)\n\n Good evening, sir. Hello, Babs.\n\n Converted to PDF by www.screentalk.org 48.\n\nBarbara has been squirming in her seat, then as if jet\n\npropelled she catapults out of it and runs to Guy, giving\n\nhim a big hug and a smack on the cheek.\n\n BARBARA\n\n Something awful has happened, Guy.\n\n SENATOR\n\n (firmly)\n\n Sit down, Barbara.\n\nSubdued, she sits down. But Guy remains standing.\n\n SENATOR\n\n (finding it difficult\n\n to begin)\n\n There seems to be no way of\n\n diplomatically breaking tragic news.\n\n I'm sorry, Guy, to be the one to\n\n tell you. It concerns your wife.\n\n She's been murdered.\n\nGuy stares woodenly at the Senator, is if hypnotized.\n\n BARBARA\n\n The police have been using everything\n\n but radar to locate you.\n\n SENATOR\n\n You're to call Headquarters at\n\n Metcalf.\n\nThe full impact of what has happened hits Guy once more.\n\n GUY\n\n Miriam...murdered.\n\n ANNE\n\n (with inner tension)\n\n She was...strangled.\n\nSlowly Guy's eyes meet hers. They are remembering what he\n\nsaid on the phone: "I could strangle her." He sinks into a\n\nchair. The Senator is quite distressed.\n\nDuring the following scene Barbara quietly goes about the\n\nbusiness of pouring drinks and serving them. She knows\n\neveryone's preference.\n\nConverted to PDF by www.screentalk.org 49.\n\n SENATOR\n\n (wrylt, to Guy)\n\n It happened on an island in an\n\n amusement park. It was sort of a\n\n lovers lane, I believe. A rather\n\n sordid atmosphere.\n\n BARBARA\n\n (quickly, to Guy)\n\n Miriam went there with two boys.\n\n They were the ones who found her.\n\n So they're not suspects. But you\n\n probably will be.\n\n SENATOR\n\n Young lady, we can't overlook the\n\n fact that murder is at our doorsteps.\n\n But I forbid you to drag it into the\n\n living room!\n\n BARBARA\n\n (wide-eyed)\n\n Let's not fool ourselves. The police\n\n will say Guy wanted Miriam out of\n\n the way so he could marry Anne. In\n\n a crime of this sort the police first\n\n go after the husband, and Guy had\n\n every motive.\n\n SENATOR\n\n (aghast)\n\n Motive?\n\n GUY\n\n (quietly)\n\n She's right. Whichever way you look\n\n at it...I'm in a spot.\n\n SENATOR\n\n (disconcerted but\n\n whistling in dark)\n\n Oh come now, my boy. I'm sure you\n\n have nothing to worry about.\n\n BARBARA\n\n (flatly)\n\n If he hasn't an alibi for nine-thirty\n\n tonight he has plenty to worry about.\n\n Converted to PDF by www.screentalk.org 50.\n\n ANNE\n\n (who hasn't taken\n\n anxious eyes off Guy)\n\n You can tell them where you were,\n\n can't you, Guy?\n\n GUY\n\n (wearily)\n\n At nine-thirty I was on the train\n\n from New York to Washington.\n\n SENATOR\n\n (relieved)\n\n There you are.\n\n BARBARA\n\n Who saw you? Did you speak to anyone?\n\n You'll need a Witness, you know.\n\n GUY\n\n (as if it didn't matter)\n\n Yes, I spoke to someone.\n\n SENATOR\n\n (hopefully)\n\n Anyone you know?\n\n GUY\n\n No. His name was Collins. He is a\n\n professor.\n\n SENATOR\n\n (brightening)\n\n Harvard.\n\n GUY\n\n University of Virginia.\n\nThe Senator's expression says: "Well, that's not too bad."\n\nCLOSEUP ANNE\n\nHer face shows her relief that Guy can account for his time.\n\n ANNE\n\n Then everything's's all right.\n\n Converted to PDF by www.screentalk.org 51.\n\nBACK TO SCENE\n\n BARBARA\n\n Not quite. Detectives play a game\n\n called Motive, Motive, Who'd got the\n\n Motive.\n\n ANNE\n\n (near the breaking\n\n point)\n\n I'm sick of hearing that word!\n\n BARBARA\n\n He'll still have to answer questions.\n\n SENATOR\n\n Routine. Pure routine.\n\n GUY\n\n I'm afraid there'll be a lot of\n\n reporters at your front door in the\n\n morning.\n\n BARBARA\n\n Daddy doesn't mind a little scandal.\n\n He's a senator.\n\n ANNE\n\n (answering Guy's look)\n\n It can't be helped, darling. It is\n\n not your fault. It's not as though\n\n anyone can say you had something to\n\n do with it.\n\n GUY\n\n Someone might say it...I'd do anything\n\n to keep you all out of this mess.\n\n SENATOR\n\n Profit by my experience, Guy. Never\n\n lose any sleep over accusations.\n\n (an afterthought)\n\n Unless they can be proved, of course.\n\n We'll help all we can. Dreadful\n\n business, dreadful. That poor\n\n unfortunate girl.\n\n BARBARA\n\n (flatly)\n\n She was a tramp.\n\n Converted to PDF by www.screentalk.org 52.\n\n SENATOR\n\n (pontificially)\n\n She was a human being. let me remind\n\n you that even the most unworthy of\n\n us has the right to life and the\n\n pursuit of happiness.\n\n BARBARA\n\n (unimpressed)\n\n From what I hear, she pursued it in\n\n all directions.\n\n SENATOR\n\n Barbara!\n\n ANNE\n\n Father, it's getting terribly late,\n\n and Guy looks so tired...\n\n SENATOR\n\n (quickly)\n\n Of course, of course. Back to bed,\n\n Barbara.\n\n BARBARA\n\n (ignoring this - to\n\n Anne and Guy)\n\n Well, you two. Nothing stands in\n\n your way now. You can be married\n\n right away. Think of it -- you're\n\n free!\n\nCLOSE TWO ANNE AND GUY\n\nlook at one another with a growing realization of what\n\nMiriam's death actually means to their happiness -- they are\n\nfree.\n\nBACK TO SCENE\n\nThe Senator firmly urges Barbara to the door.\n\n SENATOR\n\n (to Barbara)\n\n One doesn't always have to say what\n\n one thinks!\n\n BARBARA\n\n (sweetly)\n\n Father, I'm not a politician.\n\n Converted to PDF by www.screentalk.org 53.\n\nThe Senator gives her a gentle but firm push out of sight.\n\n SENATOR\n\n You won't forget that call, Guy?\n\n Captain Turley.\n\n GUY\n\n Yes sir. Goodnight.\n\nBarbara pokes her head quickly around the door.\n\n BARBARA\n\n I still think it would be wonderful\n\n to have a man love you so much he'd\n\n kill for you.\n\n (she ducks out)\n\nTWO SHOT\n\nLeft alone, Guy and Anne embrace. Anne's nervous tension\n\ncomes to the surface in a flood of relief.\n\n ANNE\n\n I told myself over and over I was\n\n being silly, but there was one\n\n horrible moment tonight when the\n\n news came through. I kept remembering\n\n what you shouted telephone from\n\n Metcalf.\n\n GUY\n\n That I could strang...\n\nAnne quickly puts her fingers over his mouth.\n\n ANNE\n\n Don't even say it. Forget you ever\n\n said it. Even more terrifying than\n\n the murder itself, Guy, was the awful\n\n thought that if you had anything to\n\n do with it we'd be separated, -perhaps\n\n forever. I'd never see you again.\n\n I couldn't bear it.\n\n DISSOLVE TO:\n\nLONG SHOT MAIN STREET OF METCALF DAY\n\nwith its customary mid-afternoon activity.\n\n LAP DISSOLVE TO:\n\n Converted to PDF by www.screentalk.org 54.\n\nEXT. METCALF POLICE HEADQUARTERS DAY\n\nA knot of people are hanging around the entrance, including\n\na few newspaper photographers. There is a rush of interest\n\nwhen a taxi pulls up and Guy steps out of it. Guy pushes\n\nhis way through the people. Two or three bulbs flash. There\n\nis a murmur from the crowd and we hear Guy's name. He passes\n\ninto the entrance.\n\nINT. CORRIDOR OF POLICE HEADQUARTERS\n\nGuy comes into the corridor from the street and approaches\n\ntwo policemen who are standing nearby.\n\n GUY\n\n Captain Turley's office?\n\nOne of the policemen gestures to a door at the right. Guy\n\ncrosses and enters.\n\nINT. RECEPTION ROOM OUTSIDE CAPTAIN TURLEY'S OFFICE\n\nAt one side of the room is a young police sergeant seated at\n\na typewriter. A group of people are seated in chairs lined\n\nagainst the opposite wall.\n\nGuy enters, crosses to the sergeant at the desk.\n\n GUY\n\n Captain Turley is expecting me. Guy\n\n Haines.\n\n SERGEANT\n\n Just a moment, Mr. Haines.\n\nHe rises and goes into an adjoining room.\n\nCLOSEUP GUY\n\nHe now has time to take stock of the waiting people. He\n\ncatches his breath when he sees:\n\nCLOSEUP MRS. JOYCE\n\nMiriam's mother, dressed all in black, is seated in one of\n\nthe chairs. She has been staring at the floor, but brings\n\nher eyes up slowly to glare at Guy with a look of burning\n\nhatred.\n\n Converted to PDF by www.screentalk.org 55.\n\n MRS. JOYCE\n\n (a fierce whisper)\n\n You'll pay for this!\n\nCLOSEUP MR. HARGREAVES\n\nMr. Hargreaves from the music shop looks across at Guy,\n\nattempts in awkward nod but is very embarrassed.\n\nCLOSEUP GUY\n\nGuy nods in returns.\n\nMED. SHOT\n\nThe two boys who were with Miriam at the amusement park.\n\nThey look at Guy with interest.\n\nMED. SHOT GUY\n\nHe looks about him uncomfortably, then turns suddenly as he\n\nsees:\n\nMED. SHOT\n\nSeated behind Guy, apart from the others who are waiting, is\n\nProfessor Collins, Guy's drunken companion on the train of\n\nthe night before. The professor is completely sober now,\n\ndignified and erect. He has removed his glasses to polish\n\nthem and does not react to Guy's presence.\n\nCLOSEUP GUY\n\nGuy starts with a smile of recognition to say, "How do you\n\ndo?" but at that moment he hears the door open and his name\n\ncalled:\n\n SERGEANT'S VOICE\n\n Will you come in, please, Mr. Haines?\n\nMED. SHOT\n\nGuy breaks away from his uncompleted greeting to the professor\n\nand goes through the door to Captain Turley's office, followed\n\nby the eyes of the waiting people.\n\n Converted to PDF by www.screentalk.org 56.\n\nINT. CAPTAIN TURLEY'S OFFICE\n\nCAPTAIN TURLEY is conscientious, methodical and always polite.\n\nHe puts aside photographs and records and rises from behind\n\nhis desk as Guy comes in. A detective lieutenant, CAMPBELL,\n\nis attending a coffee maker. Their expressions are grave by\n\ncontrast with Guy's confident attitude after seeing the\n\nprofessor in the waiting room.\n\n CAPTAIN TURLEY\n\n Good of you to be so prompt, Mr.\n\n Haines. This is Lieutenant Campbell.\n\n (the two nod to each\n\n other)\n\n Won't you sit down?\n\n GUY\n\n Thank you, sir.\n\n (he sits)\n\n CAPTAIN TURLEY\n\n I know you're a busy man, so we won't\n\n detain you any longer than\n\n necessary...Now you already been\n\n good enough to tell us where you\n\n were last evening, and we've managed\n\n to locate the gentleman you spoke\n\n with on the train.\n\nTurley signals to Campbell to call the professor in.\n\n GUY\n\n (brightening)\n\n Yes. I saw him outside.\n\n CAMPBELL\n\n (at open door)\n\n Will you come in please, professor?\n\nCLOSEUP GUY\n\nHe looks up eagerly.\n\nMED. SHOT\n\nProfessor Collins comes in and sits in a chair opposite Guy.\n\n TURLEY\n\n Professor Collins, this is Mr. Haines.\n\n He was with you on the train last\n\n night.\n\n Converted to PDF by www.screentalk.org 57.\n\nThe professor studies Guy for a moment, then awkwardly turns\n\nto Turley.\n\n COLLINS\n\n I'm terribly sorry, but I really\n\n don't remember meeting this gentleman.\n\nCLOSEUP GUY\n\nSurprised. His confident expression fades.\n\nCLOSEUP PROFESSOR COLLINS\n\nHe turns from the captain to Guy.\n\n COLLINS\n\n (apologetically)\n\n Unfortunately, I remember very little\n\n about the journey from New York...You\n\n see, there had been a little\n\n celebration --\n\nMED. SHOT GROUP\n\nGuy interrupts with a slight note of impatience.\n\n GUY\n\n But we were sitting opposite each\n\n other in the observation car! You\n\n were singing a song about a goat --\n\n COLLINS\n\n (incredulously)\n\n A goat?\n\n GUY\n\n (urgently)\n\n And calculus. You were going over a\n\n speech you'd made.\n\nTurley and Campbell are watching closely.\n\n COLLINS\n\n I was? I'm sorry, Mr. Halnes.\n\n (shakes his head)\n\n I certainly must have celebrated! I\n\n can't remember you at all.\n\n Converted to PDF by www.screentalk.org 58.\n\nCLOSEUP GUY\n\nMomentarily Guy is frustrated, then he turns quietly to\n\nTurley.\n\n GUY\n\n (calmly, logically)\n\n Captain, is it so important whether\n\n or not Professor Collins remember\n\n me? Surely, the important thing is\n\n that I've been able to name a man\n\n who was on the train with me. You've\n\n been able to find him. Isn't that\n\n proof of where I was at nine-thirty\n\n last night?\n\nGuy asks this question with a look of near triumph that he\n\nhas clearly established his alibi.\n\n DISSOLVE TO:\n\nINT. BURTON LIVING ROOM EVENING\n\nThe Burtons are having coffee. Barbara has been glancing\n\nthrough a new murder mystery with a lurid cover. As Guy\n\nenters, Anne rises to greet him.\n\n ANNE\n\n Hello, darling. Have you had your\n\n dinner?\n\n GUY\n\n On the train.\n\n ANNE\n\n You weren't in Metcalf all this time?\n\n We expected you hours ago.\n\n BARBARA\n\n (flatly)\n\n I didn't. They sometimes throw a\n\n suspect in the can and keep him there\n\n all night.\n\n SENATOR\n\n (after a disapproving\n\n glance at Barbara)\n\n Sit down, Guy. Sit down. Give him\n\n some coffee, Anne.\n\n (MORE)\n\nConverted to PDF by www.screentalk.org 59.\n\n SENATOR (CONT'D)\n\n (back to Guy)\n\n You had no trouble with the police\n\n of course, once they verified your\n\n alibi?\n\n GUY\n\n (morosely)\n\n When an alibi is full of bourbon,\n\n sir, it can't stand up.\n\n BARBARA\n\n You mean the professor was boiled?\n\n GUY\n\n Completely. He didn't remember me.\n\n ANNE\n\n But, you knew he was on the train!\n\n Wasn't that enough to prove you were\n\n on it, too?\n\n GUY\n\n Apparently not at the right time.\n\n They suggested I could have caught\n\n the train at Baltimore after Miriam\n\n was murdered. They had it all worked\n\n out --\n\n (taps his head)\n\n in their timetables.\n\n ANNE\n\n (growing indignant\n\n and increasingly\n\n nervous)\n\n That's ridiculous. They're acting\n\n as if you were guilty.\n\n BARBARA\n\n (somewhat subdued and\n\n trying to be\n\n comforting)\n\n Everything will be all right, Anne.\n\n The police were just being thorough --\n\n (she's unsure of\n\n herself, and defers\n\n to the senator)\n\n Weren't they, daddy?\n\n SENATOR\n\n I certainly hope so.\n\n (to Guy)\n\n What is your next step?\n\n Converted to PDF by www.screentalk.org 60.\n\n GUY\n\n (wryly)\n\n Whatever it is, the police will know\n\n it. They gave me a present -- come\n\n take a look.\n\nHe crosses to the window, lifts the curtain slightly, then\n\nturns back to the others.\n\n GUY\n\n (continuing)\n\n My guardian angel.\n\nThe group move to look out the window, the senator with\n\nreluctance.\n\nLONG SHOT EXT. STREET FROM THEIR VIEWPOINT\n\nThrough the window we see the figure of a man across the\n\nstreet. He is lighting a cigarette and strolling up and\n\ndown.\n\nBACK TO GROUP\n\n BARBARA\n\n (impressed)\n\n You're being tailed!\n\n GUY\n\n (turning to them)\n\n That's Leslie Hennessy. He works\n\n sixteen hours a day. Somebody else\n\n takes over for the next eight.\n\n (drops the curtain,\n\n turns back into room)\n\n As a matter of fact, Hennessy's a\n\n very nice fellow.\n\n BARBARA\n\n Shouldn't we ask him in for Coffee --\n\n or something?\n\nNobody bothers to answer her. The Senator is disturbed, but\n\nconfident of his own prestige as he goes back to his coffee.\n\n SENATOR\n\n I'll have him called off immediately\n\n of course.\n\n Converted to PDF by www.screentalk.org 61.\n\n GUY\n\n (calmly)\n\n I'm afraid where I go, Hennessy goes.\n\n Even to the Senate.\n\n SENATOR\n\n (Pausing with his cup\n\n hallway to his mouth)\n\n Is he likely to -- picket my office?\n\n GUY\n\n Very likely.\n\nThe Senator's cup is suddenly back on its saucer and he is\n\non his feet, pacing nervously.\n\n SENATOR\n\n I would suggest, Guy, for your own\n\n peace of mind, of course, that you\n\n work here at the house for a few\n\n days.\n\n (a pause)\n\n It would be less embarrassing for\n\n you.\n\nGuy has been looking at Anne and is concerned at the worry\n\non her face. He nods in assent to the Senator's suggestion,\n\nbut puts his hand over Anne's.\n\n GUY\n\n (hopelessly)\n\n Then what about practicing? Perhaps\n\n I'd better forget Forest Hills?\n\n SENATOR\n\n My dear boy, wouldn't it look rather --\n\n awkward -- if you suddenly canceled\n\n all your plans.\n\n ANNE\n\n He's right, Guy. You mustn't do\n\n anything that would look suspicious.\n\n You've got to carry on as though\n\n nothing has happened.\n\n BARBARA\n\n (pointing out the\n\n window)\n\n Escorted by Mr. Hennessy.\n\nThe are crestfallen again. RANDALL, the manservant, has\n\nentered with the telephone.\n\n Converted to PDF by www.screentalk.org 62.\n\n RANDALL\n\n A call for you, Mr. Haines. They\n\n say it is urgent.\n\nThe phone is plugged in to a connection and Guy crosses the\n\nroom and picks up the receiver. The Burtons watch him.\n\n GUY\n\n Hello --\n\nINT. TELEPHONE BOOTH BIG HEAD CLOSEUP OF BRUNO\n\nHis face wears the most affable expression.\n\n BRUNO\n\n Hello, Guy. I tried your apartment,\n\n but --\n\n (pause)\n\n Why, Guy, this is Bruno!\n\nINT. BURTON LIVING ROOM\n\nGuy hangs up the telephone quickly. He looks at the others,\n\nawkwardly tries to explain:\n\n GUY\n\n Must be some mistake. It wasn't for\n\n me.\n\nHis embarrassment grows as Anne looks at him with a puzzled\n\nexpression.\n\n FADE OUT.\n\nFADE IN\n\nEXT. WASHINGTON STREET APPROACHING JEFFERSONS MEMORIAL DAY\n\nGuy and HENNESSY are walking along the street together, CAMERA\n\nMOVING WITH THEM. Their relationship is most friendly. Guy\n\ncarries a briefcase. Hennessy is an amiable but not gullible\n\nyoung man in his early thirties. He knows his job, is well\n\ngroomed, well educated, and well liked.\n\n GUY\n\n Well, I suppose I was pretty lucky\n\n to be seeded fifth, really.\n\n Converted to PDF by www.screentalk.org 63.\n\n HENNESSY\n\n I've never seen the Forest Hillss\n\n tournament before. I'm looking\n\n forward to it.\n\n GUY\n\n (wryly)\n\n Do you mean we'll be going there\n\n together, Hennessy?\n\n HENNESSY\n\n Oh, don't worry. This thing will be\n\n cleared up by that time.\n\n (changes the subject)\n\n Ever thought of turning professional,\n\n Guy?\n\n GUY\n\n I won't have to do that. When I'm\n\n through with tennis. I'll be going\n\n into politics, I hope.\n\n HENNESSY\n\n (aghast)\n\n Politics! It's a good thing for you\n\n I don't report that to the chief.\n\nHe turns to light a cigarette. As he does, Guy gives a barely\n\nperceptible start at what he sees offscene.\n\nLONG SHOT JEFFERSON MEMORIAL FROM GUYS VIEWPOINT\n\nThe tiny figure of a man is standing at the base of the tall\n\nwhite column. The figure lifts in arm and waves. Instinct\n\ntells us that this is Bruno. Hennessy is still mumbling his\n\nopinion of politics.\n\n HENNESSY'S VOICE\n\n If he knew you were getting into\n\n that rat-race --\n\nTWO SHOT GUY AND HENNESSY\n\nGuy turns his back on Bruno's figure and looks frantically\n\ntoward to street, wanting to get away.\n\n HENNESSY\n\n -- He'd put ten men on your trail.\n\n He says --\n\n Converted to PDF by www.screentalk.org 64.\n\n GUY\n\n (interrupts)\n\n Let's take this cab. It's getting\n\n late.\n\nHe hails a taxi which is cruising by, and they start to get\n\nin. Guy directs the driver.\n\n GUY\n\n Pentagon Building, please.\n\n HENNESSY\n\n Oh, no, not there! I always get\n\n lost.\n\nINT. TAXI CLOSE SHOT\n\nGuy turns and looks out of the window.\n\nLONG SHOT JEFFERSON MEMORIAL\n\nfrom Guy's viewpoint, shot through the cab window. Again we\n\nsee the solitary figure of Bruno looking after Guy and\n\nbeginning to recede with the background as the cab starts\n\noff.\n\n DISSOLVE TO:\n\nINT. GUY'S APARTMENT NIGHT\n\nAs Guy comes in from outside, there is a note on the floor\n\nthat has been pushed under the door. Guy picks it up, stares\n\nat it for a minute before he opens it. He takes out a\n\nhandwritten note and reads it with an expression of disgust.\n\nINSERT NOTE (IN GUY'S HANDS)\n\nIT READS:\n\nDear Guy:\n\nWe have to meet and make plans.\n\nCall me at Arlington ----.\n\nTime's getting short.\n\n Bruno\n\nThe handwriting is sprawling and erratic, embellished with\n\nconceited flourishes.\n\n Converted to PDF by www.screentalk.org 65.\n\nMEDIUM SHOT\n\nGuy looks off for a moment with set face, then tearing the\n\nnote into shreds, crosses to a small desk, lights a match\n\nand holds it to the fragments, letting them burn and fall\n\ninto an ash tray.\n\nGuy looks off for a moment with set face, then tearing the\n\nnote into shreds, crosses to a small desk, lights a match\n\nand holds it to the fragments, letting them burn and fall\n\ninto an ash tray.\n\n DISSOLVE TO:\n\nLONG SHOT EXT. MELLON GALLERY LATE AFTERNOON\n\nCAMERA is in a low setup, to take in the sign across the\n\ndoorway which identifies the gallery. Hennessy stands in\n\nthe foreground in front of the building, on duty.\n\n LAP DISSOLVE TO:\n\nINT. MELLON GALLERY\n\nGuy and Anne are walking slowly through a more or less\n\ndeserted room of the gallery. Their manner is relaxed and\n\nintimate.\n\n ANNE\n\n Well, we'd better be getting back.\n\n GUY\n\n We've actually been alone for an\n\n hour. Seems almost indecent. You\n\n like?\n\n ANNE\n\n (softly)\n\n I like.\n\n GUY\n\n I was beginning to feel like a\n\n goldfish.\n\n ANNE\n\n So was I. When we build our house,\n\n darling, we won't even have glass\n\n windows. No doorbells, no newspapers,\n\n no telephone --\n\n Converted to PDF by www.screentalk.org 66.\n\n GUY\n\n No Hennessy.\n\n ANNE\n\n (suddenly serious)\n\n How long can it go on?\n\n GUY\n\n I don't know. I suppose until they\n\n find out who did it.\n\n ANNE\n\n We'll be happier then, won't we?\n\n GUY\n\n I suppose so.\n\nAnne looks it him, surprised at his lack of enthusiasm.\n\nThey walk on out of the picture.\n\nA figure steps out from behind a pillar in the main hall of\n\nthe gallery, near the spot from which they have disappeared.\n\nIt is Bruno. He calls.\n\n BRUNO\n\n (softly)\n\n Guy!\n\nAnne stops and looks back. Guy knows who it is and would\n\nnot turn but that he is forced to by Anne's action. He takes\n\na few steps towards Bruno.\n\nCLOSEUP\n\nAnne watches Guy approach this stranger. She looks downward\n\nat Bruno's tie pin.\n\nCLOSEUP\n\nBruno's tie pin, bearing his name, gleams in the light.\n\nCLOSEUP\n\nAnne reads the name on the tie pin.\n\nTWO SHOT\n\nGuy comes up to Bruno, steps in front of him.\n\n Converted to PDF by www.screentalk.org 67.\n\n GUY\n\n (muttering harshly)\n\n Will you stop pestering me!\n\n BRUNO\n\n But Guy, you haven't called me. My\n\n father's leaving for Florida the end\n\n of this week --\n\n GUY\n\n (interrupts)\n\n You crazy fool! There's a detective\n\n outside. He'll see us together!\n\n BRUNO\n\n (brushing this off)\n\n Oh, they can't have anything on you.\n\n (looking past Guy)\n\n Isn't that Anne Burton? Slight\n\n improvement over Miriam -- eh, Guy?\n\n GUY\n\n Stay away from me, I tell you!\n\nHe leaves Bruno abruptly to rejoin Anne. Bruno looks after\n\nhim, a little hurt.\n\nTWO SHOT\n\nGuy rejoins Anne and they start to walk away.\n\n ANNE\n\n Who was it, Guy?\n\n GUY\n\n (unnerved)\n\n I never saw him before. Just some\n\n tennis fan.\n\nAnne looks at him a little oddly. He seems unduly concerned\n\nabout a casual stranger.\n\nCLOSEUP ANNE\n\nHer face is troubled.\n\n FADE OUT.\n\n Converted to PDF by www.screentalk.org 68.\n\nFADE IN\n\nINT. MORTON STUDY MED. SHOT\n\nGuy and a secretary have set up office in the Morton study.\n\nAs the scene opens the secretary is handing Guy a large\n\nenvelope.\n\n SECRETARY\n\n Here's a special delivery, Mr. Haines.\n\n It's marked personal.\n\nAs Guy is opening the envelope, Barbara speaks to him from\n\natop a library ladder. She is getting a book from one of\n\nthe top shelves of a bookcase, which is next to a window.\n\n BARBARA\n\n Are you getting in any practice today,\n\n Guy?\n\n GUY\n\n (as he takes out a\n\n large folded sheet\n\n of paper and glances\n\n at it, mystified)\n\n Yes, if I can get a court at the\n\n club.\n\nAs Guy's hands unfold the paper and hold it for moment, we\n\nsee that it is a diagrammed plan of the grounds and the\n\nInterior of the Anthony house. There are dotted lines along\n\nthe upper hall, with an arrow which points to one room and\n\nwhere Bruno has indicated in his handwriting, "My father's\n\nroom." Over this we hear the voices of Barbara and the\n\nsecretary:\n\n SECRETARY'S VOICE\n\n Barbara, who are you waving at?\n\n BARBARA'S VOICE\n\n Mr. Hennessy. I think it is a shame\n\n Daddy won't let us have him in the\n\n house to sit down. Have you met him\n\n yet, Louise?\n\n SECRETARY'S VOICE\n\n No.\n\n BARBARA'S VOICE\n\n He is awfully cute.\n\n Converted to PDF by www.screentalk.org 69.\n\nMED. SHOT\n\nGuy frowns, quickly folds the paper up and stuffs it into\n\nhis pocket. He looks off abstractedly.\n\nCLOSEUP SECRETARY\n\nShe looks at Guy sympathetically.\n\n SECRETARY\n\n Is anything wrong, Mr. Haines?\n\nCLOSEUP GUY\n\nHer voice breaks his reverie. He answers her with a forced\n\nsmile.\n\n GUY\n\n No, thank you, Louise.\n\n FADE OUT.\n\nFADE IN\n\nTENNIS COURT AT WASHINGTON COUNTRY CLUB\n\nThere are twenty or thirty people sitting in the bleacher\n\nseats opposite the umpire's chair. A game of mixed doubles\n\nis in progress.\n\nMED. SHOT AT THE ENTRANCE TO THE COURT\n\nGuy appears, carrying his racquets. His partner for the\n\nforthcoming game, and one or two other players, are close\n\nby.\n\nCLOSER SHOT\n\nGuy looks about him. Several people are looking at him\n\nawkwardly or avoiding his eyes. He moves self-consciously\n\naway, and the CAMERA PANS HIM around the court to the umpire's\n\nchair.\n\nMED. SHOT\n\nA couple of women players whisper something about Guy as he\n\ngoes past them.\n\n Converted to PDF by www.screentalk.org 70.\n\n FIRST WOMAN\n\n I didn't think he'd show up after\n\n what happened.\n\n SECOND WOMAN\n\n And miss all the publicity?\n\nMED. SHOT\n\nAs Guy stands at the umpire's chair, the umpire glances down\n\nand gives him a rather embarrassed greeting.\n\nCLOSEUP GUY\n\nHe looks across at the watching crowd.\n\nMED. SHOT FROM GUY'S VIEWPOINT\n\nThe heads of the people in the bleachers move from side to\n\nside, to follow the play on the court. One head is not\n\nmoving. It is staring at Guy. It is Bruno.\n\nAt this moment, we hear the umpire calling, "Game, set and\n\nmatch" to the winning mixed doubles pair.\n\nCLOSEUP GUY\n\nHis expression becomes set.\n\nLONG SHOT\n\nThe mixed doubles couples complete their handshaking at the\n\nnet and move off the court. We see Guy move up to the base\n\nline while the other player takes his position for the\n\npreliminary knock-up.\n\nMED. SHOT\n\nAs Guy casually knocks the ball across the net, he glances\n\nagain toward Bruno.\n\nMED. SHOT FROM GUY'S VIEWPOINT\n\nBruno is making his way out of the small stand.\n\n Converted to PDF by www.screentalk.org 71.\n\nCLOSEUP GUY\n\nPerplexed and apprehensive as to what Bruno may be up to.\n\nHe hears his opponent's voice.\n\n PLAYER'S VOICE\n\n Ready, Guy?\n\nGuy shakes off his abstraction and poises himself to receive\n\nthe ball.\n\n LAP DISSOLVE TO:\n\nMED. SHOT PASSAGEWAY LEADING TO TERRACE\n\nWe see Guy coming alone, having fInIshed his game. He is\n\ncarrying his rackets, wears a towel around his neck, etcetera.\n\nHe walks into foreground, into CLOSEUP, and suddenly stops\n\nshort at what he sees:\n\nMED. SHOT FROM GUY'S VIEWPOINT\n\nThe group at the table comprising Bruno, Anne and the two\n\nFrench people. Bruno is preening himself as the others laugh\n\nuproariously, obviously at something Bruno has said. Anne\n\ncatches sight of Guy and smiles at him.\n\nCLOSE SHOT GUY\n\nCAMERA MOVES WITH HIM as he comes forward toward the table.\n\nMED. SHOT GROUP AT TABLE\n\nAs Guy comes into the scene. He stands staring.\n\n ANNE\n\n Guy, darling -- this is Mr. Antony --\n\n a friend of Monsieur and Madame\n\n Darville...\n\n (to Bruno)\n\n Guy Haines.\n\nCLOSEUP GUY\n\nHe gives a weak acknowledgment in Bruno's direction, realizing\n\nthat Bruno has wormed his way into the group and that he\n\nmust accept the introduction.\n\n Converted to PDF by www.screentalk.org 72.\n\nMEDIUM SHOT\n\nBruno half rises, smiles affably at Guy, reaches out his\n\nhand. Guy is forced to shake hands with him\n\n BRUNO\n\n I've been a fan of yours for a long\n\n time, Mr. Haines. In fact, I follow\n\n everything you do.\n\n MME. DARVILLE\n\n Mr. Antony has been telling us such\n\n charming stories... Very funny.\n\nCLOSEUP GUY\n\nHe gives another weak little smile.\n\nMED. SHOT\n\nIn response to the Frenchwoman's attentive and eager\n\nexpression, Bruno leans forward on the table and starts saying\n\nsomething more in extremely fluent French.\n\nCLOSEUP ANNE\n\nShe is staring at Bruno with a new expression.\n\nCLOSEUP FROM ANNE'S VIEWPOINT\n\nBruno's coat has spread open a bit, and his tie pin bearing\n\nthe name "Bruno" is resting on the edge of the table.\n\nCLOSEUP ANNE\n\nShe becomes aware that this is the man she has seen call to\n\nGuy in the art museum, that they have met before. Her eyes\n\nturn a little in Guy's direction, though she does not look\n\nat him.\n\nCLOSEUP GUY\n\nHe is still watching Bruno talk to the French couple. Guy\n\nis unaware of Anne's looks. Suddenly his attention is\n\narrested by the sound of Barbara's voice calling him.\n\n Converted to PDF by www.screentalk.org 73.\n\n BARBARA'S VOICE\n\n Guy!\n\nHe turn his head and CAMERA PANS him to Barbara, who is\n\nstanding a few steps from the table beckoning to him.\n\n BARBARA\n\n (Sotto voce)\n\n I've just been talking to your shadow.\n\n (very impressed)\n\n Guy, did you know Mr. Hennessy helped\n\n crack that axe murder I was reading\n\n about? You know, the one where the\n\n body was cut up and hidden in the\n\n butcher shop? He was locked in the\n\n ice box with the left leg for six\n\n hours!\n\n GUY\n\n He pulls those yarns right out of\n\n his hat, Babs.\n\nCLOSEUP GUY\n\nHe gives a sharp look back toward Bruno. There is more\n\nlaughter coming from the French couple at the table.\n\nCLOSE SHOT GROUP AT TABLE FROM GUY'S VIEWPOINT\n\nBruno is occupied with his French joke, but Anne is looking\n\nat Guy strangely.\n\nTWO SHOT GUY AND BARBARA\n\nGuy turns back to Barbara. Barbara looks with interest toward\n\nBruno.\n\n BARBARA\n\n Who's the nice looking Frenchman\n\n with the Darvilles?\n\n GUY\n\n He's not French. His name's Antony.\n\nBarbara steps toward the table.\n\nMED. SHOT AT TABLE\n\nas Barbara joins the group.\n\n Converted to PDF by www.screentalk.org 74.\n\n BARBARA\n\n How do you do, Madame Darville.\n\n Monsieur.\n\nThey looks up.\n\nCLOSEUP BRUNO\n\nBruno stops in the middle of some French to stare at Barbara.\n\nHer voice continues.\n\n BARBARA'S VOICE\n\n How are you?\n\n FRENCH COUPLES' VOICES\n\n Delightful to see you. How sweet\n\n you look, Miss Barbara.\n\nCLOSE SHOT BARBARA FROM BRUNO VIEWPOINT\n\n BARBARA\n\n I hope you aren't forgetting our\n\n little party on Thursday, Madame.\n\nFrom Bruno's viewpoint, as Barbara speaks,CAMERA MOVES IN\n\nCLOSER until to faintest impression of the merry-go-round\n\nfills the screen with the effect of whirling around Barbara's\n\nhead. Her glasses seem to glint until her eyes are\n\nobliterated by the glare.\n\nMED. SHOT THE GROUP\n\n MME. DARVILLE\n\n We are planning on it?\n\n M. DARVILLE\n\n But of course.\n\nAll talk dies out as all eyes turn to Bruno, who is staring\n\nat Barbara. Except Anne's, who is saying quietly to Bruno:\n\n ANNE\n\n This is my sister Barbara. Barbara,\n\n this is Mr. Antony.\n\nCLOSEUP BRUNO\n\nHe does not acknowledge the introduction immediately. He is\n\nstill staring at Barbara. Then he nods abstractedly.\n\n Converted to PDF by www.screentalk.org 75.\n\nCLOSEUP ANNE\n\nShe is looking at Bruno, wondering what mystery lies behind\n\nthis strange individual and why he and Guy have disclaimed\n\nany previous acquaintance.\n\n FADE OUT.\n\nFADE IN\n\nINT. GUY'S APARTMENT NIGHT\n\nCLOSEUP A LUGER PISTOL HELD IN GUY'S HANDS\n\nCAMERA PULLS BACK TO SHOW Guy staring down at it. He is\n\npartially dressed for an evening party, in black bow tie but\n\nwithout his jacket. He leans forward to take up a letter\n\nfrom among brown paper wrappings on the table.\n\nINSERT: LETTER\n\n Dear Guy --\n\n Just two more days left. We must\n\n get together for final details.\n\nThe note, in Bruno's handwriting, is unsigned.\n\nCLOSEUP GUY\n\nHe stares down at the note. At this moment there is a knock\n\nat the door.\n\nMED. SHOT\n\nGuy hastily gather together the gun, the note and the\n\nwrappings and puts them in a dresser drawer. He crosses to\n\nthe door and opens it. Hennessy enters, carrying a topcoat.\n\n GUY\n\n Hiya, Hennessy. Won't keep you out\n\n late tonight.\n\n (getting into his\n\n dinner jacket)\n\n With Forest Hills coming up tomorrow,\n\n I've got to get some sleep.\n\n Converted to PDF by www.screentalk.org 76.\n\n HENNESSY\n\n (helping himself to a\n\n cigarette)\n\n That's too bad. Hammond takes over\n\n in a couple of hours. I'd like to\n\n see him earn his salary.\n\nGuy turns to the dresser drawer in which he has put the note\n\nand the gun, maneuvering his body between the dresser and\n\nHennessy's view. He takes out a handkerchief, closes the\n\ndrawer, sticks the handkerchief in his pocket, speaking as\n\nhe does so.\n\n GUY\n\n Doesn't that bloodhound over relax?\n\n He sticks so close he's beginning to\n\n grow on me -- like a fungus.\n\n HENNESSY\n\n (mildly)\n\n He thinks you're a very suspicious\n\n character. He doesn't trust anybody!\n\n Not even himself.\n\nGuy is eager to get out of the room, and Hennessy is\n\nmaddeningly slow in his movements.\n\n GUY\n\n Come on.\n\n (indicating at Hennessy\n\n overcoat)\n\n Don't forget your sleeping bag.\n\n HENNESSY\n\n (taking his time)\n\n Yeah, If I have to wait too long on\n\n the sidewalk my feet get cold. And\n\n if I sit too long on those stone\n\n steps, my --\n\nGuy has the door open and eases Hennessy toward the haIl.\n\n GUY\n\n (quickly)\n\n Don't worry. Since you told Barbara\n\n Burton about the icebox, you're her\n\n favorite charity. She'll send the\n\n butler out with something to defrost\n\n you.\n\n HENNESSY\n\n (grinning)\n\n Cute kid.\n\n Converted to PDF by www.screentalk.org 77.\n\nHe's gone, and with a last glance at the dresser, Guy goes\n\nout and closes the door.\n\n LAP DISSOLVE TO:\n\nEXT. BURTON HOUSE LONG SHOT NIGHT\n\nThe street outside the Burton house is lined with cars and\n\nlimousines. Various guests are arriving.\n\nMED. SHOT\n\nOn the opposite side of the street we see Hennessy, now\n\nwearing his topcoat. He looks bored as he glances across\n\nthe street to the house.\n\n LAP DISSOLVE TO:\n\nINT. BURTON HOUSE BIG HEAD CLOSEUP OF ANNE\n\nHer face is troubled. CAMERA BEGINS TO PULL BACK. We see\n\nnow that the reception is in progress and that Anne stands\n\nbeside her father to greet the arriving guests. CAMERA PULLS\n\nBACK FURTHER to show us a full view of a very crowded\n\nWashington gathering Many white ties and tails and decollete\n\nin evidence. Many accents. Even some foreign languages are\n\nbeing spoken. Music and chatter in the b.g.\n\nCLOSE SHOT\n\nAnne and the Senator are still greeting new arrivals. Anne's\n\nmanner is somewhat preoccupied. She glances around as she\n\nspeaks, as though looking for someone.\n\n ANNE\n\n (to new arrival)\n\n Thank you so much, Mr. Lindsay.\n\n We'll look forward to it.\n\nPANNING SHOT FROM ANNE'S VIEWPOINT\n\nTHE CAMERA PASSES various groups of guests in conversation\n\nincluding Guy and Barbara who are together. From this\n\ndistance we cannot hear what they are saying. CAMERA\n\nCONTINUES TO the front door. It opens to admit a new arrival.\n\nIt is Bruno. He wears white tie and tails, looking very\n\nelegant.\n\n Converted to PDF by www.screentalk.org 78.\n\nWe see Guy excuse himself from Barbara, cross to Bruno and\n\nspeak to him angrily, obviously asking, "What are you doing\n\nhere?" Bruno, however, greets Guy with a smile then turns\n\nfrom him, unperturbed and bland. He sees Anne and moves\n\ntoward her, smiling.\n\nCLOSEUP ANNE\n\nAs Bruno comes in her direction, Anne's expression shows her\n\nmystification and concern about Bruno's presence and about\n\nGuy's attitude toward him.\n\nMED. SHOT\n\nBruno comes up to Anne and the Senator. He gives a slight\n\nbow to the Senator; then puts his hand out to Anne.\n\n BRUNO\n\n Good evening, Miss Burton.\n\nThe Senator looks inquiringly. Anne makes the introduction.\n\n ANNE\n\n This is Mr. Antony, father.\n\n SENATOR\n\n How do you do, sir.\n\n BRUNO\n\n I'd like to talk to you sometime,\n\n Senator, about my idea of harnessing\n\n the life force. It will make atomic\n\n power look like the horse and buggy.\n\n (the Senator and Anne\n\n are beginning to\n\n look at him in\n\n amazement)\n\n I'm already developing my faculty\n\n for seeing millions of miles. And,\n\n Senator, can you imagine being able\n\n to smell a flower on the planet of\n\n Mars? I'd like to lunch with you\n\n some day soon and tell you more about\n\n it.\n\nInterrupted by new arrivals, Bruno moves away out of the\n\npicture, with a charming smile to Anne.\n\nThe Senator greets the new guests with open mouth and simply\n\nshakes their hands while glancing off in direction of the\n\ndeparting Bruno.\n\n Converted to PDF by www.screentalk.org 79.\n\n DOWAGER\n\n (to Senator)\n\n So nice to see you, my dear Senator.\n\n SENATOR\n\n Ah yes, indeed -- I beg your pardon?\n\nShe realizes he hasn't heard a word she's said and haughtily\n\nmoves on. The Senator turns to Anne.\n\n SENATOR\n\n (still looking after\n\n Bruno)\n\n I don't remember inviting that young\n\n man. Who is he?\n\n ANNE\n\n A friend of the Darvilles.\n\n SENATOR\n\n He has an unusual personality.\n\n Provocative.\n\nCLOSEUP ANNE\n\nShe looks off in Bruno's direction extremely disturbed at\n\nthis new aspect of the mysterious stranger.\n\nCLOSEUP GUY\n\nHe is watching Bruno.\n\nMED. SHOT\n\nGuy sees Bruno join a group of several ladies who are seated\n\non a settee and a couple of older men who are standing by.\n\nA waiter comes along with a tray of drinks. Bruno takes\n\none.\n\nCLOSEUP BARBARA\n\nShe comes from the same direction that Guy came. She stops\n\nshort as she sees:\n\nMED. SHOT FROM BARBARA'S VIEWPOINT\n\nBruno is now heartily joining in conversation with one of\n\nthe elderly gentlemen.\n\n Converted to PDF by www.screentalk.org 80.\n\nCLOSE SHOT BRUNO AND GROUP\n\nBruno talking to an elderly, dignified gentlemen.\n\n BRUNO\n\n But tell me, Judge, after you've\n\n sentenced a man to the chair, isn't\n\n it difficult to go and eat your dinner\n\n after that?\n\n JUDGE\n\n Young man, when a murderer is caught,\n\n he must be tried. When he is\n\n convicted, he must be sentenced.\n\n When he is sentenced to death, he\n\n must be executed.\n\n BRUNO\n\n Quite impersonal, isn't it, sir?\n\n JUDGE\n\n So it is. Besides, it doesn't happen\n\n every day.\n\nAt this moment, Anne comes into the scene. She hesitates as\n\nshe hears Bruno's answer.\n\n BRUNO\n\n So few murderers are caught.\n\nThe Judge moves out of the way. Bruno smiles blandly at the\n\nladies. One of them speaks to him.\n\n MRS. CUNNINGHAM\n\n Well, Mr. Antony, you seem very\n\n interested in the subject of murder.\n\nAnne looks more troubled, then moves on out of the scene.\n\n BRUNO\n\n No more than anyone else. No more\n\n than you, for instance.\n\n MRS. CUNNINGHAM\n\n Me? I'm not interested in murder.\n\nBruno pulls up a chair to face the two woman on the settee,\n\nsits down, straddling the seat, to look at them over the\n\nback of the chair and settle down for a nice conversation.\n\nConverted to PDF by www.screentalk.org 81.\n\n BRUNO\n\n (his tone is teasing)\n\n Oh, come now, everyone's interested\n\n in that. Everyone would like to put\n\n someone out of the way. Now surely,\n\n Madame, you're not going to tell me\n\n that there hasn't been a time when\n\n you wanted to dispose of someone.\n\n Your husband, for instance.\n\n MRS. CUNNINGHAM\n\n (laughs)\n\n Good heavens, no!\n\n BRUNO\n\n (playfully)\n\n Ah ah!\n\n (shaking a finger at\n\n her)\n\n Are you sure? Do you mean to tell\n\n me there wasn't a tiny moment - when\n\n you'd been made really angry? And\n\n what did you say?\n\n MRS. CUNNINGHAM\n\n (squirms, giggling)\n\n Well...\n\n BRUNO\n\n There you are, you see! There you\n\n are! All right, now you're going --\n\n to do a murder. How are you going\n\n to do it? This is the fascinating\n\n part -- how are you going to do it...I\n\n didn't get your name?\n\n MRS. CUNNINGHAM\n\n Mrs. Cunningham.\n\n BRUNO\n\n Mrs. Cunningham, how are you going\n\n to do it?\n\n MRS. CUNNINGHAM\n\n (entering into the\n\n spirit of the play)\n\n Well, I suppose I'll have to get a\n\n gun from somewhere.\n\n BRUNO\n\n (shakes his head)\n\n Tssk, tssk. Oh no, Mrs. Cunningham.\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 82.\n\n BRUNO (CONT'D)\n\n Bang, bang, all over the place.\n\n Blood everywhere?\n\nThe other woman joins in:\n\n MRS. ANDERSON\n\n What about a little poison?\n\n BRUNO\n\n Ah! That's better, that's better.\n\n Mrs.....?\n\n MRS. ANDERSON\n\n Anderson.\n\n BRUNO\n\n (he is thoroughly\n\n enjoying himself)\n\n That's better, Mrs. Anderson. But\n\n Mrs. Cunningham is in a dreadful\n\n hurry. Poison could take...let's\n\n see...ten to twelve weeks, if poor\n\n Mr. Cunningham is to die from natural\n\n causes.\n\n MRS. CUNNINGHAM\n\n I have a wonderful Idea! I can take\n\n him out in the car and when I get to\n\n a lonely spot, knock him on the head\n\n with a hammer, pour gasoline over\n\n him and over the car and start the\n\n whole thing ablaze.\n\n BRUNO\n\n (looks at her\n\n deprecatingly)\n\n And then have to walk all that way\n\n home?\n\nMrs. Anderson laughs.\n\n BRUNO\n\n No, I have the best way, and the\n\n best tools.\n\n (he holds out his\n\n hands and shows them)\n\n Simple, silent, and quick. The silent\n\n part being the most important. Let\n\n me show you what I mean.\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 83.\n\n BRUNO (CONT'D)\n\n (he raises his hands\n\n toward Mrs.\n\n Cunningham's throat,\n\n then stops a moment\n\n to ask)\n\n You don't mind if I borrow your neck\n\n for a moment do you?\n\n MRS. CUNNINGHAM\n\n (giggles)\n\n Well, it's not for long.\n\n BRUNO\n\n Oh! no.\n\n (he takes a drink and\n\n puts his glass down)\n\n Now, when I nod my head, just see if\n\n you can cry out, and I bet you can't.\n\n (he places his hands\n\n around Mrs.\n\n Cunningham's neck)\n\n Now with my two thumbs...you see\n\n that's where I'll be able to prevent\n\n any sound coming from you. Now,\n\n just wait for the nod of my head.\n\nCLOSEUP BRUNO\n\nAs he starts to Press her neck, his eyes wander from the\n\nface of his "victim" to someone else off scene.\n\nMED. SHOT BARBARA\n\nShe is watching this rather unorthodox demonstration. The\n\nCAMERA MOVES UP until her head fills the screen. Her glasses\n\nglint in the light.\n\nCLOSEUP BRUNO\n\nHe is now transfixed. His breathing becomes heavy. A strange\n\nexpression comes over his face. He still stares off at\n\nBarbara.\n\nMED. SHOT BARBARA\n\nWe see the whirling merry-go-round spinning around her head.\n\n Converted to PDF by www.screentalk.org 84.\n\nBIG HEAD CLOSEUP BRUNO\n\nHe now seem to have almost gone into a trance. Over the\n\nshot we begin to HEAR a strangled cry, and a broken\n\nexclamation, then Mrs. Anderson's voice.\n\n MRS. ANDERSON'S VOICE\n\n Mr. Antony! Mr. Antony!\n\n ANOTHER VOICE\n\n Stop him! Stop him!\n\nCLOSEUP\n\nBruno's wrists and hands and the neck of his victim. We can\n\njust see Mrs. Cunningham's chin at the top of the screen.\n\nHer head is tossing from side to side. Her hands are\n\nclutching at Bruno's wrists. The hands of the other two\n\nwomen, also in the picture, are pulling at Bruno's wrists.\n\nMrs. Cunningham's hands begin to slide off. Her head drops\n\nback.\n\nOver this we HEAR cries of:\n\n VOICES\n\n Stop him!\n\n Help, somebody!\n\n Pull him off!\n\n Mr. Antony! Mr. Antony!\n\nCLOSEUP BRUNO\n\nHis body is swaying slightly at the various efforts to drag\n\nhim away from Mrs. Cunningham. His eyes begin to close, and\n\nslowly he falls away from the picture in a dead faint on the\n\nfloor.\n\nMEDIUM SHOT\n\nThere is a rush of people around Mrs. Cunningham, who is\n\nbreathing frantically, her eyes opening and closing. A couple\n\nof women are feebly slapping her hands, someone else is\n\nfanning her face.\n\nMEDIUM SHOT\n\nThe Senator and Guy rush into the picture. They look at the\n\nfallen Bruno. They search around for an explanation. Other\n\nman come in ad they start to pick Bruno up.\n\n Converted to PDF by www.screentalk.org 85.\n\n GUY\n\n Bring him this way.\n\nGuy gives a quick look in direction of Mrs. Cunningham, sees\n\nthat she is being attended to.\n\nMEDIUM SHOT\n\nAnne rushes into the picture. She sees Bruno being helped\n\nto his feet; then turns her attention to Mrs. Cunningham,\n\nwho has now somewhat recovered. Mrs. Cunningham is helped\n\nto the settee. There is a babble of women's voices trying\n\nto explain what has happened.\n\n ANNE\n\n (thru the babble)\n\n Bring her upstairs.\n\nAs the two groups pass off in different directions, the few\n\npeople who ran into the scene late are asking the others\n\nwhat the disturbance is. "What's wrong?" "Did she faint?"\n\n"I didn't see anything." "What happened to him?" "Somebody\n\nhurt?" But one small figure stands in the clear. It is\n\nBarbara, She is still transfixed by what she has seen. Her\n\nhands are trembling. CAMERA MOVES SLOWLY IN ON HER. We see\n\nthat her lips are trembling, too, and in her eyes frightened\n\ntears are welling. Her breath is heavy.\n\nINT. STUDY\n\nBruno is stretched out on a settee. He is completely out.\n\nHis collar and tie are open. Two or three of the male guests\n\nare just leaving the room. The Senator remains behind for a\n\nmoment with Guy.\n\n SENATOR\n\n I thought he was a bit weird when he\n\n arrived. Who is he?\n\n GUY\n\n I hardly know him, sir.\n\n SENATOR\n\n Get him out of here as soon as you\n\n decently can -- will you. This is a\n\n nice item for the gossips. First\n\n thing you know, they'll be talking\n\n about orgies. I'd better get back...\n\n GUY\n\n Yes, sir.\n\n Converted to PDF by www.screentalk.org 86.\n\nThe Senator leaves. Guy stands over Bruno's outstretched\n\nfigure.\n\nMEDIUM SHOT\n\nBruno is now half awake. Almost without seeing Guy, he\n\nstaggers to his feet and begins to make his way to the door.\n\nGuy advances, and with a sharp thrust, pushes Bruno back on\n\nthe settee.\n\nBruno looks and sees Guy clearly for the first time.\n\n BRUNO\n\n What happened? I was on a merry-go-\n\n round somewhere. It made me dizzy.\n\nGuy moves forward, and thrusting his hand in Bruno's open\n\nshirt, pulls him to his feet. Bruno ignores Guy's violence\n\nand remain puzzled.\n\n GUY\n\n (disgusted)\n\n You're a mad, crazy maniac, and you\n\n ought to be locked-up! Now will you\n\n get out of here and let me alone?\n\n BRUNO\n\n But, Guy --\n\nGuy smashes Bruno in the jaw, in utter disgust, and knocks\n\nhim back onto the settee. Bruno looks up from his sprawled\n\nposition, a dull look in his eye.\n\n BRUNO\n\n You shouldn't have done that, Guy.\n\n GUY\n\n (subsiding)\n\n Come on -- pull yourself together.\n\n Do your tie up.\n\nBruno staggers to his feet. He fumbles at his collar. As\n\nhe crosses to him, CAMERA MOVES IN to a CLOSER SHOT.\n\n GUY\n\n Here -- let me.\n\nHe fixes Bruno's shirt and collar together and quickly ties\n\nhis white bow. Bruno stands swaying like a small boy as Guy\n\ndoes this.\n\n Converted to PDF by www.screentalk.org 87.\n\nCAMERA PANS WITH THEM as Guy starts to escort Bruno from the\n\nroom.\n\n GUY\n\n Have you got a car here?\n\n BRUNO\n\n (mumbling)\n\n Driver's outside.\n\nThey pass trough door into the hallway.\n\nINT. HALL MEDIUM SHOT\n\nOne or two of the guests turn their heads as Guy takes Bruno\n\nacross to the front door.\n\nCLOSE SHOT\n\nBarbara appears in the hallway, coming from the crowded\n\nsitting room. She watches the two men go out the front door.\n\nMEDIUM SHOT\n\nBruno and Guy going out the front door. The man-servant\n\ndoes not close it immediately, so we are able to HEAR the\n\ncall for Mr. Antony's car.\n\nCLOSEUP BARBARA\n\nShe turns her head and looks up the stairs. Barbara has not\n\nquite recovered from her ordeal. She hurries forward to\n\ngreet Anne who is hurrying down the stairs.\n\nTWO SHOT\n\nCAMERA PANS DOWN with Anne as she descends the last few steps.\n\nBarbara enters the picture and the two girls meet at the\n\nfoot of the stairs.\n\n ANNE\n\n What's the matter, Barbara? Did you\n\n see it happen? Did you see it --\n\n all?\n\n Converted to PDF by www.screentalk.org 88.\n\nCLOSEUP BARBARA\n\n BARBARA\n\n (still shaken)\n\n He looked at me! His hands were on\n\n her throat, but, he was strangling\n\n me!\n\nCLOSEUP ANNE\n\n ANNE\n\n (aghast)\n\n How do you mean?\n\nTWO SHOT\n\n BARBARA\n\n He was looking at her first. Then\n\n he looked over at me. He went into\n\n a sort of trance\n\n (shudders)\n\n He looked horrible!\n\n (reflectively)\n\n He thought he was murdering me.\n\nCLOSEUP ANNE\n\nShe looks away, with growing consciousness of the situation\n\nTWO SHOT\n\n BARBARA\n\n Anne, why me? Why me? What did I\n\n have to do with it?\n\nAnne is extremely concerned and thoughtful. Suddenly she\n\ngets an idea and with a pat on Barbara's arm, asks hurriedly:\n\n ANNE\n\n Do you know where Guy is?\n\n BARBARA\n\n He went out with that man!\n\nAnne hurries to the front door and passes through.\n\n Converted to PDF by www.screentalk.org 89.\n\nEXT. HOUSE\n\nAnne comes out onto the steps and looks around. She stops\n\nshort as she sees:\n\nLONG SHOT EXT. STREET FROM ANNE'S VIEWPOINT\n\nThere are cars lined up outside on the street. One limousine\n\nis pulling up in the center, two figures at the passenger\n\ndoor. One is climbing in. The other is Guy.\n\nCLOSEUP ANNE\n\nShe calls out urgently:\n\n ANNE\n\n Guy!\n\nCLOSE SHOT\n\nGuy turns and closes the door.\n\nMEDIUM SHOT FROM ANNE'S VIEWPOINT\n\nThe limousine moves off and Guy comes toward her.\n\nMEDIUM SHOT\n\nAnne comes down the steps and intercepts Guy on the sidewalk.\n\nShe leads him along a few paces and then stops and faces\n\nhim.\n\nCLOSE TWO SHOT\n\nAnne nods off in the direction of the departed Bruno and\n\nspeaks in a desperate, low voice.\n\n ANNE\n\n You didn't meet him for the first\n\n time the other day, did you, Guy?\n\nGuy stares at her for a moment.\n\n GUY\n\n You mean when you introduced us at\n\n the club?\n\n Converted to PDF by www.screentalk.org 90.\n\n ANNE\n\n Yes. Did you notice how he stared\n\n at Barbara that day?\n\n GUY\n\n (awkwardly)\n\n Well, I didn't -- particularly --\n\n ANNE\n\n (breaks in)\n\n He stared at her again tonight --\n\n while his hands were around Mrs.\n\n Cunningham's throat.\n\nGuy looks at Anne with an expression of growing fear and\n\nalarm. She goes on inexorably:\n\n ANNE\n\n What did Miriam look like, Guy.\n\n GUY\n\n (awkwardly)\n\n Well, why do you ask me? You've\n\n seen her pictures in the paper.\n\n ANNE\n\n Go on, I want you to tell me.\n\n GUY\n\n (haltingly)\n\n Well, she was dark, not too tall,\n\n rather pretty --\n\n ANNE\n\n What else?\n\n GUY\n\n What else is there?\n\n ANNE\n\n She wore glasses, didn't she?\n\n GUY\n\n Yes.\n\n ANNE\n\n She looked a lot like Barbara, didn't\n\n she?\n\nGuy suddenly begins to realize what Anne is getting at.\n\nAnne lowers her head, deliberately avoids looking at Guy, as\n\nshe asks:\n\nConverted to PDF by www.screentalk.org 91.\n\n ANNE\n\n How did you get him to do it, Guy.\n\n GUY\n\n I get him to do it?\n\n ANNE\n\n He killed Miriam, didn't he? Tell\n\n me, Guy!\n\n GUY\n\n Yes.\n\n (suddenly bursting\n\n out)\n\n He's a maniac. I met him on the\n\n train going to Metcalf. He had a\n\n crazy scheme about exchanging murders.\n\n I do his murder and he do mine.\n\n ANNE\n\n (quietly)\n\n What do you mean -- your murder,\n\n Guy?\n\n GUY\n\n Well, he'd read about me in the paper.\n\n He knew about Miriam -- and about\n\n you. He suggested that if he got\n\n rid of Miriam for me, I should kill\n\n his father.\n\n ANNE\n\n You must have realized he was talking\n\n a lot of nonsense!\n\n GUY\n\n Of course! I didn't give it another\n\n thought. And now a lunatic wants me\n\n to kill his father.\n\n ANNE\n\n (beginning to believe)\n\n It's too fantastic!\n\n GUY\n\n (grimly)\n\n Yes, isn't it?\n\n ANNE\n\n You mean you've known about Miriam\n\n all this time?\n\n Converted to PDF by www.screentalk.org 92.\n\n GUY\n\n Since the first night. He gave me\n\n her glasses.\n\n ANNE\n\n Why didn't you call the police?\n\n GUY\n\n (bitterly)\n\n And have them say what you did --\n\n "Mr. Haines, how did you get him to\n\n do it?" And Bruno would say we'd\n\n planed it together.\n\n ANNE\n\n Oh, Guy -- what can we do?\n\n GUY\n\n I don't know, Anne...I don't know.\n\n ANNE\n\n (With an anxious look\n\n across the street)\n\n Guy, hadn't we better go inside?\n\n Your friend Hennessy's watching us.\n\n (she Shudders)\n\n GUY\n\n (sadly)\n\n You see, Anne, that's why I didn't\n\n want you to know anything about this.\n\n I wanted to protect all of you --\n\n your father, Barbara. And now that\n\n you know, you're acting guilty, too.\n\n ANNE\n\n (desperately)\n\n Oh, if we could only talk to father\n\n or someone about it.\n\n GUY\n\n No, that's no good, Anne. I mustn't\n\n drag anyone else into this mess.\n\n Come on. Let's go in.\n\nThey go toward the house.\n\n CUT TO:\n\n Converted to PDF by www.screentalk.org 93.\n\nTWO SHOT ACROSS THE STREET\n\nAs Hennessy watches Anne and Guy go toward the house, his\n\nrelief, HAMMOND, comes up. Hammond's a zealous, hard-eyed\n\nsleuth.\n\n HENNESSY\n\n (a little glum)\n\n Hello, Hammond.\n\n HAMMOND\n\n You look worried. What's the matter?\n\n HENNESSY\n\n You'd better keep on your toes.\n\n Something funny's going on.\n\n DISSOLVE TO:\n\nINT. GUY'S APARTMENT LATER THAT NIGHT\n\nStill in his dinner clothes, Guy is seated in deep thought\n\nnear the telephone, wrestling with his problem. There is an\n\nopen telephone directory in front of him. He comes to a\n\ndecision, picks up the telephone and dials a number. He\n\nwaits for the answer, then:\n\n GUY\n\n Bruno? Yes, yes, it's Guy...I've\n\n decided to do what you want. I'll\n\n make that little visit to father....\n\n (listens a moment)\n\n Tonight.\n\n (listens another moment)\n\n Yes, I want to get this thing over\n\n with, can you leave the house again,\n\n Bruno?\n\n (pause)\n\n You'd better stay out till daylight.\n\nGuy hangs up, rises and starts to move with purpose for his\n\nnight's activities.\n\n DISSOLVE TO:\n\nINT. GUY'S APARTMENT NIGHT\n\nGuy is sitting at the table. He is dressed differently,\n\nhaving changed from his dinner clothes to a sack suit. There\n\nis only one lamp lighted in the room. Guy presents a grim\n\npicture.\n\n Converted to PDF by www.screentalk.org 94.\n\nHe is studying the plan of Bruno's house, and he picks up\n\nthe key Bruno sent along with it. Finally he looks at his\n\nwatch, then folds the plan and puts it in his pocket with\n\nthe key. He rises, crosses to the chest of drawers, opens\n\nthe top drawer.\n\nINSERT: THE OPEN DRAWER\n\nGuy's hands take out the Luger. His hand then picks up\n\nMiriam's glasses from the drawer, holds them a moment. He\n\nis about to put them back, then decides to take them along,\n\nputs them into his pocket.\n\nMED. SHOT\n\nCAMERA PANS GUY across to the window. He parts the curtains\n\nslightly and looks out.\n\nMED. SHOT ON STREET (FROM GUY'S VIEWPOINT)\n\nHammond is lighting a cigarette as he strolls in front of\n\nthe house.\n\nINT. GUY'S APARTMENT\n\nGuy crosses to his door, which he opens surreptitiously.\n\nMED. SHOT CORRIDOR\n\nGuy glances down the stairs, then closes the door behind him\n\nquietly and moves away to a window at the turn of the stairs.\n\nEXT. FIRE ESCAPE\n\nGuy comes out of the window onto the second floor fire escape.\n\nHe creeps stealthily down and emerges into a narrow alleyway.\n\nHe steps back into the shadows for a moment when he sees:\n\nLONG SHOT FROM GUY'S VIEWPOINT (PROCESS)\n\nThe strolling figure of Hammond on the far side of the street.\n\n Converted to PDF by www.screentalk.org 95.\n\nMED. SHOT\n\nGuy turns away and is soon lost in the darkness of the street.\n\n LAP DISSOLVE TO:\n\nEXT. A TALL PAIR OF ELABORATE IRON GATES NIGHT\n\nWe are on the inside of the gates. We see them swing open\n\nslightly and the figure of Guy edges through them.\n\nCLOSE SHOT\n\nGuy leaves the gates ajar and then, taking the plan of Bruno's\n\nhouse from his pocket, and the key, he looks toward the house.\n\nEXT. STEPS LONG SHOT NIGHT\n\nThis is a long flight of steps. Moonlit. They are lined\n\nwith tall black cypress trees which throw their shadows across\n\nthe steps. Guy moves out of one shadow, into another and\n\ncarefully starts up the stairs.\n\nAT THE DOOR\n\nHe pauses, looks about for a moment and listens. Then he\n\nputs the key into the lock, finding it with his flashlight.\n\nThe door opens a few inches. He turns off the flash, and\n\nenters.\n\nINT. ANTONY HOME ENTRANCE HALL\n\nAs Guy moves in soundlessly and closes the door. He looks\n\ntoward the stairs which are in shadow.\n\nMED. SHOT\n\nGuy starts up the stairs slowly. He carries his flashlight\n\nand the plan.\n\nAT THE TOP OF THE STAIRS THE DOG\n\nA huge shadow lies it the head of the stairs. As Guy comes\n\nslowly up the stairs, the Great Dane looks down at him.\n\n Converted to PDF by www.screentalk.org 96.\n\nGUY ON THE STAIRS\n\nHe reacts to the sight of the dog, stops an instant, and\n\nturn on his flashlight. The heavy massive face of the dog\n\nlooks straight down at him. Guy turns off the flashlight\n\nand after a moment of indecision starts slowly up the stairs\n\nonce more, the dog watching every step he takes.\n\nUPPER HALLWAY\n\nGuy comes up the last few stairs and still the dog hasn't\n\nmoved. Guy slowly edges past him and the Great Dane's head\n\nturns to watch him.\n\nGUY\n\nmoving quietly along the hallway, approaches two doors. He\n\ntakes out his flash and identifies the door with his plan.\n\nINSERT:\n\nThe plan shows two doors in relation to the stairway. The\n\nfirst one is clearly marked: "MY room." The adjoining door\n\nis marked: "My FATHER'S room."\n\nCLOSE SHOT GUY\n\nHe pauses at the first door, then passes it quietly, walking\n\non to the next one. He turns the knob soundlessly and passes\n\nthrough into the room.\n\nINT. ANTONY BEDROOM LONG SHOT\n\nThe room is in darkness except for the dim outline of the\n\nrecumbent figure in the bed. We hear Guy's voice, in a loud\n\nwhisper:\n\n GUY\n\n Mr. Antony!\n\nThe figure stirs.\n\nANOTHER ANGLE\n\nGuy takes a stop closer to the bed.\n\n Converted to PDF by www.screentalk.org 97.\n\n GUY\n\n (urgently)\n\n Mr. Antony! Don't be alarmed -- but\n\n I must talk to you about your son.\n\n About Bruno. Mr. Antony!\n\nThe figure on the bed turns and a hand stretches out toward\n\na bedside light. The light goes on with a sudden glare.\n\nCLOSEUP FACE OF BRUNO IN THE LIGHT (LOW CAMERA)\n\nThe low CAMERA throws a vast shadow up on the wall behind\n\nhim, creating a grimace of his smile.\n\n BRUNO\n\n Yes, Mr. Haines?\n\nCLOSEUP GUY\n\nHis face is dead.\n\nMED. SHOT\n\nBruno rises from the bed and sits on the and of it. He is\n\nfully dressed, just as he was at the party, in white tie and\n\ntails.\n\n BRUNO\n\n (politely)\n\n My father isn't home tonight, Mr.\n\n Haines.\n\n (smiles grimly at\n\n Guy's surprise)\n\n I was about to tell you that over\n\n the phone. But you came to such a\n\n sudden decision. I wondered why.\n\n GUY\n\n (recovering quickly)\n\n Since you sent me a key to your house,\n\n I decided to use it -- to make a\n\n little social call on your father.\n\n I thought he'd be Interested to know\n\n he his a lunatic son.\n\nThe faintest flicker of Bruno's eyes indicates the intensity\n\nof his reaction. He stares hard at Guy.\n\n Converted to PDF by www.screentalk.org 98.\n\n BRUNO\n\n Then a I correct, Mr. Haines, in\n\n assuming that you have no intention\n\n of going ahead with our arrangement?\n\n GUY\n\n No intention whatsoever. I never\n\n had.\n\n BRUNO\n\n I see. You won't have any further\n\n use for the key, then, Mr. Haines.\n\n (he holds out his\n\n hand and Guy gives\n\n him the key)\n\n Thank you very such.\n\nAs Bruno continues to stare at him, Guy takes out the Luger.\n\nFor a moment a look of fear comes into Bruno's face as he\n\nthinks Guy will probably shoot him. After a pause, Guy tosses\n\nthe gun on the bed.\n\n GUY\n\n Or this.\n\nBruno's relief turns again to menace. He picks up the gun\n\nand fingers it nervously.\n\n GUY\n\n (kindly)\n\n Look, Bruno. You're terribly sick.\n\n (haltingly)\n\n I don't know whether it's possible\n\n for you to realize it or not. I\n\n don't know much about these things,\n\n Bruno. But why don't you go someplace\n\n where you can get some treatment?\n\n Not only for your own sake, Bruno,\n\n but you can't go on causing more and\n\n more destruction to anyone you happen\n\n to meet.\n\nBruno pays no attention. He rises.\n\nTWO SHOT\n\nGuy's arguments have made no impression on Bruno whatsoever.\n\nHe fingers the gun.\n\n BRUNO\n\n I don't like to be doublecrossed.\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 99.\n\n BRUNO (CONT'D)\n\n I have a murder on my conscience,\n\n but it's not my murder, Mr. Haines --\n\n it's yours. And as you're the one\n\n to profit, I think you should be the\n\n one to pay for it.\n\nFor an instant his nervous hands seem to be struggling with\n\nthe urge to kill Guy.\n\n GUY\n\n (gives up)\n\n Well, I guess it's no use, Bruno.\n\n We sees to have nothing further to\n\n discuss.\n\nBruno goes to the door in silent acquiescence and opens it\n\nfor Guy to pass through.\n\nINT. HALLWAY MED. SHOT\n\nGuy walks toward the stairs, tense and apprehensive. Bruno\n\nis following him, still holding the gun. When the Great\n\nDane sees Bruno it gets to its feet, as if waiting for a\n\ncommand.\n\nGuy starts down the stairs but Bruno stays where he is, the\n\ndog beside him. Gay turns and looks back it this tableaux\n\nof menace.\n\n BRUNO\n\n Don't worry. I'm not going to shoot\n\n you, Mr. Haines. It might disturb\n\n mother.\n\n (with a feeling of\n\n power)\n\n I'm a very clever follow. I'll think\n\n of something better than that. Much\n\n better.\n\nLONG SHOT\n\nBruno remains in the foreground of the scene as Guy proceeds\n\non down the stairs. We see him open the front door and pass\n\nthrough.\n\n LAP DISSOLVE TO:\n\n Converted to PDF by www.screentalk.org 100.\n\nEXT. STREET ACROSS GUY'S APARTMENT EARLY\n\nCLOSE SHOT HENNESSY AND HAMMOND MORNING\n\nHennessy is relieving Hammond who has kept watch on Guy's\n\napartment night.\n\n HAMMOND\n\n (in the middle of his\n\n story)\n\n He came back at three twenty-five.\n\n I didn't even know he'd given me the\n\n slip until his 'phone kept ringing\n\n for about half an hour. Nobody sleeps\n\n that sound. So I got the janitor to\n\n let me in. No Haines.\n\n HENNESSY\n\n (to himself)\n\n Wonder where he went?\n\n HAMMOND\n\n We'll probably hear of another dame\n\n murdered.\n\n HENNESSY\n\n (puzzled)\n\n Shut up. I'd better contact Metcalf.\n\n I should think this calls for more\n\n questioning of Mr. Haines.\n\n HAMMOND\n\n Questioning? Nuts! Let's take him\n\n in.\n\n HENNESSY\n\n My dear Mr. Hamond, how many times\n\n do I have to tell you that we have\n\n nothing conclusive on Haines? There's\n\n no evidence that he was ever at the\n\n scene of the crime. Can't you get\n\n that into your thick head?\n\n (quietly)\n\n Now stay put till I get back.\n\nAs he starts away --\n\n FADE OUT.\n\n Converted to PDF by www.screentalk.org 101.\n\nFADE IN\n\nINT. ANTONY LIVING ROOM LATE MORNING\n\nAnne and Mrs. Antony are in the middle of a conversation.\n\nAnne's manner is tense and purposeful, Mrs. Antony's much\n\nless serious.\n\n MRS. ANTONY\n\n Oh, now, Miss Burton, really! I\n\n know Bruno's been in some very awkward\n\n scrapes, but nothing so ridiculous\n\n as a murder.\n\n (she gives a short\n\n little laugh)\n\n ANNE\n\n (desperately)\n\n But, Mrs. Antony, you've got to make\n\n him do something about this. Don't\n\n you see that just one word from him\n\n would extricate Guy from this dreadful\n\n situation?\n\n MRS. ANTONY\n\n (lightly)\n\n Oh, but Miss Burton, I'm sure this\n\n thing must be some practical joke.\n\n You know, Bruno sometimes goes too\n\n far.\n\n (girl to girl)\n\n Of course I shouldn't be saying this\n\n to an outsider, but sometimes he's\n\n terribly irresponsible and gets into\n\n all kinds of escapades.\n\n ANNE\n\n But don't you understand, Mrs. Antony --\n\n your son is responsible for a woman's\n\n death.\n\n MRS. ANTONY\n\n (drawing herself up\n\n with some hauteur)\n\n Did Bruno tell you this?\n\n ANNE\n\n Of course not, Mrs. Antony.\n\n MRS. ANTONY\n\n (that settles it)\n\n Well, there you are.\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 102.\n\n MRS. ANTONY (CONT'D)\n\n (she sighs and rises,\n\n winding it up)\n\n Well, Miss Burton, it was very nice\n\n of you to call. You must excuse me\n\n now. I must get back to my painting.\n\n Do you care for painting, Miss Burton?\n\n I find it so soothing.\n\n (shakes Anne's hand)\n\n You must come again sometime.\n\nShe goes out. Anne is left helpless, standing in the middle\n\nof the room. She picks up her purse and is about to go when\n\nshe hears a voice:\n\n BRUNO'S VOICE\n\n Oh, Miss Burton!\n\nAnne turns back in direction of the voice. CAMERA PULLS\n\nBACK until we can see the feet of Bruno protruding from behind\n\na chair in which he is sitting. He has obviously heard the\n\nentire conversation between Anne and his mother. Bruno rises.\n\nHe is in dressing gown and pajamas.\n\n BRUNO\n\n I'm afraid mother wasn't very helpful,\n\n was she?\n\n (he strolls toward\n\n Anne)\n\n You know she hasn't been well for a\n\n long time. She's a little -- how\n\n shall I say -- confused.\n\n (shakes his head\n\n commiseratingly)\n\n Poor mother.\n\nAnne is too stunned to speak.\n\n BRUNO\n\n You know, I'm very upset with Guy.\n\n He shouldn't have sent you on an\n\n errand like this.\n\n ANNE\n\n Guy doesn't know I'm here, Mr. Antony.\n\n BRUNO\n\n He's been leading you up the garden\n\n path, I'm afraid. He must be very\n\n desperate to try to involve me.\n\n I've been protecting him ever since\n\n we had that conversation on the train\n\n and he told me how he hated his wife.\n\n Converted to PDF by www.screentalk.org 103.\n\nBruno is now standing near the window a little apart from\n\nAnne, with his back to him. He takes something out of the\n\npocket of his dressing gown and looks down at it in his hand.\n\nIt is Guy's lighter. Suddenly he stuffs it back his pocket\n\nand turn back to Anne.\n\n BRUNO\n\n Why, do you know, Miss Burton, he\n\n tried to get me to go back to the\n\n island one night after dark and pick\n\n up his lighter so the police wouldn't\n\n find it? He dropped it there, you\n\n know, when -- well, that night.\n\nAnne's horror is growing.\n\n BRUNO\n\n The whole thing's been worrying me\n\n so much. But of course I couldn't\n\n do it, Miss Burton. It would have\n\n been too risky. And besides, it\n\n would have made me an accessory.\n\nAnne stares at this insane man and sinks on the settee. She\n\nstarts to cry in sheer frustration. Bruno goes to her\n\nsympathetically.\n\n BRUNO\n\n Miss Burton, I know how you feel.\n\nHe puts his hand on her shoulder. Anne flings it off. There\n\nis an awkward pause as Bruno looks down at her. Then he\n\nbegins to look around restlessly.\n\n BRUNO\n\n Miss Burton, you must excuse me. I\n\n have an urgent appointment.\n\n (looks it his watch)\n\n I must go up and change. Now, I\n\n really must go...if you'll excuse\n\n me...\n\nHe turns, starts out of the room and up the stairs in the\n\nhall. Anne watches him.\n\nSTAIRWAY FROM ANNE'S VIEWPOINT\n\nBruno turns and waves to Anne from the landing, then goes on\n\nup the stairs.\n\n Converted to PDF by www.screentalk.org 104.\n\nINT. LIVING ROOM MED. SHOT\n\nAnne slowly rises, a lonely figure in the large room, and\n\nmakes her way out.\n\n DISSOLVE TO:\n\nLONG SHOT FOREST HILL STADIUM\n\nGrouped. A game is in progress.\n\nMED. SHOT A TERRACE NEAR THE MAIN STADIUM (PROCESS)\n\nwhere people get refreshments. There are various table with\n\numbrellas.\n\nMED. SHOT AT ONE OF TABLE (PROCESS)\n\nAnne and Guy are seated at the table.\n\n ANNE\n\n ...And he said that if the police\n\n found your lighter there, that's all\n\n they'd need -- something to prove\n\n you were at the scene of the murder.\n\n GUY\n\n (grimly)\n\n That big lie about my wanting him to\n\n get it back means he's going to put\n\n my lighter on that island!\n\n ANNE\n\n (urgently)\n\n Guy, you'll have to get there before\n\n he does. You won't have time to\n\n play, You'd better tell them.\n\n (she nods her head in\n\n the direction of the\n\n center court)\n\n GUY\n\n Darling, if that loudspeaker announces\n\n that I'm not going to play, Hennessy\n\n bound to be suspicious He'd keep me\n\n from ever getting near Metcalf.\n\n ANNE\n\n Then I'll go.\n\n Converted to PDF by www.screentalk.org 105.\n\n GUY\n\n (quickly)\n\n No, darling.\n\n (he puts his hand on\n\n hers and speaks\n\n firmly, with concern\n\n for her safety as\n\n well as for his own\n\n situation)\n\n You stay right here and help me give\n\n Hennessy the slip after the match.\n\n ANNE\n\n But, Guy, that'll be too late!\n\n GUY\n\n (getting a thought)\n\n Didn't Bruno say that I wanted him\n\n together there one night after dark?\n\n ANNE\n\n Yes.\n\n GUY\n\n Well, that's what's in his mind now.\n\n He's not going to expose himself in\n\n broad daylight, If I can finish off\n\n this match in three sets, I'll still\n\n get there in time.\n\nREYNOLDS, Guy's opponent, enters scene behind Guy's chair.\n\n REYNOLDS\n\n We're on in a few minutes, Guy.\n\n (to Anne)\n\n How are you, Miss Morton.\n\nAnne acknowledges his greeting with a nod.\n\n GUY\n\n Okay, Tim. Be right with you.\n\nReynolds leaves Anne and Guy rise, and as they walk toward\n\nthe stadium, we can see Guy start to speak to Anne in a\n\nwhisper.\n\nENTRANCE TO COVERED STAND ALREADY SHOT\n\nHennessy and Hammond are standing by.\n\n Converted to PDF by www.screentalk.org 106.\n\n HAMMOND\n\n Well, if Turley said to pick him up\n\n for questioning, let's pick him up!\n\n HENNESSY\n\n Let him have his game first, Hammond.\n\n HAMMOND\n\n (sourly)\n\n This is the first time I ever waited\n\n for a murder suspect to play tennis\n\n before I pulled him in. When the\n\n boys it headquarters heir about this\n\n they'll send me orchids.\n\nGuy and Anne come into the scene just as the players from\n\nthe previous match emerge. They pass through, nodding to\n\nHennessy.\n\n HENNESSY\n\n Good luck, Guy.\n\nGuy is so preoccupied with his grim doesn't nod to Hennessy\n\nuntil Anne nudges him.\n\nINSIDE THE STAND MED. SHOT\n\nAnne is reluctant to leave Guy who must now join his opponent,\n\nReynolds.\n\n GUY\n\n You got it straight?\n\n (ANNE nods)\n\n Just make sure Barbara has everything\n\n ready as soon as the third set starts.\n\nHe goes on to the court, and Anna goes to her box.\n\nMED. SHOT\n\nAnne joins Barbara in the box. She starts to whisper\n\nsomething to her.\n\nLONG SHOT\n\nGuy and Reynolds complete their warm-up as the umpire\n\nannounces that Guy is to serve. The game starts.\n\n Converted to PDF by www.screentalk.org 107.\n\nEXT. ANTONY HOME\n\nA taxi is at the front door. Bruno is descending the steps.\n\nHe gets into the cab, which moves off.\n\nFOREST HILLS MED. SHOT ANNOUNCER'S BOOTH (PROCESS)\n\nOver the shoulder of the announcer WE SEE the game in progress\n\nthrough the window of his booth.\n\n ANNOUNCER\n\n --It looks like an interesting match\n\n with Haines constantly charging the\n\n net -- not like Haines at all -- to\n\n press so early in the game...\n\nMED. SHOT TEN COURT\n\nGuy and his opponent, Reynolds, in play. Guy scores a point.\n\nCLOSEUP THE UMPIRE\n\nHe announces game to Haines.\n\nMED. LONG SHOT\n\nWe see the two men change ends and come toward the Umpiri's\n\nchair. Reynolds stops to take a drink of water. Guy, with\n\nan impatient glance at him, moves over to the passing line\n\nand waits, the CAMERA going with him.\n\nEXT. WASHINGTON STREET\n\nA taxicab is seen coming along.\n\nMED. SHOT INSIDE CAB (PROCESS)\n\nBruno is sitting with in unlighted cigarette in his mouth.\n\nCAMERA MOVES IN until he is in big CLOSEUP. His eyes look\n\ndown. There is the SOUND of a click, then, Guy's lighter\n\ncomes up into the picture held against the cigarette.\n\n LAP DISSOLVE TO:\n\n Converted to PDF by www.screentalk.org 108.\n\nFADE IN\n\nLONG SHOT FOREST HILLS STADIUM\n\nGrouped. A game is in progress.\n\nMED. SHOT\n\nA terrace where people get refreshments. There are various\n\ntables with umbrellas.\n\nMED. SHOT AT ONE OF TABLE (PROCESS)\n\nGuy is seated. He has his rackets with him and is waiting\n\nhis turn to start his match. An official is talking to him\n\nbut Guy keeps looking around as if expecting someone.\n\n OFFICIAL\n\n Well, at least there'd be a trip to\n\n Australia, if you made it.\n\n GUY\n\n (absently)\n\n We'll know more about that by the\n\n end of the week...\n\n (his face brightens\n\n as he sees Anne)\n\nAnne hurries in, nods briefly to the official who has started\n\nto leave, and sits down.\n\n OFFICIAL\n\n They're close to the finish, Guy\n\n GUY\n\n Be right there.\n\n (turns to Anne)\n\n I was afraid you wouldn't get here.\n\n Wish me luck, darling.\n\nHe makes a move as if to follow official toward the stadium,\n\nbut Anne puts hand on his arm.\n\n ANNE\n\n (quickly and urgently)\n\n Guy, listen to me, If I sound all\n\n mixed up I can't help it. I -- I'm\n\n scared.\n\n GUY\n\n What about?\n\n Converted to PDF by www.screentalk.org 109.\n\n ANNE\n\n That's just it. I don't know. It's\n\n Bruno. I talked to him, Guy --\n\nGuy stares at her, takes a quick look toward the stadium,\n\nthen gives Anne his full attention.\n\n ANNE\n\n He acted peculiar -- as if he could\n\n put the murder right in your lap,\n\n and not involve himself at all.\n\n GUY\n\n (shaking his head)\n\n He'd drag himself into it, -- and\n\n Bruno loves Bruno. I'm all right so\n\n long as he thinks I have an alibi\n\n for that night.\n\n (noticing the stricken\n\n look on Anne is face)\n\n He knows?\n\nAnne nods slowly.\n\n GUY\n\n (grimly)\n\n Then he'll think of something. He\n\n said he would.\n\n ANNE\n\n Guy, has he anything that the police\n\n could trace to you --\n\n (quoting Bruno)\n\n Any little thing.\n\n GUY\n\n My cigarette lighter. He said once\n\n he could have left it on the islands\n\n as evidence\n\n (a pause)\n\n But he wouldn't do that. Not in\n\n broad day light.\n\n ANNE\n\n (trying to think)\n\n But he's going somewhere, Guy. He\n\n told his mother --\n\n GUY\n\n (tensely)\n\n Metcalf -- did he say Metcalf?\n\n Converted to PDF by www.screentalk.org 110.\n\n ANNE\n\n No, -- I don't think so. Oh, why\n\n can't I remember -- he said such\n\n crazy things!\n\n GUY\n\n (tensely)\n\n Try to think, Anne!\n\n VOICE\n\n (OFFSCENE)\n\n Guy Haines! -- Reynolds!\n\nWhile Anne is frantically trying to remember, Guy turns\n\ntoward, the stadium and gives a signal of "Be right there."\n\n ANNE\n\n Something about the moon -- he said\n\n he had an appointment with the moon.\n\nGuy's shoulders droop with disappointment.\n\n GUY\n\n That's no help. But I can't take\n\n any chances. I've got to get that\n\n lighter -- somehow.\n\nREYNOLDS, Guy's opponent, ENTERS SCENE behind Guy's chair.\n\n REYNOLDS\n\n Okay, Guy. We're on.\n\nHe walks away. Anne and Guy rise, following him.\n\n GUY\n\n I'll have to default.\n\n ANNE\n\n And have Hennessy and that other one\n\n right at your heels?\n\nGuy's expression says she's right, as they walk toward the\n\nstadium.\n\nENTRANCE TO COVERED STAND\n\nHennessy and Hammond, the two detectives, are standing by.\n\n HAMMOND\n\n First time I ever waited for a killer\n\n to play tennis before I nabbed him!\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 111.\n\n HAMMOND (CONT'D)\n\n When the boys at headquarters hear\n\n about this they'll send me an orchid!\n\n HENNESSY\n\n We got our orders. We take him in --\n\n after the match.\n\nGuy and Anne come INTO THE SCENE just as the players from\n\nthe previous match emerge. They pass through, nodding to\n\nHennessy.\n\n HENNESSY\n\n (a little sadly)\n\n Good luck, Guy!\n\nGuy gives him a thank-you nod. Hammond rolls his eyes in\n\ndisgust at Hennessy's politeness.\n\nINSIDE THE STAND MED. SHOT\n\nAnne is about to turn to her box but she is reluctant to\n\nleave Guy, who must now join his opponent, Reynolds. As\n\ntheir eyes hold, in mutual helplessness, Guy suddenly stares\n\nat her with realization.\n\n GUY\n\n The moon! You said he had an\n\n appointment --\n\nAnne looks puzzled as Guy looks up at the sun, then at his\n\nwatch.\n\n GUY\n\n Then he is going to Metcalf. But he\n\n has to wait until it gets dark --\n\n (with frantic haste,\n\n he thinks quickly,\n\n then murmurs to Anne)\n\n Listen, Anne, as soon as the third\n\n set starts, tell Barbara --\n\nMED. CLOSE SHOT REYNOLDS\n\nwaiting at the bottom of steps to the stand. Guy joins his\n\nopponent, and Anne goes to her box. Guy and Reynolds move\n\nonto the court amid the rounds of applause that greet them.\n\n Converted to PDF by www.screentalk.org 112.\n\nMEDIUM SHOT ANNE JOINS BARBARA\n\nIn the box. She starts to whisper something to her.\n\nLONG SHOT\n\nGuy and Reynolds complete their warm-up as the umpire\n\nannounces that Guy is to serve. The game starts.\n\nEXT. ANTONY HOME\n\nA taxi is at the front door. Bruno is descending the steps.\n\nHe gets into the cab, which moves off.\n\nFOREST HILLS MED. SHOT ANNOUNCER'S BOOTH\n\nOver the shoulder of the announcer WE SEE the game in progress\n\nthrough the window of his booth.\n\n ANNOUNCER\n\n It looks like an interesting match --\n\n with Haines out to blast Reynolds\n\n into a fast fight, -- not like Haines\n\n at all -- to press so early in the\n\n game...\n\nMED. SHOT THE COURT\n\nGuy and his opponent, Reynolds, in play. Guy scores a point.\n\nCLOSEUP THE UMPIRE\n\nHe announces game to Haines.\n\nMED. LONG SHOT\n\nWe see the two men change ends and come toward the Umpire's\n\nchair. Reynolds stops to take a drink of Water. Guy, with\n\nan impatient glance it him, moves over to the passing line\n\nand waits, the CAMERA going with him.\n\nEXT. WASHINGTON STREET\n\nA taxicab is seen coming along.\n\n Converted to PDF by www.screentalk.org 113.\n\nMED. SHOT INSIDE TAXI CAB\n\nBruno is sitting with an unlighted cigarette in his mouth.\n\nCAMERA MOVES IN until he is in big CLOSEUP. His eyes look\n\ndown. There is the SOUND of a click, then Guy's lighter\n\ncomes up into the picture held against the cigarette.\n\n LAP DISSOLVE TO:\n\nINT. ANNOUNCER'S BOOTH FOREST HILL\n\nThe announcer is broadcasting the progress of the match and\n\nwe learn from him that this first set is nearly finished.\n\nLONG SHOT THE COURT\n\nGuy and Reynolds in play.\n\nMED. SHOT\n\nAnne and Barbara sitting in their box watching the play\n\nanxiously.\n\nMED. SHOT\n\nAt the entrance to the covered stand. The two detectives\n\nHennessy and Hammond, are watching. Hammond is bored by\n\nthis game.\n\n HAMMOND\n\n Stupid game. You'd never get me\n\n into them short pants. I'd feel\n\n naked.\n\n HENNESSY\n\n (his eyes intent on\n\n the game)\n\n You'd feel naked in an Eskimo suit --\n\n if you weren't wearing your badge.\n\nMED. SHOT\n\nGuy playing hard but holding his own.\n\nMED. SHOT\n\nReynolds, his opponent, playing back.\n\n Converted to PDF by www.screentalk.org 114.\n\nLONG SHOT\n\nThe big crowd watching.\n\nMED. SHOT\n\nGuy scores point over Reynolds.\n\nMED. SHOT\n\nThere is general applause from the crowd in the covered stand\n\nas we HEAR the Umpire's announcement.\n\n UMPIRE'S VOICE\n\n (O.S.)\n\n Mr. Haines wins the first set.\n\nEXT. UNION STATION WASHINGTON D.C.\n\nWe see Bruno get out of a cab and pass into the depot.\n\nLONG SHOT FOREST HILLS\n\nThe game in process.\n\nMED. SHOT\n\nA nearer view of the game.\n\nCLOSE SHOT GUY IN PLAY\n\nvolleying with Reynolds.\n\nCLOSE SHOT\n\nReynolds playing the covered stand people are concentrating.\n\nMED. SHOT\n\nGuy misses a point and the game. He and Reynolds make for\n\nthe Umpire's chair. We HEAR the Umpire announce.\n\n UMPIRE'S VOICE\n\n Game to Mr. Reynolds. Games are two\n\n all...Second set.\n\n Converted to PDF by www.screentalk.org 115.\n\nINT. UNION STATION WASHINGTON, D.C.\n\nBruno is casually waiting for the train. He stands near a\n\nnews-stand reading a paper.\n\nINSERT:\n\nWe see that the paper is open at the sports page. There is\n\na picture of Guy among other tennis players. WITH A DISSOLVE\n\nthe whole character of this page changes with the exception\n\nof Guy's picture, which becomes surrounded with large type,\n\nannouncing the arrest of Guy Haines for the murder of his\n\nwife Miriam. A sub-heading tells of Guy's cigarette lighter\n\nfound at the scene of the crime. All this DISSOLVES AWAY\n\nand the page becomes once more the sports section.\n\nCLOSEUP\n\nBruno looks up with satisfaction.\n\nLONG SHOT FOREST HILLS\n\nThe crowd watching.\n\nMED. SHOT\n\nGuy and Reynolds in play.\n\nMED. SHOT\n\nGuy playing hard.\n\nMED. SHOT\n\nReynolds playing back.\n\nCLOSEUP\n\nThe Umpire watching the game. Suddenly he announces:\n\n UMPIRE\n\n Game to Mr. Reynolds. Games are\n\n three all... second set.\n\n Converted to PDF by www.screentalk.org 116.\n\nINT. CLUB CAR ON TRAIN\n\nBruno is now seated in his accustomed place in the club car.\n\nHis gloved fingers are quietly toying with Guy's lighter. A\n\npassenger next to him asks:\n\n PASSENGER\n\n May I have a light, please?\n\nBruno looks at him for a moment and then at the lighter.\n\nWith great deliberation he puts the lighter away in his pocket\n\nand takes out book-matches. Lighting a match, he holds it\n\nto his fellow passenger's cigarette.\n\nLONG SHOT FOREST HILLS\n\nThe game as seen from under the covered stand.\n\nMED. SHOT\n\nAnne and Barbara very tense.\n\nCLOSEUP GUY\n\nabout to serve, looks anxiously across the court.\n\nCLOSEUP THE CLOCK\n\nCLOSEUP GUY\n\nas he serves.\n\nCLOSEUP REYNOLDS\n\nreturns.\n\nCLOSEUP BALL\n\nhits the net.\n\nCLOSEUP UMPIRE\n\nannounces.\n\n Converted to PDF by www.screentalk.org 117.\n\n UMPIRE\n\n Second set to Haines. Haines leads\n\n two sets to love.\n\nThere is a round of applause. We see the heads of the two\n\nplayers reach the Umpire's chair. Guy, very anxious still,\n\nas he wipes his neck with a towel.\n\nINT. COVERED STAND CLOSE SHOT ANNE BARBARA\n\nAnne is speaking.\n\n ANNE\n\n If he wins this next set -- you'd\n\n better have everything ready.\n\n (takes bill from her\n\n purse and hands it\n\n to Barbara)\n\n Here -- give the driver this ten\n\n dollars.\n\n BARBARA\n\n (puzzled)\n\n I wish understood what this is all\n\n about!\n\n ANNE\n\n (urgently)\n\n You don't have to understand, just\n\n do it. And for heaven's sake, act\n\n natural.\n\nBarbara nods and goes along.\n\nENTRANCE TO COVERED STAND\n\nBarbara smiles winningly at Hennessy as she goes through.\n\nHer interpretation of "acting natural" is exaggerated and\n\nrather comical. Hammond's eyes narrow as he looks after her\n\nsuspiciously.\n\nLONG SHOT\n\nThe game in progress. Guy starts the next set. He serves.\n\nMED. SHOT\n\nReynolds returns.\n\n Converted to PDF by www.screentalk.org 118.\n\nMED. SHOT\n\nGuy volleys.\n\nMED. SHOT\n\nReynolds puts the ball in the air.\n\nCLOSE SHOT\n\nGuy smashes.\n\nCLOSE SHOT\n\nThe ball hits the net.\n\nCLOSEUP UMPIRE\n\n UMPIRE\n\n Love fifteen.\n\nLONG SHOT THE CROWD\n\nWe HEAR the smash of the ball and the voice of the Umpire.\n\n UMPIRE'S VOICE\n\n (O.S.)\n\n Love thirty.\n\nCLOSEUP ANNE\n\nlooking very worried. Again the call of the Umpire.\n\n UMPIRE'S VOICE\n\n (O.S.)\n\n Double fault. Love forty.\n\nINT. THE ANNOUNCER'S BOOTH\n\nThe announcer telling his listeners that Guy Haines seems to\n\nbe a little reckless.\n\n ANNOUNCER\n\n -- Haines hasn't let up his terrific\n\n pace for an instant, smashing every\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 119.\n\n ANNOUNCER (CONT'D)\n\n ball with a recklessness we've never\n\n seen in his playing. It's beginning\n\n to look as if he doesn't care whether\n\n he wins or loses because he's in a\n\n hurry - an awfully big hurry ---\n\n LAP DISSOLVE TO:\n\nEXT. METCALF STATION\n\nWe see Bruno alight from the train. He makes his way in the\n\ndirection of the town.\n\nMED. SHOT METCALF STATION\n\nAs Bruno comes toward us, he stands on the sidewalk and then\n\ntakes the lighter from his pocket once more. At this moment\n\na hurrying passenger on his way to the depot accidentally\n\njogs Bruno's elbow. The lighter flies from his hand.\n\nCLOSE SHOT\n\nWe see it fall through the bars of a grating by the sidewalk.\n\nCLOSEUP BRUNO\n\nlooks down in dismay.\n\nFOREST HILLS MED. SHOT\n\nThe game in progress. Guy and his opponent playing hard.\n\nGuy misses a point. We HEAR the Umpire's call.\n\n UMPIRE'S VOICE\n\n (O.S)\n\n Game to Mr. Reynolds. Mr. Reynolds\n\n leads five games to three in the\n\n third set.\n\nEXT. METCALF STATION\n\nBruno is leading a porter toward the grating, pulling him by\n\nthe arm. They reach the drain.\n\n Converted to PDF by www.screentalk.org 120.\n\n BRUNO\n\n Down there -- my -- my cigarette --\n\n (catches himself --\n\n not wanting to say\n\n "cigarette lighter")\n\n case. It's very valuable.\n\n PORTER\n\n (peering down)\n\n Down here?\n\n BRUNO\n\n You've got to get this grating up\n\n right away.\n\nTwo passersby enter.\n\n FIRST PASSERBY\n\n What's the trouble?\n\n BRUNO\n\n (yelling)\n\n Can't we do something...!\n\n (to passerby)\n\n I dropped my cigarette case.\n\n PORTER\n\n (looking down)\n\n Mightn't be any good, mister.\n\n Probably gone down the storm drain.\n\n BRUNO\n\n (horrified)\n\n Storm drain?\n\n FIRST PASSERBY\n\n On the other hand, it might have\n\n lodged on the edge.\n\n SECOND PASSERBY\n\n Don't they have a trap down there --\n\n like under a sink?\n\n BRUNO\n\n (excited)\n\n Don't just stand here -- do something!\n\n PORTER\n\n (calmly)\n\n Guess we could phone the city\n\n engineer, all right.\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 121.\n\n PORTER (CONT'D)\n\n Worst he could do would be to tell\n\n me to take a running jump and --\n\n (Bruno grabs his arm.\n\n Porter shakes Bruno\n\n off)\n\n Relax, mister.\n\n BRUNO\n\n I don't want to relax.\n\nHe goes on his knees and forces his arm down the drain.\n\nINT. THE ANNOUNCER'S BOOTH FOREST HILLS\n\n ANNOUNCER\n\n (with great excitement)\n\n This is more than a tennis game,\n\n ladies and gentlemen -- it's a\n\n desperate fight with Guy Haines\n\n playing as if his life depended on\n\n it!\n\nMED. SHOT\n\nGuy is volleying.\n\nMED. SHOT\n\nReynolds lobs.\n\nCLOSEUP\n\nGuy smashes.\n\nCLOSE SHOT\n\nReynolds lobs again.\n\nCLOSE SHOT\n\nGuy smashes.\n\nCLOSE SHOT\n\nReynolds misses and the ball hits inside the line.\n\n Converted to PDF by www.screentalk.org 122.\n\nCLOSEUP\n\nThe Umpire calling.\n\n UMPIRE\n\n Game to Mr. Haines. Mr. Reynolds\n\n leads five games to four...third\n\n set.\n\nEXT. METCALF STATION MED. SHOT\n\nA few more passersby have stopped to watch Bruno, whose arm\n\nis pushed through the grating.\n\nCLOSEUP\n\nBruno's face -- straining.\n\nCLOSEUP\n\nUnder the grating Bruno's hand is groping. His fingers are\n\na long way from the lighter.\n\nLONG SHOT FOREST HILLS\n\nwith the game in progress.\n\nMED. SHOT EXT. CLUB\n\nA taxi has pulled up. Barbara gets out.\n\nCLOSE SHOT\n\nShe takes the ten dollar bill from her purse and passes it\n\nto the driver. She gives a final look inside the cab.\n\nCLOSEUP\n\nOn the seat are Guy's everyday pants, laid out.\n\nMED. SHOT\n\nBarbara hurries out of the picture toward the club.\n\n Converted to PDF by www.screentalk.org 123.\n\nLONG SHOT\n\nThe crowd watching.\n\nCLOSEUP\n\nThe tense face of Anne.\n\nCLOSEUP\n\nThe Umpire is somewhat impressed.\n\nINT. THE ANNOUNCER'S BOOTH CLOSEUP\n\nThe announcer is telling his listeners that the score is now\n\nsix-five in favor of Haines. That he has pulled up\n\nwonderfully and only needs one more game to win the match.\n\nEXT. COVERED STAND ENTRANCE\n\nBarbara, very nervous but trying to "act natural", passes\n\nHennessy and Hammond. Hammond's eyes again follow her, but\n\nHennessy is intent on the game.\n\nMED. SHOT FEATURING BOX\n\nAs Barbara joins Anne, she gives her a surreptitious signal\n\nby ringing her thumb and forefinger, indicating everything\n\nis set.\n\nCLOSE SHOT\n\nGuy now playing hard.\n\nCLOSEUP\n\nHis racket smashing at the ball.\n\nCLOSEUP\n\nReynolds and his racket hitting the ball back.\n\n Converted to PDF by www.screentalk.org 124.\n\nCLOSEUP THE UMPIRE CALLING\n\n UMPIRE\n\n Advantages, Mr. Haines.\n\nCLOSEUP\n\nGuy serving.\n\nCLOSEUP\n\nHis ball hitting the racket.\n\nCLOSEUP\n\nThe ball in the net.\n\nCLOSEUP\n\nA second ball hitting the net. The Umpire's voice calling:\n\n UMPIRE'S VOICE\n\n (O.S)\n\n Duece!\n\nEXT. METCALF STATION\n\nA LOW SHOT ON Bruno bent over the grating and the onlookers\n\nbehind him.\n\nBIG HEAD CLOSEUP BRUNO\n\nstraining and panicky.\n\nCLOSEUP\n\nUnder the grating, Bruno's fingers get near the lighter, and\n\nin their groping, they knock the lighter off the ledge, onto\n\nthe ledge below.\n\nCLOSEUP\n\nBruno's horror-stricken face.\n\n Converted to PDF by www.screentalk.org 125.\n\nFOREST HILLS MED. SHOT\n\nGuy still playing.\n\nCLOSE SHOT\n\nBarbara standing with Hennessy, watching. We HEAR the score.\n\n UMPIRE'S VOICE\n\n (O.S)\n\n Advantage, Mr. Reynolds.\n\nCLOSEUP ANNE\n\nunable to bear the suspense. She glances O.S.\n\nMED. SHOT\n\nThe waiting cab.\n\nCLOSE SHOT\n\nGuy and Reynolds in play.\n\n UMPIRE'S VOICE\n\n (O.S)\n\n Score is deuce.\n\nCLOSE SHOT\n\nReynolds serves.\n\nCLOSE SHOT\n\nGuy volleys. He waits for the return ball. He misses it.\n\n UMPIRE'S VOICE\n\n (O.S)\n\n Advantage, Mr. Reynolds.\n\nEXT. METCALF STATION\n\nANGLE SHOOTING THROUGH the grating at CLOSEUP BRUNO'S HEAD\n\nAND SHOULDERS staining.\n\n Converted to PDF by www.screentalk.org 126.\n\nCLOSEUP\n\nUnder the grating, Bruno's fingers go lower and lower,\n\nstraining to reach the lighter, which is still a few inches\n\nout of reach.\n\nFOREST HILLS MED. SHOT\n\nGuy is volleying with Reynolds.\n\nINT. ANNOUNCER'S BOOTH\n\nHe is very excited.\n\n ANNOUNCER\n\n -- Haines hasn't let up for a moment.\n\n If he wins this set, he wins the\n\n whole match!\n\nCLOSEUP ANNE AND BARBARA\n\nin their box. They are extremely tense.\n\nMED. SHOT\n\nGuy slams hard a shot that wins him the game.\n\nLONG SHOT CROWD\n\napplauding and shouting.\n\nCLOSE SHOT ANNE AND BARBARA\n\nAt an urgent signal from Anne, Barbara hurries out as if she\n\nknew what she had to do.\n\nLONG SHOT\n\nGuy shakes hands with his opponent, and then hurries across\n\nto Anne in the stand. He leans over the front of the box.\n\nWhile congratulating him outwardly, she whispers something\n\nto him. He leaves his racket with her and hurries away.\n\n Converted to PDF by www.screentalk.org 127.\n\nMED. SHOT STAND ENTRANCE\n\nA block of people leaving cut off Hennessy's view. Barbara\n\ntries desperately to keep his attention off Guy.\n\n BARBARA\n\n (breathlessly)\n\n Isn't it wonderful, Mr. Hennessy?\n\n He won! It calls for a celebration.\n\n Anne says you must have dinner with\n\n us. Just the family, and you, and\n\n Guy.\n\n HENNESSY\n\n (awkwardly)\n\n Sorry I can't make it. Business.\n\n BARBARA\n\n But Guy is your business. You'll be\n\n with him, won't you?\n\n HENNESSY\n\n (a little grimly)\n\n Yeah -- I'll be with Guy.\n\nMED. SHOT\n\nGuy moving along the front of the stand making for another\n\nexit.\n\nCLOSE SHOT\n\nBarbara takes it for granted that Hennessy will accept her\n\ninvitation.\n\n BARBARA\n\n Guy says you love steak -- rare,\n\n Medium, or well-done?\n\n HENNESSY\n\n I sure wish I could --\n\nSEMI CLOSEUP\n\nHammond is looking off. He calls into the stand.\n\n HAMMOND\n\n Hennessy!\n\nHe points off toward Guy.\n\n Converted to PDF by www.screentalk.org 128.\n\nMED. SHOT\n\nGuy is hurrying toward the public entrance of the stand.\n\nSEMI CLOSEUP\n\nHennessy and Hammond move off, leaving a dismayed Barbara.\n\nSEMI LONG SHOT\n\nGuy hurrying under the stand toward the waiting cab.\n\nMED. SHOT\n\nThe two men hurrying after him.\n\nEXT. CLUB\n\nGuy goes to the waiting cab and gets in. The cab moves off.\n\nMED. SHOT\n\nThe two men hurry out of the club and stand helplessly looking\n\nafter the departing cab. They hurry out of the picture.\n\nCLOSE SHOT\n\nWe see them grab another car. It is a chauffeur-driven\n\nlimousine. Hammond jumps in front and seats himself beside\n\nthe driver. Hennessy hops in the back. The car moves off.\n\nINT. LIMOUSINE TWO SHOT\n\nHennessy finds himself seated by an old dowager about seventy-\n\nfive years of age. She looks startled for a moment and almost\n\nrecoils from him. He shows her his badge.\n\n HENNESSY\n\n If you'll pardon us, madam, we need\n\n your help. We're chasing a man.\n\nThe old lady's eyes light up.\n\n DOWAGER\n\n How exciting.\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 129.\n\n DOWAGER (CONT'D)\n\n (she leans forward\n\n and calls to the\n\n chauffeur)\n\n Hurry, O'Toole! Hurry!\n\nShe leans back and maintains her air of excitement as she\n\nlooks across at Hennessy.\n\nCLOSE SHOT INSIDE THE TAXI\n\nGuy is busy changing his pants. He glances over his shoulder.\n\nINT. CAR\n\nThe two men looking ahead toward Guy.\n\nEXT. METCALF STATION\n\nCLOSEUP BRUN0'S FACE - ANGLE SHOOTING UP to get the peering\n\nfaces behind him. Bruno still frantically trying to reach\n\nthe lighter.\n\nCLOSEUP\n\nUnder the grating Bruno's fingers slowly closing in on the\n\nlighter. They barely manage to grasp it.\n\nCLOSEUP\n\nBRUN0'S FACE -- triumphant.\n\nCLOSEUP\n\nBruno's fist, holding the lighter, comes through the grating.\n\nCLOSE SHOT\n\nBruno straightens up. CAMERA BACK as all the onlookers turn\n\ntheir heads in his direction.\n\n ONLOOKER\n\n You sure must think a lot of that --\n\n Whatever it is.\n\n Converted to PDF by www.screentalk.org 130.\n\nBruno doesn't answer. With the lighter in his closed fist,\n\nhe darts through the crowd, the people looking after him.\n\nLONG SHOT\n\nThe sun is much lower.\n\nINT. CLUB CAR\n\nGuy is now glancing at his watch. The sun is behind him and\n\nvery much lower.\n\nEXT. AMUSEMENT PARK\n\nBruno is looking at his watch and then across at the sky.\n\nLONG SHOT FROM HIS VIEWPOINT\n\nThe last trace of the setting sun has gone.\n\nEXT. METCALF STATION MED. SHOT\n\nGuy is stepping off the train. He crosses to a waiting taxi,\n\nCAMERA FOLLOWING him.\n\nCLOSE SHOT\n\n GUY\n\n (to the driver)\n\n The amusement park, quick.\n\nAs he gets in the Cab, we go to --\n\nCLOSE SHOT MAN\n\nwatching Guy get into taxi. As we hear the taxi drive away,\n\nthe man hurries across to a waiting police car.\n\nCLOSE SHOT\n\nHe puts his head in the side window and tells the two waiting\n\ndetectives where Guy has gone.\n\n MAN\n\n Amusement park.\n\n Converted to PDF by www.screentalk.org 131.\n\nWe see one of the detectives take up a microphone as the car\n\ndrives off.\n\nEXT. AMUSEMENT PARK\n\nIt is now getting dark.\n\nMED. SHOT\n\nBruno leaves his spot at the side of the tent and ambles\n\nover toward the queue of people waiting for boats.\n\nCLOSE SHOT BRUNO\n\njoining the queue. He glances ahead of him.\n\nMED. SHOT FROM HIS VIEWPOINT\n\nWe see the light above the pay booth go on, shedding a\n\ndownward glare.\n\nCLOSE SHOT BRUNO\n\npulls his hat a little further over his, eyes. Some new\n\narrivals join the queue behind him.\n\nINT. TAXI\n\nGuy looking anxiously ahead on his way to the amusement park.\n\nAMUSEMENT PARK ENTRANCE\n\nWe see a police car arrive. One uniformed man and two\n\ndetectives get out of the car and make their way toward the\n\nentrance. One of to detectives stands at the entrance while\n\nthe other two hurry into the grounds.\n\nMED. SHOT\n\nGuy's taxi arrives.\n\nMED. SHOT\n\nAcross the street, another police car arrives.\n\n Converted to PDF by www.screentalk.org 132.\n\nMED. SHOT\n\nAt Guy is paying his cab fare, he glances around him.\n\nMED. SHOT FROM HIS VIEWPOINT\n\nHe sees one police car.\n\nCLOSE SHOT GUY\n\ngives a furtive glance around while waiting for his change.\n\nMED. SHOT ANOTHER POLICE CAR\n\nMED. SHOT\n\nGuy cautiously makes his way toward the entrance to the\n\nAmusement Park.\n\nMED. SHOT\n\nGuy passes the waiting detective and looks off. From his\n\nviewpoint we see:\n\nMED. SHOT THE TWO DETECTIVES\n\nwho were at the station indicate Guy is the man.\n\nMED. SHOT\n\nOne detective turns away and starts to follow Guy.\n\nCLOSE SHOT BRUNO\n\nin the queue of people. He is edging slowly along. He is\n\nabout ten people away from the entrance. He suddenly looks\n\nahead and sees.\n\nFROM HIS VIEWPOINT\n\nThe uniformed man and the detective are talking casually to\n\nthe boat men in charge of the concession.\n\n Converted to PDF by www.screentalk.org 133.\n\n DETECTIVE\n\n (to boatman)\n\n The killer is here tonight. So keep\n\n your eyes open and the minute you\n\n see him, let us know.\n\nCLOSE SHOT\n\nThe boatman looks at them with an expression of alarm.\n\nCLOSE SHOT BRUNO\n\nbegins to look a little uneasy. We see him begin to mentally\n\ndeliberate.\n\nMED. SHOT\n\nGuy, threading his way through the crowds, conscious that he\n\nis being followed, but nevertheless, on the lookout for Bruno.\n\nCLOSE SHOT BRUNO\n\nmoving along the line. CAMERA MOVES IN until his head and\n\nshoulders fill the screen. He is now coming within range of\n\nthe flood-lit pay-box. The light seems to creep up across\n\nhis chest and slowly reveal his face. He lowers his head.\n\nMED. SHOT\n\nThe boatman begins to look along the queue. There is an\n\nexpression of growing recognition on his face.\n\nMED. SHOT\n\nBruno sees this, makes a decision and casually deserts the\n\nqueue of people.\n\nMED. SHOT\n\nThe boatman hurries across to the uniformed man and begins\n\nto talk to him excitedly, looking in Bruno's direction.\n\nMED. SHOT GUY\n\nComing along and looking for Bruno. His eyes light up.\n\n Converted to PDF by www.screentalk.org 134.\n\nSEMI-LONG SHOT FROM HIS VIEWPOINT\n\nWe see Bruno making his way from the queue of people.\n\nCLOSE SHOT GUY\n\ncalls to Bruno.\n\n GUY\n\n Hey, Bruno.\n\nCLOSE SHOT BRUNO\n\ngives a quick glance back, sees Guy then he turns and looks\n\noff in another direction.\n\nSEMI-LONG SHOT\n\nThe uniformed man and the boatman approaching him.\n\nCLOSE SHOT\n\nBruno hurries on. He stop short as he sees.\n\nSEMI-LONG SHOT FROM HIS VIEWPOINT\n\nAnother uniformed man.\n\nMED. SHOT\n\nBruno starts to run.\n\nMED. SHOT\n\nGuy starts to run after him.\n\nMED. SHOT\n\nBruno is seen to jump on a merry-go-round, which is just\n\nstarting. Its pace is already pretty fast.\n\nMED. SHOT\n\nGuy runs toward Bruno.\n\n Converted to PDF by www.screentalk.org 135.\n\nCLOSE SHOT\n\n DETECTIVE\n\n Haines! Hold it! Hold it!\n\nThe detective pulls out his gun and starts to run after Guy.\n\nSEMI-LONG SHOT\n\nGuy jumps on the merry-go-round after Bruno. Its speed is\n\nso great that he nearly gets flung off.\n\nCLOSE SHOT\n\nThe detective fires at Guy.\n\nCLOSE SHOT\n\nThe man running the machine in the center of the merry-go-\n\nround is suddenly hit in the shoulder.\n\nCLOSE SHOT\n\nHis hand, which is on the starting lever, jerks it down.\n\nMED. SHOT\n\nThe detective, after Guy, jumps on the machine but is flung\n\noff on his back.\n\nFULL SHOT\n\nThe merry-go-round has now started to increase the speed.\n\nCLOSE SHOT\n\nBruno at the far side is trying to jump off, but it's going\n\ntoo fast.\n\nLONG SHOT FROM HIS VIEWPOINT\n\nWe see the hard ground whizzing past him. Everything seems\n\nto be a blur. We get a glimpse of screaming women and the\n\ncrowds rushing up from the midway.\n\n Converted to PDF by www.screentalk.org 136.\n\nCLOSE SHOT BRUNO\n\nHe turns and glances over his shoulder.\n\nMED. SHOT FROM HIS VIEWPOINT\n\nGuy is threading his way between the rising and falling\n\nhorses. Guy gets right up close to him.\n\nTWO SHOT\n\nAs Guy comes near to Bruno, the latter turns on him and starts\n\nto attack him.\n\n BRUNO\n\n I want to get off of here! Let me\n\n off of here! It makes me dizzy.\n\n GUY\n\n Stop it, Bruno. Give me my lighter,\n\n Bruno!\n\nMED.SHOT\n\nAgainst the whirling background of the merry-go-round, Turley\n\nand Campbell rush up as the detective struggles to his feet,\n\nslightly hurt. The noise from the calliope is very loud.\n\n CAMPBELL\n\n (to Turley, puzzled;\n\n indicating the merry-\n\n go-round)\n\n Who's the man he's fighting with on\n\n there?\n\nAt this moment the boatman rushes up.\n\n BOATMAN\n\n (excited)\n\n There he is! That's the one! That's\n\n the one who killed her!\n\n TURLEY\n\n Of course he is. We know that.\n\nCLOSE SHOT ON MERRY-GO-ROUND\n\nGuy and Bruno in a struggle. Guy has to protect himself\n\nfrom a madman whose hands attempt to reach his throat.\n\n Converted to PDF by www.screentalk.org 137.\n\nThey are staggering across between the rising and falling\n\nhorses.\n\nMED. SHOT OUTSIDE MERRY-GO-ROUND\n\nA detective turns to the group around hIm.\n\n DETECTIVE\n\n Get somebody to come and stop that\n\n thing!\n\nAn elderly man in soiled work clothes speaks up.\n\n WORKMAN\n\n I'll handle it.\n\nImmediately the workman heads straight for the merry-go-round\n\nand starts to crawl under it on his stomach.\n\n DETECTIVE\n\n (calls after him)\n\n Hey! Be careful! Stop!\n\nA second detective speaks to him quizzically.\n\n 2ND DETECTIVE\n\n Well, do you want to do it yourself?\n\nThe first detective leans over and looks off toward the\n\nworkman who is continuing his slithering way under the\n\nmachine, then straightens up.\n\n 1ST DETECTIVE\n\n (changing his mind)\n\n No. I think he'll make it all right.\n\nMED. SHOT GUY AND BRUNO\n\nBruno swings around till his back is to us. He pushes Guy\n\ntoward the edge, but Guy manages to grab the rein of the\n\nnearest horse. The momentum of the machine swings Guy around\n\nagainst the horse, whose big head towers in the f.g. Bruno,\n\non this side of the horse pushes forward and tries to grab\n\nthe reins from Guy's hand. He tries to slash at Guy's face.\n\nThe back of Bruno's head is toward us during this. Guy\n\nsuddenly leans out across the horse and smashes his fist\n\nagainst Bruno's face. Bruno's head goes back until it is in\n\nthe f.g. in a upside-down position.\n\n Converted to PDF by www.screentalk.org 138.\n\nMED. SHOT\n\nTHE CAMERA IS LOW so that we get the effect of Bruno falling\n\ninto the CAMERA from Guy's blow!\n\nMED. SHOT\n\nIn the f.g. is a young boy of four years. He is excited by\n\nthe speed of the ride and laughs at the fight with great\n\nenjoyment. He sees this by suddenly glancing over his\n\nshoulder. In the b.g. Guy and Bruno are continuing their\n\nfight. Bruno rises. Guy staggers after him. Bruno again\n\nleaps upon Guy. The two men sway toward the CAMERA until\n\nBruno gets alongside the little boy. The boy now shows some\n\nanxiety. The three figures now fill the screen with the\n\nhorses' heads in the f.g. Bruno is forced against the little\n\nboy, who now, alarmed, beats Bruno on the cheek with one\n\nhand, the other holding onto the brass rail in front. Bruno\n\nstops and with a sweep of his arm, knocks the little boy off\n\nthe horse onto the floor below. The little boy, in falling,\n\ngrabs the horse's rein or stirrup.\n\nCLOSE SHOT\n\nGuy breaks away from Bruno and dives around the back of the\n\nhorse to grab the little boy.\n\nCLOSE SHOT\n\nAs Guy grabs the boy, he staggers forward with him to a small\n\ngondola. Bruno leaps onto his back but Guy manages to put\n\nthe boy in the gondola.\n\nCLOSE SHOT UNDERNEATH THE WHIRLING MERRY-GO-ROUND\n\nThe boat man is making slow progress.\n\nFROM HIS VIEW POINT\n\nWe see his goal. It is the wounded mechanic in the center,\n\nwho is slightly stirring. All during this the base of the\n\nmerry-go-round is skimming above the back and head of the\n\nboat man.\n\n Converted to PDF by www.screentalk.org 139.\n\nBACK ON MERRY-GO-ROUND\n\nThe two men are now in a clinch. Guy tries to fight off the\n\nmaddened Bruno. They are flung between the horses, bouncing\n\none against the other, almost half way around the merry-go-\n\nround.\n\nCLOSE SHOT BRUNO AND GUY\n\nAgain they struggles between two horses. On each side of\n\nthem are two young screaming girls. The two bounce from one\n\nhorse to the other.\n\nCLOSE SHOT\n\nThe calliope has little figures and these boat away on their\n\ncymbals almost as though they are applauding what's going\n\non.\n\nCLOSE SHOT\n\nUnderneath the merry-go-round, the boat man has made further\n\nprogress. He is creeping inch by inch. His nose starts to\n\nrun. He starts to fumble for a dirty piece of handkerchief.\n\nHe blows his nose and then moves on.\n\nCLOSE SHOT\n\nBack above the two men swinging past the two girls on their\n\nhorse and they both crash to the floor underneath another\n\nhorse, upon which is riding side-saddle, a mother and her\n\nthree-year-old little girl.\n\nCLOSEUP\n\nThe two big heads of the men, battling. The two men roll\n\nunderneath the horse's hoofs, which are seen rising and\n\nfalling. They get right underneath one horse.\n\nCLOSEUP\n\nGuy has turned over on his back and his eyes look up.\n\n Converted to PDF by www.screentalk.org 140.\n\nCLOSEUP FROM HIS VIEWPOINT\n\nWe see the big horse's head above and its hoofs coming down\n\ntoward the CAMERA and filling the screen. We get a faint\n\nimpression of the screaming mother hugging her child to her\n\nbreast, above.\n\nBIG CLOSEUP THE HORSE'S HOOFS\n\nstriking Guy's head.\n\nCLOSE SHOT\n\nGuy wrenches himself out of this position. He rolls away\n\nfrom the CAMERA right to the edge of the merry-go-round. He\n\nmanages to grab a rail.\n\nMED. SHOT\n\nGuy's body is flung out horizontally. We see the crowd behind\n\nback-up for fear of being knocked over. The screw of tension\n\nincrease. Over this comes the sound of an approaching\n\nambulance siren.\n\nCLOSE SHOT\n\nBruno edges himself toward Guy. He is hanging on to the\n\nreins of a horse. His feet manage to roach Guy's knuckles.\n\nCLOSEUP BRUNO'S VICIOUS EXPRESSION\n\nCLOSEUP BRUNO'S FEET\n\nkicking at Guy's knuckles.\n\nCLOSEUP GUY'S AGONIZED EXPRESSION\n\nMED. SHOT\n\nA flash of the horror-stricken faces of the spectators seen\n\nthrough the whirling machine.\n\n Converted to PDF by www.screentalk.org 141.\n\nCLOSEUP\n\nMachinery and the lever that was pulled on too fast. The\n\nBoatman's hand comes up into the picture and pulls the lever\n\nover.\n\nLONG SHOT\n\nThe sudden braking causes the whole merry-go-round to topple\n\nover with a grinding roar.\n\nLONG SHOT FROM HIGH ANGLE\n\nThe merry-go-round his keeled over. For a moment we don't\n\nknow who has survived. There is a surge of people milling\n\nand shouting. Those who have jumped back out of the way\n\nwhen the merry-go-round toppled, now rush forward again as\n\nthe cloud of dust settles. From the midway in the background\n\nothers are running forward.\n\nMED. LONG SHOT\n\nDistraught parents try to force their way to their children\n\nwho were on the merry-go-round, but are hold back from the\n\nwreckage by police.\n\nMED. CLOSE SHOT\n\nGuy is somewhat stunned from his fall. He is helped to his\n\nfeet by some men in the crowd. His knuckles are bleeding.\n\nIn the background people are rushing about. The crowd is in\n\nuproar as women and children are helped from the wreckage.\n\nOfficials and uniformed policemen pushing back the surge of\n\nthe crowd.\n\n AD LIBS\n\n Get back. Get back there. Give us\n\n room here.\n\nTurley and Campbell rush in to Guy.\n\n TURLEY\n\n Are you all right, Haines?\n\n GUY\n\n Yes, I think so.\n\n Converted to PDF by www.screentalk.org 142.\n\nGuy is surrounded by police and Campbell stands at his elbow.\n\nAt this moment the boatman runs in. One of the detectives\n\nis with him.\n\n DETECTIVE\n\n Mr. Turley! Mr. Turley!\n\n (indicating boatman)\n\n He says this isn't the man we want.\n\n (with a nod in Guy's\n\n direction)\n\n It's the other one -- the one he was\n\n fighting with.\n\n TURLEY\n\n (stops to give his\n\n full attention to\n\n this unexpected bit\n\n of information)\n\n What do you mean, this isn't the --\n\n (turns to Guy, not\n\n quite taking it in)\n\n Not Haines?\n\n (back to boatman)\n\n But you said he was. You pointed\n\n him out.\n\n BOATMAN\n\n No, I didn't, sir. I've never seen\n\n this man before in my life. I meant\n\n the other one.\n\nThe detective who was holding Guy instinctively relax his\n\nhold on Guy's arm. Turley turns to Guy, puzzled.\n\n TURLEY\n\n What is this all about, Haines? Did\n\n you know he killed your wife?\n\n GUY\n\n (nods)\n\n He has my cigarette lighter and wanted\n\n to plant it there on the island to\n\n pin the whole thing on me.\n\n (urgently)\n\n Let me talk to him. Let me show\n\n you. Where is he?\n\n ANOTHER DETECTIVE\n\n Over here.\n\nHe leads the way. They follow.\n\n Converted to PDF by www.screentalk.org 143.\n\nMED. CLOSE SHOT\n\nas Guy and Turley enter to the spot where Bruno is pinned\n\nunder the overturned machine. He is caught between two of\n\nthe horses, the head of one of them across his chest. Bruno's\n\nhead sags back somewhat, but is resting on pieces of debris.\n\nA uniformed policeman looks up from Bruno to Turley:\n\n POLICEMAN\n\n This one's in a pretty bad way, Mr.\n\n Turley.\n\nGuy is shocked at the sight of Bruno.\n\n GUY\n\n (looking down at Bruno)\n\n Can't you get that stuff off him?\n\n POLICEMAN\n\n No, they've done everything they can\n\n until the crane comes.\n\nBruno opens his eyes and sees Guy.\n\n BRUNO\n\n Hello, Guy.\n\nTurley has leaned forward to look at the helpless Bruno.\n\n BRUNO\n\n (weakly nodding at\n\n Turley)\n\n Who's that?\n\n GUY\n\n This is Mr. Turley, Chief of Police.\n\n BRUNO\n\n (with a half smile)\n\n So they got you at last, eh, Guy?\n\nGuy looks around desperately, frustrated for a moment as\n\nTurley eyes him stonily. Then he turns again to Bruno.\n\n GUY\n\n (rather gently)\n\n Can you talk a little? Can you tell\n\n the chief you have my lighter?\n\n Converted to PDF by www.screentalk.org 144.\n\n BRUNO\n\n (with a faint,\n\n quizzical smile)\n\n I haven't got it. It's still on the\n\n island where you left it.\n\nGuy looks around helplessly to Turley, who looks back at him\n\nsuspiciously.\n\n DETECTIVE\n\n (looking down at Bruno)\n\n I think he's going.\n\nTurley leans over to look.\n\nCLOSE SHOT BRUNO'S FIST FROM TURLEY'S VIEWPOINT\n\nAs Bruno is dying, his closed fist slowly starts to open.\n\n DETECTIVE'S VOICE\n\n He's finished.\n\nGuy's lighter is now revealed in Bruno's open hand.\n\nMED. SHOT GROUP\n\nTurley takes the lighter from the dead Bruno's hand. Guy is\n\nwatching him. Turley straightens up and holds the lighter\n\nout to him.\n\n TURLEY\n\n Is this your lighter, Haines?\n\nGuy nods without speaking, and with a half look in Bruno's\n\ndirection.\n\n TURLEY\n\n Well, you were right.\n\n (sticks the lighter\n\n into his own pocket)\n\n I'd better keep this for the time\n\n being.\n\n (in a friendly tone)\n\n We can clear the whole thing out the\n\n morning. How about staying in town\n\n over night, Haines? I imagine you\n\n have a lot to tell me. Nine o'clock,\n\n all right?\n\n Converted to PDF by www.screentalk.org 145.\n\n GUY\n\n (nods)\n\n Okay, Mr. Turley. Thanks.\n\nTurley turns back to the group around Bruno. Guy looks down\n\nfor a moment at Bruno, then speaks to the boatman, who is\n\nstanding nearby.\n\n GUY\n\n Can you tell me where there's a\n\n telephone?\n\n BOATMAN\n\n (indicating)\n\n There's one up near the entrance.\n\n (with a look back to\n\n the dead Bruno)\n\n Who was he, Bud?\n\nGuy looks back sympathetically in Bruno's direction, speaks\n\nwithout looking at the boatman.\n\n GUY\n\n Bruno. Bruno Antony.\n\n (reminiscently and a\n\n little\n\n compassionately,\n\n remembering what\n\n Bruno had said of\n\n himself)\n\n A very clever fellow.\n\nHe moves off through the crowd.\n\n DISSOLVE TO:\n\nINT. BURTON STUDY NIGHT\n\nAnne, Barbara and the Senator are sitting silently in the\n\nattitudes of waiting. The telephone rings. Anne is instantly\n\non her feet. Barbara and the Senator watch her anxiously as\n\nshe goes to answer it.\n\n ANNE\n\n (into phone)\n\n Hello...\n\n (impatiently)\n\n Yes, operator, yes!\n\n (waits a moment, then\n\n eagerly:)\n\n Guy?\n\n (MORE)\n\n Converted to PDF by www.screentalk.org 146.\n\n ANNE (CONT'D)\n\n (a pause, then she\n\n closes her eyes with\n\n heartfelt relief.\n\n Another pause, then:)\n\n Yes, darling, yes. Of course I'll\n\n be there...Goodbye.\n\nShe hangs up, turns slowly, to face Barbara and her father.\n\nHer expression is one of intense relief.\n\n ANNE\n\n Guy'll be back tomorrow.\n\n (overcome with emotion\n\n she has difficulty\n\n in speaking)\n\n He wants me to take him some things.\n\nWith a sob, Barbara flings herself into Anne's arms. As she\n\ncries, Anne strokes her head comfortingly. Then with a half-\n\nchoked sobs Anne, too, begins to cry. She speaks through\n\nher tears, looking over Barbara's shoulder at her father.\n\n ANNE\n\n He says he looks silly in his tennis\n\n clothes.\n\nThe Senator eyes them a moment, then speaks a little wryly:\n\n SENATOR\n\n I presume from all those tears that\n\n you have had good news.\n\n DISSOLVE TO:\n\nINT. PARLOR OF TRAIN NEXT DAY\n\nAnne and Guy are sitting quietly together. Opposite them is\n\na man in a clerical collar who is reading a sports magazine.\n\nOn the cover is a picture of a tennis player in action. The\n\nman looks over the top of his magazine at Guy, with\n\nrecognition. He leans forward.\n\n CLERIC\n\n I beg your pardon, but aren't you\n\n Guy Haines?\n\n GUY\n\n (uncomfortably)\n\n Yes.\n\n Converted to PDF by www.screentalk.org 147.\n\nGuy and Anne exchange a brief look, rise hurriedly and start\n\nto walk away before the conversation can go any farther.\n\nThe cleric looks after them with a frown and a puzzled shrug\n\nof his shoulders, as if to say, "Did I say something wrong?"\n\n FADE OUT.\n\n THE END\n\n |
| 17 | script_sunsetblvd..txt | deed3ee1-dae | Script | <b><!--\n\n</b>if (window!= top)\n\ntop.location.href=location.href\n\n<b>// -->\n\n</b>\n\nSunset Boulevard\n\n\tSUNSET BOULEVARD\n\n Charles Brackett\n\n Billy Wilder\n\n D.M. Marshman, Jr.\n\n March 21,1949\n\n SEQUENCE "A" \n\n A-l-4 START the picture with the actual street sign:\n\n SUNSET BOULEVARD, stencilled on a curbstope.\n\n In the gutter lie dead leaves, scraps of paper,\n\n burnt matches and cigarette butts. It is early\n\n morning.\n\n Now the CAMERA leaves the sign and MOVES EAST, the\n\n grey asphalt of the street filling the screen. As\n\n speed accelerates to around 40 m.p.h., traffic de-\n\n marcations, white arrows, speed-limit warnings, man-\n\n hole covers, etc., flash by. SUPERIMPOSED on all\n\n this are the CREDIT TITLES, in the stencilled style\n\n of the street sign.\n\n Over the scene we now hear MAN'S VOICE\n\n sirens. Police squad cars Yes, this is Sunset\n\n hurtle toward the camera, Boulevard, Los Angeles,\n\n turn off the road into a California. It's about\n\n driveway with squealing five o'clock in the\n\n brakes. Dismounted motor- morning. That's the\n\n cycle cops stand directing Homicide Squad, com-\n\n the cars in. plete with detectives\n\n and newspaper men.\n\n A-5 PATIO AND POOL OF A murder has been re-\n\n MANSION ported from one of those\n\n great big houses in the\n\n The policemen and news- ten thousand block.\n\n paper reporters and You'll read all about\n\n photographers have it in the late editions,\n\n jumped out of the cars I'm sure. You'll get\n\n and are running up to it over your radio,\n\n the pool, in which a and see it on tele-\n\n body is seen floating. vision -- because an\n\n Photographers' bulbs old-time star is in-\n\n flash in rapid suc- volved. one of the big-\n\n cession. gest. But before you\n\n hear it all distorted\n\n and blown out of\n\n proportion, before those\n\n Hollywood columnists\n\n get their hands on it,\n\n maybe you'd like to\n\n hear the facts, the\n\n whole truth...\n\n A-6 FLASH OF THE BODY\n\n MAN'S VOICE\n\n Angle up through the If so, you've come to the\n\n water from the bottom right party... You see,\n\n of the pool, as the the body of a young man\n\n body floats face down- was found floating in the\n\n ward. It is a well- pool of her mansion, with\n\n dressed young man. two shots in his back and\n\n one in his stomach. No-\n\n body important, really.\n\n Just a movie writer with\n\n a couple of "B" pictures\n\n to his credit. The poor\n\n dope. He always wanted a\n\n pool Well, in the end\n\n he got himself a pool --\n\n SLOW DISSOLVE TO: only the price turned out\n\n to be a little high...\n\n Let's go back about six\n\n A-7 HOLLYWOOD, SEEN FROM months and find the day\n\n THE HILLTOP AT IVAR when it all started.\n\n & FRANKLIN STREETS\n\n It is a crisp sunny I was living in an\n\n day. The voice con- apartment house above\n\n tinues speaking as Franklin and Ivar.\n\n CAMERA PANS toward Things were tough\n\n the ALTO NIDO APART- at the moment. I hadn't\n\n MENT HOUSE, an ugly worked in a studio for\n\n Moorish structure ofsat a long time. So I\n\n stucco, about four there grinding\n\n stories high. CAMERA out original stories,\n\n MOVES TOWARD AN OPEN two a week. Only I\n\n WINDOW on the third seemed to have lost\n\n floor, where we look my touch. Maybe they\n\n in on JOE GILLIS' APART- weren't original\n\n MENT. Joe Gillis, bare- enough. Maybe they\n\n footed and wearing no- were too original.\n\n thing but an old bath- All I know is they\n\n robe. is sitting on didn't sell.\n\n the bed. In front of\n\n him. on a straight\n\n chair, is a portable\n\n typewriter. Beside\n\n him, on the bed, is a\n\n dirty ashtray and a\n\n scattering of type\n\n written and pencil-\n\n marked pages. Gillis\n\n is typing. with a\n\n pencil clenched bet-\n\n ween his teeth.\n\n A-8 JOE GILLIS' APARTMENT\n\n It is a one-room affair with an unmade Murphy bed\n\n pulled out of the wall at which Gillis sits typing.\n\n There are a couple of worn-out plush chairs and a\n\n Spanish-style, wrought-iron standing lamp. Also a\n\n small desk littered with books and letters, and a\n\n chest of drawers with a portable phonograph and some\n\n records on top. On the walls are a couple of repro-\n\n ductions of characterless paintings, with laundry\n\n bills and snapshots stuck in the frames. Through an\n\n archway can he seen a tiny kitchenette, complete with\n\n unwashed coffee pot and cup, empty tin cans, orange\n\n peels, etc. The effect is dingy and cheerless --\n\n just another furnished apartment. The buzzer SOUNDS.\n\n GILLIS\n\n Yeah.\n\n The buzzer SOUNDS again. Gillis gets up and opens\n\n the door. Two men wearing hats stand outside one of\n\n them carrying a briefcase.\n\n NO. 1\n\n Joseph C. Gillis?\n\n GILLIS\n\n That's right.\n\n The men ease into the room. No. 1 hands Gillis a\n\n business card.\n\n NO. 1\n\n We've come for the car.\n\n GILLIS\n\n What car?\n\n NO. 2\n\n (Consulting a paper)\n\n 1946 Plymouth convertible. Calif-\n\n ornia license 97 N 567.\n\n NO. 1\n\n Where are the keys?\n\n GILLIS\n\n Why should I give you the keys?\n\n NO. 1\n\n Because the company's played ball\n\n with you long enough. Because\n\n you're three payments behind. And\n\n because we've got a Court order.\n\n Come on -- the keys.\n\n NO. 2\n\n Or do you want us to jack it up\n\n and haul it away?\n\n GILLIS\n\n Relax, fans. The car isn't here.\n\n NO. 1\n\n Is that So?\n\n GILLIS\n\n I lent it to a friend of mine.\n\n He took it up to Palm Springs.\n\n NO. 1\n\n Had to get away for his health,\n\n I suppose.\n\n GILLIS\n\n You don't believe me? Look in\n\n the garage.\n\n NO. 1\n\n Sure we believe you, only now we\n\n want you to believe us. That car\n\n better be back here by noon tomorrow,\n\n or there's going to be fireworks.\n\n GILLIS\n\n You say the cutest things.\n\n The men leave. Gillis GILLIS' VOICE\n\n stands pondering beside Well, I needed about two\n\n the door for a moment. hundred and ninety dollars\n\n Then he walks to the and I needed it real\n\n center of the room and, quick, or I'd lose my car.\n\n with his back to the It wasn't in Palm Springs\n\n CAMERA, slips into a and it wasn't in the\n\n pair of gray slacks. garage. I was way ahead\n\n There is a metallic of the finance company.\n\n noise as some loose\n\n change and keys drop\n\n from the trouser pockets.\n\n As Gillis bends over to\n\n pick them up, we see that\n\n he has dropped the car\n\n keys, identifiable be-\n\n cause of a rabbit's\n\n foot and a miniature\n\n license plate attached\n\n to the key-ring. Gillis\n\n pockets the keys and as\n\n he starts to put on a\n\n shirt\n\n DISSOLVE TO:\n\n A-9 EXTERIOR OF RUDY'S GILLIS' VOICE\n\n SHOESHINE PARLOR (DAY) \n\n I knew they'd be coming\n\n A small shack-like build- around and I wasn't tak-\n\n ing, it stands in the ing any chances, so I\n\n corner of a public park- kept it a couple of\n\n ing lot. Rudy, a blocks away in a parking\n\n colored boy, is giving lot behind Rudy's Shoe-\n\n a customer a shine. shine Parlor. Rudy\n\n never asked any quest-\n\n ions. He'd just look at\n\n your heels and know the\n\n score.\n\n PAN BEHIND the shack to GILLIS' CAR, a yellow 1946\n\n Plymouth convertible with the top down. Gillis enters\n\n the SHOT. He is wearing a tweed sport jacket, a tan\n\n polo shirt, and moooasins. He steps into the car and\n\n drives it off. Rudy winks after him.\n\n A-10 THE ALLEY NEXT TO SIDNEY'S\n\n MEN'S SHOP ON BRONSON AVE. GILLIS' VOICE\n\n I had an original story\n\n Gillis drives into the kicking around Paranount.\n\n alley and parks his car My agent told me it was\n\n right behind a delivery dead as a doornail. but\n\n truck. PAN AND FOLLOW I knew a big shot over\n\n HIM as he gets out, walks there who'd always liked\n\n around the corner into me, and the time had\n\n Bronson and then toward come to take a little\n\n the towering main gate of advantage of it. His\n\n Paramount. A few loafers, name was Sheldrake. He\n\n studio cops and extras are was a smart producer,\n\n lounging there. with a set of ulcers to\n\n prove it.\n\n DISSOLVE TO:\n\n A-11 SHELDRAKE'S OFFICE\n\n It is in the style of a Paramount executive's office --\n\n mahogany, leather, and a little chintz. On the\n\n walls are some large framed photographs of Paramount\n\n stars, with dedications to Mr. Sheldrake. Also a\n\n couple of framed critics' awards certificates, and an\n\n Oscar on a bookshelf. A shooting schedule chart is\n\n thumb-tacked into a large bulletin board. There are\n\n piles or scripts, a few pipes and, somewhere in the\n\n background, some set models.\n\n Start on Sheldrake. He is about 45. Behind his wor-\n\n ried face there hides a coated tongue. He is en-\n\n gaged in changing the stained rilter cigarette in\n\n his Zeus holder.\n\n SHELDRAKE\n\n All right, Gillis. You've got\n\n five minutes. What's your story\n\n about?\n\n GILLIS\n\n It's about a ball player, a rookie\n\n shortstop that's batting 347. The\n\n poor kid was once mixed up in a hold-\n\n up. But he's trying to go straight --\n\n except there's a bunch of gamblers\n\n who won't let him.\n\n SHELDRAKE\n\n So they tell the kid to throw the\n\n World Series, or else, huh?\n\n GILLIS\n\n More or less. Only for the end\n\n I've got a gimmick that's real good.\n\n A secretary enters, carrying a glass or milk.\n\n She opens a drawer and takes out a bottle of pills for\n\n Sheldrake.\n\n SHELDRAKE\n\n Got a title?\n\n GILLIS\n\n Bases Loaded. There's a 4O-page\n\n outline.\n\n SHELDRAKE\n\n (To the secretary)\n\n Get the Readers' Department and\n\n see what they have on Bases Loaded.\n\n The secretary exits. Sheldrake takes a pill and\n\n washes it down with some milk.\n\n GILLIS\n\n They're pretty hot about it\n\n over at Twentieth, but I\n\n think Zanuck's all wet. Can\n\n you see Ty Power as a\n\n GILLIS (cont'd)\n\n shortstop? You've got the best\n\n man for it right here on this lot.\n\n Alan Ladd. Good change of pace for\n\n Alan Ladd. There's another thing:\n\n it's pretty simple to shoot. Lot\n\n of outdoor stuff. Bet you could\n\n make the whole thing for under a\n\n million. And there's a great little\n\n part for Bill Demarest. One of the\n\n trainers, an oldtime player who\n\n got beaned and goes out of his head\n\n sometimes.\n\n The door opens and Betty Schaefer enters -- a clean-\n\n cut, nice looking girl of 21, with a bright, alert\n\n manner. Dressed in tweed skirt, Brooks sweater and\n\n pearls, and carrying a folder of papers. She puts\n\n them on Sheldrake's desk, not noticing Gillis, who\n\n stands near the door.\n\n BETTY\n\n Hello, Mr. Sheldrake. On that Bases\n\n Loaded. I covered it with a 2-page\n\n synopsis.\n\n (She holds it out)\n\n But I wouldn't bother.\n\n SHELDRAKE\n\n What's wrong with it?\n\n BETTY\n\n It's from hunger.\n\n SHELDRAKE\n\n Nothing for Ladd?\n\n BETTY\n\n Just a rehash of something that\n\n wasn't very good to begin with.\n\n SHELDRAKE\n\n I'm sure you'll be glad to meet\n\n Mr. Gillis. He wrote it.\n\n Betty turns towards Gillis, embarrassed.\n\n SHELDRAKE\n\n This is Miss Kramer.\n\n BETTY\n\n Schaefer. Betty Schaefer. And\n\n right now I wish I could crawl\n\n into a hole and pull it in after\n\n me.\n\n GILLIS\n\n If I could be of any help...\n\n BETTY\n\n I'm sorry, Mr. Gillis, but I\n\n just don't think it's any good.\n\n I found it flat and banal.\n\n GILLIS\n\n Exactly what kind of material do\n\n you recommend? James Joyce?\n\n Dostoosvsky?\n\n SHELDRAKE\n\n Name dropper.\n\n BETTY\n\n I just think pictures should say\n\n a little something.\n\n GILLIS\n\n Oh, you're one of the message\n\n kids. Just a story won't do.\n\n You'd have turned down Gone With the\n\n Wind.\n\n SHELDRAKE\n\n No, that was me. I said, Who\n\n wants to see a Civil War picture?\n\n BETTY\n\n Perhaps the reason I hated Bases\n\n Loaded is that I knew your name.\n\n I'd always heard you had some\n\n talent.\n\n GILLIS\n\n That was last year. This year\n\n I'm trying to earn a living.\n\n BETTY\n\n So you take Plot 27-A, make it\n\n glossy, make it slick --\n\n SHELDRAKE\n\n Carefull Those are dirty words!\n\n You sound like a bunch of New\n\n York critics. Thank you, Miss\n\n Schaefer.\n\n BETTY\n\n Goodbye, Mr. Gillis.\n\n GILLIS\n\n Goodbye. Next time I'll write\n\n The Naked and the Dead.\n\n Betty leaves.\n\n SHELDRAKE\n\n Well, seems like Zanuck's got\n\n himself a baseball picture.\n\n GILLIS\n\n Mr. Sheldrake, I don't want you\n\n to think I thought this was going\n\n to win any Academy Award.\n\n SHELDRAKE\n\n (His mind free-wheeling)\n\n Of course, we're always looking\n\n for a Betty Hutton. Do you see\n\n it as a Betty Hutton?\n\n GILLIS\n\n Frankly, no.\n\n SHELDRAKE\n\n (Amusing himself)\n\n Now wait a minute. If we made\n\n it a girls' softball team, put\n\n in a few numbers. Might make a\n\n cute musical: It Happened in\n\n the Bull Pen -- the story of a\n\n Woman.\n\n GILLIS\n\n You trying to be funny? -- because\n\n I'm all out of laughs. I'm over a\n\n barrel and I need a job.\n\n SHELDRAKE\n\n Sure, Gillis. If something should\n\n come along -\n\n GILLIS\n\n Along is no good. I need it now.\n\n SHELDRAKE\n\n Haven't got a thing.\n\n GILLIS\n\n Any kind of assignment. Additional\n\n Dialogue.\n\n SHELDRAKE\n\n There's nothing, Gillis. Not\n\n even if you were a relative.\n\n GILLIS\n\n (Hating it)\n\n Look, Mr. Sheldrake, could you\n\n let me have three hundred bucks\n\n yourself, as a personal loan?\n\n SHELDRAKE\n\n Could I? Gillis, last year some-\n\n body talked me into buying a ranch\n\n in the valley. So I borrowed money\n\n from the bank so I could pay for\n\n the ranch. This year I had to\n\n mortgage the ranch so I could keep\n\n up my life insurance so I could\n\n borrow on the insurance so I could\n\n pay my income tax. Now if Dewey\n\n had been elected -\n\n GILLIS\n\n Goodbye, Mr. Sheldrake.\n\n DISSOLVE TO:\n\n A-12 EXT. SCHWAB'S DRUG STORE\n\n (EARLY AFTERNOON ACTIVITY) GILLIS' VOICE\n\n After that I drove down\n\n MOVE IN toward drug store to headquarters. That's\n\n and the way a lot of us think\n\n about Schwab's Drug Store.\n\n DISSOLVE TO: Actors and stock girls and\n\n waiters. Kind of a\n\n combination office,Kaffee-\n\n A-13 INT. SCHWAB'S DRUG STORE Klatsch and waiting room.\n\n Waiting, waiting for the\n\n The usual Schwabadero gravy train.\n\n crowd sits at the fount-\n\n ain, gossips at the\n\n cigar-stand, loiters by\n\n the magazine display.\n\n MOVE IN towards the TWO\n\n TELEPHONE BOOTHS. In I got myself ten nickels\n\n one of them sits Gillis, and started sending out\n\n a stack of nickels in a general S.O.S. Couldn't\n\n front of him. He's get hold of my agent,\n\n doing a lot of talking naturally. So then I\n\n into the telephone, called a pal of mine,name\n\n hanging up, dropping of Artie Green -- an awful\n\n another nickel, dialing, nice guy, an assistant\n\n talking again. director. He cquld let me\n\n have twenty, but twenty\n\n wouldn't do.\n\n GILLIS' VOICE (Cont.)\n\n Then I talked to a couple of\n\n yes men at Twentieth. To me\n\n they said no. Finally I\n\n located that agent of mine, the\n\n big faker. Was he out digging\n\n up a job for poor Joe Gillis?\n\n Hmph! He was hard at work in\n\n Bel Air, making with the golf\n\n clubs.\n\n Gillis hangs up with a curse, opens the door of the\n\n booth, emerges, wiping the sweat from his forehead.\n\n He walks toward the exit. He is stopped by the\n\n voice of\n\n SKOLSKY\n\n Hello, Gillis.\n\n Gillis looks around. At the fountain sits Skolsky,\n\n drinking a cup of coffee.\n\n GILLIS\n\n Hello, Mr. Skolsky.\n\n SKOLSKY\n\n Got anything for the column?\n\n GILLIS\n\n Sure. Just sold an original for\n\n a hundred grand. The Life of the\n\n Warner Brothers. Starring the Ritz\n\n Brothers. Playing opposite the\n\n Andrew Sisters.\n\n SKOLSKY\n\n (With a sour smile)\n\n But don't get me wrong -- I love\n\n Hollywood.\n\n Gillis walks out.\n\n DISSOLVE TO:\n\n A-14 THE BEL AIR GOLF LINKS\n\n On a sun-dappled green edged with tall sycamores,\n\n stands Morino, the agent, a caddy and a nondescript\n\n opponent in the background. Gillis has evidently\n\n stated his problem already.\n\n MORINO\n\n So you need three hundred dollars?\n\n Of course, I could give you three\n\n hundred dollars. Only I'm not\n\n going to.\n\n GILLIS\n\n No?\n\n MORINO\n\n Gillis, get this through your\n\n head. I'm not just your agent.\n\n It's not the ten per cent. I'm\n\n your friend.\n\n He sinks his putt and walks toward the next tee,\n\n Gillis following him.\n\n GILLIS\n\n How's that about your being my\n\n friend?\n\n MORINO\n\n Don't you know the finest things\n\n in the world have been written on\n\n an empty stomach? Once a talent\n\n like yours gets into that Mocambo-\n\n Romanoff rut, you're through.\n\n GILLIS\n\n Forget Romanoff's. It's the car\n\n I'm talking about. If I lose my\n\n car it's like having my legs out off.\n\n MORINO\n\n Greatest thing that could happen\n\n to you. Now you'll have to sit\n\n behind that typewriter. Now\n\n you'll have to write.\n\n GILLIS\n\n What do you think I've been doing?\n\n I need three hundred dollars.\n\n MORINO\n\n (Icily)\n\n Maybe what you need is another agent.\n\n He bends down to tee up his ball. Gillis turns away.\n\n DISSOLVE TO:\n\n A-15 GILLIS IN HIS OPEN CAR\n\n GILLIS' VOICE\n\n driving down Sunset As I drove back towards town\n\n towards Hollywood. He I took inventory of my pros-\n\n drives slowly. His pects. They now added up to\n\n mind is working. exactly zero. Apparently I\n\n just didn't have what it takes,\n\n and the time had come to wrap\n\n up the whole Hollywood deal\n\n and go home. Maybe if I hocked\n\n all my junk there'd be enough\n\n for a bus ticket back to Ohio,\n\n back to that thirty-five-\n\n dollar-a-week job behind the\n\n copy desk of the Dayton Evening\n\n Post, if it was still open.\n\n Back to the smirking delight\n\n of the whole office. All\n\n Gillis stops his car at right you wise guys. why don't\n\n a red light by the main you go out and take a crack at\n\n entrance to Bel Air. Hollywood? Maybe you think\n\n Suddenly his eyes fall you could -- Oh-oh!\n\n on:\n\n A-16 ANOTHER CAR\n\n It is a dark-green Dodge business coupe, also waiting\n\n for the light to change. but headed in the opposite\n\n direction. In it are the two finance company men.\n\n They spot Gillis in his car and exchange looks. From\n\n across the intersection Gillis recognizes them and\n\n pulls down the leather sunshade to screen his face.\n\n As the light changes. Gillis gives his car the gun\n\n and shoots away. The men narrowly avoid hitting\n\n another car as they make a U-turn into oncoming\n\n traffic and start after him.\n\n A-17 THE CHASE\n\n to\n\n A-21 Very short, very sharp, told in FLASHES. (Use\n\n locations on Sunset between Bel Air and Holmby Hills).\n\n The men lose Gillis around a bend, catch sight of him\n\n and then -- while they are trapped behind a slow-\n\n moving truck. he disappears again.\n\n A-22 GILLIS\n\n He is driving as fast as he dares, keeping an eye out\n\n for pursuit in his rear-view mirror. Suddenly his\n\n right front tire blows out. Gillis clutches desperately\n\n at the steering wheel and manages to turn the careening\n\n car into\n\n A-23 A DRIVEWAY\n\n It is overgrown with weeds and screened from the street\n\n by bushes and trees. Gillis stops his car about thirty\n\n feet from the street and looks back.\n\n GILLIS' VOICE\n\n Was I far enough ahead?\n\n A-24 THE OTHER CAR\n\n shoots past the driveway, still looking for Gillis.\n\n A-25 GILLIS\n\n He watches his pursuers GILLIS' VOICE\n\n shoot past and out of Yeah...\n\n sight. He opens the\n\n door and looks down at I had landed myself in the\n\n the flat tire. Then he driveway of some big mansion\n\n looks around to see that looked run-down and\n\n where he is. deserted. At the end of the\n\n drive was a lovely sight\n\n A-26 DRIVEWAY WITH GARAGE indeed -- a great big empty\n\n garage, just standing there\n\n An enormous, five-car going to waste. If ever there\n\n affair. neglected and was a place to stash away a\n\n empty-looking. limping car with a hot license\n\n number...\n\n A-27 GILLIS\n\n He gets back into his There was another occupant in\n\n car and carefully pilots that garage: an enormous\n\n the limping vehicle into foreign-built automobile. It\n\n one of the stalls. In must have burned up ten gallons\n\n the adjoining one is a to a mile. It had a 1932\n\n large, dust-covered license. I figured that's\n\n Isotta-Fraschini propped when the owners moved out...\n\n up on blocks. He closes I also figured I couldn't go\n\n the garage door and walks back to my apartment now that\n\n up the driveway. In idle those bloodhounds were on to\n\n curiosity he mounts a me. The idea was to get Artie\n\n stone staircase which Green's and stay there till I\n\n leads to the garden. could make that bus for Ohio.\n\n CAMERA IN BACK OF HIM. Once back in Dayton I'd drop\n\n At the top of the steps the credit boys a picturepost-\n\n he sees the somber pile card telling them where to\n\n of pick up the jallopy.\n\n NORMA DESMOND'S HOUSE GILLIS' VOICE\n\n It is a grandiose -- It was a great big white\n\n Italianate structure, elephant of a place. The kind\n\n mottled by the years, crazy movie people built in the\n\n gloomy, forsaken, crazy Twenties. A neglected\n\n little formal garden house gets an unhappy look.\n\n completely gone to This one had it in spades. It\n\n seed. was like that old woman in\n\n Great Expectations -- that Miss\n\n From somewhere above Haversham in her rotting wed-\n\n comes ding dress and her torn veil,\n\n taking it out on the world be-\n\n cause she'd been given the go-\n\n by.\n\n A WOMAN'S VOICE\n\n You there!\n\n Gillls turns and looks.\n\n A-28 UPSTAIRS LOGGIA\n\n Behind a bamboo blind there is a movement of\n\n a dark figure.\n\n WOMAN'S VOICE\n\n Wlly are you so late? Why have\n\n you kept me waitlng so long?\n\n A-29 GILLIS\n\n He stands flabbergasted. A new noise attracts his\n\n attention -- the creak of a heavy metal-and-glass\n\n door being opened. He turns and sees\n\n A-3O THE ENTRANCE DOOR OF THE HOUSE\n\n Max von Mayerling stands there. He is sixty, and\n\n all in black, except for immaculate white cotton\n\n gloves, shirt, high, stiff collar and a white bow\n\n tie. His coat is shiny black alpaca, his trousers\n\n ledger-atriped. He is semi-paralyzed. The left\n\n side of his mouth is pulled down, and he leans on a\n\n rubber-ferruled stick.\n\n MAX\n\n In here!\n\n Gillis enters the shot.\n\n GILLIS\n\n I just put my car in the garage.\n\n I had a blow-out. I thought --\n\n MAX\n\n Go on in.\n\n There is authority in the gesture of his white-\n\n gloved hand as he motions Gillis inside.\n\n GILLIS\n\n Look, maybe I'd better take my\n\n car --\n\n MAX\n\n Wipe your feet!\n\n Automatically, Gillis wipes his feet on an enormous\n\n shabby cocoanut mat.\n\n MAX\n\n You are not dressed properly.\n\n GILLIS\n\n Dressed for what?\n\n THE WOMAN'S VOICE\n\n Max! Have him come up, Max!\n\n MAX\n\n (Gesturing)\n\n Up the stairs!\n\n GILLIS\n\n Suppose you listen just for a\n\n minute -\n\n MAX\n\n Madame is waiting.\n\n GILLIS\n\n For me? Okay.\n\n Gillis enters.\n\n A-31 INT. NORMA DESMOND'S ENTRANCE HALL\n\n It is grandiose and grim. The whole place is one of\n\n those abortions of silent-picture days, with bowling\n\n alleys in the cellar and a built-in pipe organ, and\n\n beams imported from Italy, with California termites\n\n at work on them. Portieres are drawn before all the\n\n windows, and only thin slits or sunlight find their\n\n way in to fight the few electric bulbs which are always\n\n burning.\n\n Gillis starts up the curve of the black marble\n\n staircase. It has a wrought-iron rail and a worn\n\n velvet rope along the wall.\n\n MAX\n\n (From below)\n\n If you need help with the\n\n coffin call me.\n\n The oddity of the situation has caught Gillis'\n\n imagination. He climbs the stairs with a kind of\n\n morbid fascination. At the top he stops, undecided,\n\n then turns to the right and is stopped by\n\n WOMAN'S VOICE\n\n This way!\n\n Gillis swings around.\n\n Norma Desmond stands down the corridor next to a\n\n doorway from which emerges a flickering light. She\n\n is a little woman. There is a curious style, a\n\n great sense of high voltage about her. She is dress-\n\n ed in black house pyjamas and black high-heeled\n\n pumps. Around her throat there is a leopard-pat-\n\n terned scarf, and wound around her head a turban of\n\n the same material. Her skin is very pale, and she\n\n is wearing dark glasses.\n\n NORMA\n\n In here. I put him on my massage\n\n table in front of the fire. He\n\n always liked fires and poking at\n\n them with a stick.\n\n Gillis enters the SHOT and she leads him into\n\n A-32 NORMA DESMOND'S BEDROOM\n\n It is a huge, gloomy room hung in white brocade which\n\n has beconle dirty over the years and even slightly\n\n torn in a few places. There's a great, unmade gilded\n\n bed in the shape of a swan, from which the gold had\n\n begun to peel. There is a disorder of clothes and\n\n negligees and faded photographs of old-time stars\n\n about.\n\n In an imitation baroque fireplace some logs are burn-\n\n ing. On the massage table before it lies a small\n\n form shrouded under a Spanish shawl. At each end on\n\n a baroque pedestal stands a three-branched cande-\n\n labrum, the candles lighted.\n\n NORMA\n\n I've made up my mind we'll bury him in\n\n the garden. Any city laws against that?\n\n GILLIS\n\n I wouldn't know.\n\n NORMA\n\n I don't care anyway. I want the\n\n coffin to be white. And I want\n\n it specially lined with satin.\n\n White, or deep pink.\n\n She picks up the shawl to make up her mind about the\n\n color. From under the shawl flops down a dead arm.\n\n Gillis stares and recoils a little. It is like a\n\n child's arm, only black and hairy.\n\n NORMA\n\n Maybe red. bright flaming red.\n\n Gay. Let's make it gay.\n\n Gillis edges closer and glances down. Under the\n\n shawl he sees the sad, bearded face of a dead\n\n chimpanzee. Norma drops back the shawl.\n\n NORMA\n\n How much will it be? I warn you -\n\n don't give me a fancy price just\n\n because I'm rich.\n\n GILLIS\n\n Lady. you've got the wrong man.\n\n For the first time. Norma really looks at him\n\n through her dark glasses.\n\n GILLIS\n\n I had some trouble with my car.\n\n Flat tire. I pulled into your\n\n garage till I could get a spare.\n\n I thought this was an empty house.\n\n NORMA\n\n It is not. Get out.\n\n GILLIS\n\n I'm sorry, and I'm sorry you lost\n\n your friend, and I don't think red\n\n is the right color.\n\n NORMA\n\n Get out.\n\n GILLIS\n\n Sure. Wait a minute -- haven't\n\n I seen you -- ?\n\n NORMA\n\n Or shall I call my servant?\n\n GILLIS\n\n I know your face. You're Norma\n\n Desmond. You used to be in\n\n pictures. You used to be big.\n\n NORMA\n\n I am big. It's the pictures\n\n that got small.\n\n GILLIS\n\n I knew there was something\n\n wrong with them.\n\n NORMA\n\n They're dead. They're finished.\n\n There was a time when this busi-\n\n ness had the eyes of the whole\n\n wide world. But that wasn't good\n\n enough. Oh, nol They wanted the\n\n ears of the world, too. So they\n\n opened their big mouths, and out\n\n came talk, talk, talk...\n\n GILLIS\n\n That's where the popcorn business\n\n comes in. You buy yourself a bag\n\n and plug up your ears.\n\n NORMA\n\n Look at them in the front offices --\n\n the master minds! They took the\n\n idols and smashed them. The\n\n Fairbankses and the Chaplins and\n\n the Gilberts and the Valentinos.\n\n And who have they got now? Some\n\n nobodies -- a lot of pale little\n\n frogs croaking pish-poshl\n\n GILLIS\n\n Don't get sore at me. I'm not\n\n an executive. I'm just a writer.\n\n NORMA\n\n You are! Writing words, words!\n\n You've made a rope of words and\n\n strangled this businessl But there\n\n is a microphone right there to catch\n\n the last gurgles, and Technicolor\n\n to photograph the red, swollen tongue!\n\n GILLIS\n\n Ssh! You'll wake up that monkey.\n\n NORMA\n\n Get out!\n\n Gillis starts down the stairs.\n\n GILLIS\n\n Next time I'll bring my autograph\n\n album along, or maybe a hunk of\n\n cement and ask for your footprints.\n\n He is halfway down the staircase when he is\n\n stopped by\n\n NORMA\n\n Just a minute, you!\n\n GILLIS\n\n Yeah?\n\n NORMA\n\n You're a writer, you said.\n\n GILLIS\n\n Why?\n\n Norma starts down the stairs.\n\n NORMA\n\n Are you or aren't you?\n\n GILLIS\n\n I think that's what it says on my\n\n driver's license.\n\n NORMA\n\n And you have written pictures,\n\n haven't you?\n\n GILLIS\n\n Sure have. The last one I\n\n wrote was about cattle rustlers.\n\n Before they were through with it,\n\n the whole thing played on a\n\n torpedo boat.\n\n Norma has reached him at the bottom of the staircase.\n\n NORMA\n\n I want to ask you something.\n\n Come in here.\n\n She leads him into\n\n A-33 THE HUGE LIVING ROOM\n\n It is dark and damp and filled with black oak and\n\n red velvet furniture which looks like crappy props\n\n from the Mark of Zorro set. Along the main wall,\n\n a gigantic fireplace has been freezing for years.\n\n On the gold piano is a galaxy of photographs of\n\n Norma Desmond in her various roles. On one wall\n\n is a painting -- a California Gold Rush scene,\n\n Carthay Circle school. (We will learn later that\n\n it hides a motion picture screen.)\n\n One corner is filled with a large pipe organ, and\n\n as Norma and Gillis enter, there is a grizzly\n\n moaning sound. Gillis looks around.\n\n NORMA\n\n The wind gets in that blasted\n\n pipe organ. I ought to have\n\n it taken out.\n\n GILLIS\n\n Or teach it a better tune.\n\n Norma has led him to the card tables which stand\n\n side by side near a window. They are piled high\n\n with papers scrawled in a large, uncertain hand.\n\n NORMA\n\n How long is a movie script these\n\n days? I mean, how many pages?\n\n GILLIS\n\n Depends on what it is -- a Donald\n\n Duck or Joan or Arc.\n\n NORMA\n\n This is to be a very important\n\n picture. I have written it\n\n myself. Took me years.\n\n GILLIS\n\n (Looking at the piles\n\n of script)\n\n Looks like enough for six impor-\n\n tant pictures.\n\n NORMA\n\n It's the story or Salome. I\n\n think I'll have DeMille direct it.\n\n GILLIS\n\n Uh-huh.\n\n NORMA\n\n We've made a lot of pictures\n\n together.\n\n GILLIS\n\n And you'll play Salome?\n\n NORMA\n\n Who else ?\n\n GILLIS\n\n Only asking. I did't know\n\n you were planning a comeback.\n\n NORMA\n\n I hate that word. It is a return.\n\n A return to the millions of people\n\n who have never forgiven me for\n\n deserting the screen.\n\n GILLIS\n\n Fair enough.\n\n NORMA\n\n Salome -- what a woman! What a\n\n part! The Princess in love with\n\n a Holy man. She dances the Dance\n\n of the Seven Veils. He rejects\n\n her, so she demands his head on a\n\n golden tray, kissing his cold, dead\n\n lips.\n\n GILLIS\n\n They'll love it in Pomona.\n\n NORMA\n\n (Taking it straight)\n\n They will love it every place.\n\n (She reaches for a\n\n batch of pages from\n\n the heap)\n\n Read it. Read the scene just\n\n before she has him killed!\n\n GILLIS\n\n Right now? Never let another\n\n writer read your stuff. He\n\n may steal it.\n\n NORMA\n\n I am not afraid. Read it!\n\n NORMA (Cont'd)\n\n (Calling)\n\n Max! Max!\n\n (To Gillis)\n\n Sit down. Is there enough light?\n\n GILLIS\n\n I've got twenty-twenty vision.\n\n Max has entered.\n\n NORMA\n\n Bring something to drink.\n\n MAX\n\n Yes. Madame.\n\n He leaves. Norma turns to Gillis again.\n\n NORMA\n\n I said sit down.\n\n There is compulsion in her voice.\n\n Gillis looks at her GILLIS' VOICE\n\n and starts slowly Well. I had no pressing\n\n reading. engagement, and she'd men-\n\n tioned something to drink..\n\n Max comes in, wheeling Sometimes it's interesting\n\n a wicker tea wagon on to see just how bad bad\n\n which are two bottles o writing can be. This prom-\n\n f champagne and two ised to go the limit. I\n\n red Venetian glasses, wondered what a handwriting\n\n a box of zwieback and expert would make of that\n\n a jar of caviar. Norma childish scrawl of hers.\n\n sits on her feet. deep Max wheeled in some champagne\n\n in a chair, a gold ring and some caviar. Later, I\n\n on her forefinger with found out that Max was the\n\n a clip which holds a only other person in that\n\n cigarette. She gets up grim Sunset castle, and I\n\n and forces on Gillis found out a few other things\n\n another batch of script, about him... As for her, she\n\n goes back to her chair. sat coiled up like a watch\n\n spring, her cigarette\n\n clamped in a curious holder...\n\n I could sense her eyes on me\n\n from behind those dark\n\n glasses, defying me not to\n\n like what I read, or maybe\n\n begging me in her own proud\n\n way to like it. It meant\n\n so much to her...\n\n A-34 SHOT OF THE GILLIS' VOICE\n\n CEILING It sure was a cozy set-up.\n\n That bundle of raw nerves,and\n\n PAN DOWN to the moan- Max, and a dead monkey upstair\n\n ing organ. PAN OVER and the wind wheezing through\n\n TO THE ENTRANCE DOOR. that organ once in a while.\n\n Max opens it, and a Later on, just for comedy\n\n solemn-faced man in relief, the real guy arrived\n\n undertaker's clothes with a baby coffin. It was\n\n brings in a small all done with great dignity.\n\n white coffin. (Thru He must have been a very\n\n these shots the room important chimp. The great\n\n has been growing grandson of King Kong, maybe.\n\n duskier.)\n\n DISSOLVE TO:\n\n A-35 GILLIS It got to be eleven. I was\n\n feeling a little sick at my\n\n reading. The lamp stomach, what with that sweet\n\n beside him is now champagne and that tripe I'd\n\n really paying its been reading -- that silly\n\n way in the dark room. hodgepodge of melodramatic\n\n A lot of the manu- plots. However, by then I'd\n\n script pages are started concocting a little\n\n piled on the floor plot of my own...\n\n around his feet. A\n\n half-empty champagne\n\n glass stands on the\n\n arm of his chair.\n\n THE CAMERA SLOWLY DRAWS BACK to include Norma\n\n Desmond sitting in the dusk, just as she was before.\n\n Gillis puts down a batch of script. There is a\n\n little pause.\n\n NORMA\n\n (Impatiently)\n\n Well?\n\n GILLIS\n\n This is fascinating.\n\n NORMA\n\n Of course it is.\n\n GILLIS\n\n Maybe it's a little long and\n\n maybe there are some repetitions...\n\n but you're not a professional\n\n writer.\n\n NORMA\n\n I wrote that with my heart.\n\n GILLIS\n\n Sure you did. That's what makes\n\n it great. What it needs is a\n\n little more dialogue.\n\n NORMA\n\n What for? I can say anything I\n\n want with my eyes.\n\n GILLIS\n\n It certainly could use a pair of\n\n shears and a blue pencil.\n\n NORMA\n\n I will not have it butchered.\n\n GILLIS\n\n Of course not. But it ought to\n\n be organized. Just an editing\n\n job. You can find somebody.\n\n NORMA\n\n Who? I'd have to have somebody\n\n I can trust. When were you born --\n\n I mean, what sign of the zodiac?\n\n GILLIS\n\n I don't know.\n\n NORMA\n\n What month?\n\n GILLIS\n\n December twenty-first.\n\n NORMA\n\n Sagittarius. I like Sagittarians.\n\n You can trust them.\n\n GILLIS\n\n Thank you.\n\n NORMA\n\n I want you to do this work.\n\n GILLIS\n\n Me? I'm busy. Just finished\n\n one script. I'm due on another\n\n assignment.\n\n NORMA\n\n I don't care.\n\n GILLIS\n\n You know, I'm pretty expensive.\n\n I get five hundred a week.\n\n NORMA\n\n I wouldn't worry about money.\n\n I'll make it worth your while.\n\n GILLIS\n\n Maybe I'd better take the rest\n\n of the script home and read it -\n\n NORMA\n\n Oh no. I couldn't let it out\n\n of my house. You'll have to\n\n finish it here.\n\n GILLIS\n\n It's getting kind of late --\n\n NORMA\n\n Are you married, Mr. -- ?\n\n GILLIS\n\n The name is Gillis. I'm single.\n\n NORMA\n\n Where do you live?\n\n GILLIS\n\n Hollywood. The Alto Nido Apart-\n\n ments.\n\n NORMA\n\n There's something wrong with\n\n your car, you said.\n\n GILLIS\n\n There sure is.\n\n NORMA\n\n You can stay here.\n\n GILLIS\n\n I'll come early tomorrow.\n\n Norma takes off her glasses.\n\n NORMA\n\n Nonsense. There's room over the\n\n garage. Max will take you there...Max!\n\n THE CAMERA MOVES GILLIS' VOICE\n\n TOWARD NORMA'S FACE, She sure could say a lot of\n\n right up to her things with those pale eyes of\n\n eyes. hers. They'd been her trade\n\n mark. They'd made her the Num-\n\n ber One Vamp of another era. I\n\n remember a rather florid des-\n\n cription in an old fan magazine\n\n which said: "Her eyes are like\n\n two moonlit waterholes, where\n\n strange animals come to drink."\n\n DISSOLVE TO:\n\n A-36 SMALL STAIRCASE, LEAD- GILLIS'VOICE\n\n ING TO ROOM OVER GARAGE I felt kind of pleased with\n\n the way I'd handled the sit-\n\n Max, an electric light uation. I'd dropped the hook,\n\n bulb in his hand, is and she'd snapped at it. Now\n\n leading Gillis up. my car would be safe down\n\n Gillis carries a batch below, while I did a patch-\n\n of the manuscript. up job on the script. And\n\n there should be plenty of\n\n money in it...\n\n Max pushes open a door at the top of the stairs.\n\n MAX\n\n (Opening the door)\n\n I made your bed this afternoon.\n\n GILLIS\n\n Thanks.\n\n (On second thought)\n\n How did you know I was going to\n\n stay, this afternoon?\n\n Max doesn't answer. He walks across to the bed,\n\n screws a bulb in the open socket above it. The\n\n light goes on, revealing:\n\n A-37 A GABLED BEDROOM\n\n There are dirty windows on two sides, and dingy wall-\n\n paper on the cracked plaster walls. For furniture\n\n there is a neatly made bed, a table and a few chairs\n\n which might have been discarded from the main house.\n\n MAX\n\n This room has not been used for\n\n a long time.\n\n GILLIS\n\n It will never make house Beautiful.\n\n I guess it's O.K. for one night.\n\n Max gives him an enigmatic look.\n\n MAX\n\n (Pointing)\n\n There is the bathroom. I put in\n\n soap and a toothbrush.\n\n GILLIS\n\n Thanks.\n\n (He starts taking off\n\n his coat)\n\n Say, she's quite a character,\n\n that Norma Desmond.\n\n MAX\n\n She was the greatest. You wouldn't\n\n know. You are too young. In one\n\n week she got seventeen thousand fan\n\n letters. Men would bribe her mani-\n\n curist to get clippings from her\n\n fingernails. There was a Maharajah\n\n who came all the way from Hyderabad\n\n to get one of her stockings. Later,\n\n he strangled himself with it.\n\n GILLIS\n\n I sure turned into an interesting\n\n driveway.\n\n MAX\n\n You did, sir.\n\n GILLIS' VOICE\n\n He goes out. Gillis I pegged him as slightly\n\n looks after him, hangs cuckoo, too. A stroke maybe.\n\n his coat over a chair, Come to think of it, the\n\n walks over to the win- whole place seemed to have\n\n dow, pulls down the been stricken with a kind of\n\n rickety Venetian blind. creeping paralysis, out of\n\n As he does so, he looks beat with the rest of the\n\n down at: world, crumbling apart in\n\n slow motion ...\n\n A-38 THE TENNIS COURT OF GILLIS' VOICE\n\n THE DESMOND HOUSE There was a tennis court, or\n\n (MOONLIGHT) rather the ghost of a tennis\n\n court, with faded markings\n\n The cement surface is and sagging net ...\n\n cracked in many places,\n\n and weeds are growing\n\n high.\n\n A-39 GILLIS - IN THE WINDOW\n\n He looks away from the court to:\n\n A-40 THE DESMOND SWIMMING\n\n POOL\n\n GILLIS' VOICE\n\n There is no water in And of course she had a pool.\n\n it, and hunks of Who didn't then? Mabel Norm-\n\n mosaic which lines its and and John Gilbert must\n\n enormous basin are have swum in it ten thousand\n\n broken away. midnights ago, and Vilma Banky\n\n and Rod La Roque. It was\n\n empty now....or was it?\n\n A-41 GILLIS - IN THE WINDOW\n\n He stares down, his stomach slowly turning.\n\n A-42 THE SWIMMING POOL\n\n At the bottom of the basin a great rat is eating a\n\n decaying or,ange. From the inlet pipe crawl two\n\n other rats, who join battle with the first rat over\n\n the orange.\n\n A-43 GILLIS -IN THE WINDOW\n\n He starts away, but some- GILLIS' VOICE\n\n thing attracts his atten- There was something\n\n tion. He turns back and else going on below:\n\n looks down again. the last rites for\n\n that hairy old chimp,\n\n performed with the\n\n A-44 THE LAWN BELOW utmost seriousness --\n\n as if she were laying\n\n Norma Desmond and Max are to rest an only child.\n\n carrying the white coffin Was her life really\n\n towards a small grave as as empty as that?\n\n which has been dug in the\n\n dead turf. Norma carries\n\n one of the candelabra, all\n\n of its candles flickering\n\n in the wind. They reach\n\n the grave and lower the\n\n coffin into it. Then,\n\n Norma lighting his task\n\n with the candelabrum, Max\n\n takes a spade from the\n\n loose earth and starts\n\n filling in the grave.\n\n A-45 GILLIS - IN THE WINDOW\n\n He watches the scene be- GILLIS' VOICE\n\n low, then turns into the It was all very queer,\n\n room, goes to the door but queerer things\n\n to lock it. There is no were yet to come.\n\n key, and only a hole\n\n where the lock has been\n\n gouged out. Gillis moves\n\n a heavy overstuffed chair\n\n in front of the door, then\n\n walks towards the bed,\n\n throws himself on it,\n\n picking up some of the\n\n manuscript pages to read.\n\n DISSOLVE\n\n END OF SEQUENCE "A"\n\n SEQUENCE "B"\n\n DISSOLVE IN ON:\n\n B-1 LONG SHOT THE DESMOND\n\n HOUSE - (MORNING)\n\n The day is overcast. The SOUND: (Distant organ\n\n house is shrouded in low music - improvisations\n\n fog. on an odd, mournful\n\n theme - not too loud,\n\n continuing throughout\n\n B-2 THE TENNIS COURT, blurred the scene.)\n\n over with fog.\n\n B-3 THE EMPTY SWIMMING POOL\n\n Its dark outline even more\n\n melancholy under the misty\n\n blanket.\n\n B-4 THE ROOM OVER THE GARAGE\n\n Muted daylight seeps GILLIS' VOICE\n\n through the blinds. Gillis That night I'd had a\n\n lies on the bed, under a mixed-up dream. In it\n\n shabby quilt. The manu- was an organ grinder.\n\n script is beside him, some I couldn't see his\n\n of the pages scattered on face, but the organ\n\n the floor. He is just was all draped in\n\n opening his eyes. It takes black, and a chimp was\n\n him a moment to adjust him- dancing for pennies.\n\n self to the strange sur- When I opened my eyes,\n\n roundings. His eyes, wander- the music was still\n\n ing about the room. suddenly there... Where was\n\n stop, startled. He lifts I?\n\n himself on one elbow and\n\n stares at -\n\n B-5 THE DOOR\n\n The heavy chair he had set Oh yes, in that empty\n\n against it the night before room over her garage.\n\n has been pushed back. The Only it wasn't empty\n\n door is wide ajar. any more. Somebody\n\n had brought in all my\n\n belongings - my\n\n B-6 GILLIS books, my typewriter,\n\n my clothes...\n\n He jumps out of bed. He\n\n wears, shirt, trousers\n\n and socks. Suddenly he\n\n realizes that all his\n\n possessions have GILLIS' VOICE\n\n been brought in. In What was going on?\n\n the closet hang his\n\n shirts. His books and\n\n typewriter are neatly\n\n arranged on the table.\n\n His phonograph-radio\n\n combination is all\n\n installed. Gillis looks\n\n around startled, then\n\n sits down and starts\n\n putting on his moccasins\n\n hastily.\n\n DISSOLVE TO:\n\n B-7 A PAIR OF HANDS IN WHITE GLOVES, PLAYING THE ORGAN\n\n PULL BACK: They belong to Max von Mayerling. He\n\n is sitting erect, his bull neck taut as a wrestler's\n\n as he rights out somber chord after somber chord.\n\n He sits in a shaft of gray light coming from an open\n\n French window.\n\n Through the far archway, Gillis storms into the big\n\n room.\n\n GILLIS\n\n Hey, you -- Max -- whatever -your-\n\n name-is -- what are my things doing\n\n here?\n\n No answer.\n\n GILLIS\n\n I'm talking to you. My clothes\n\n and things are up in the room.\n\n MAX\n\n Naturally. I brought them myself.\n\n GILLIS\n\n (Furiously)\n\n Is that so!\n\n MAX\n\n Why are you so upset? Is there\n\n anything missing?\n\n GILLIS\n\n Who said you could? Who asked you to?\n\n Norma Desmond's shadow moves into the shaft of\n\n light.\n\n NORMA'S VOICE\n\n I did.\n\n Gillis looks around.\n\n On the couch by the fireplace reclines Norma Desmond,\n\n dressed in a negligee. She rises.\n\n NORMA\n\n I don't know why you should be\n\n so upset. Stop that playing,\n\n Max.\n\n (To Gillis again)\n\n It seemed like a good idea --\n\n if we are to work together.\n\n GILLIS\n\n Look, I'm supposed to fix up\n\n your script. There's nothing\n\n in the deal about my staying\n\n here.\n\n NORMA\n\n You'll like it here.\n\n GILLIS\n\n Thanks for the invitation, but\n\n I have my own apartment.\n\n NORMA\n\n You can't work in an apartment\n\n where you owe three months' rent.\n\n GILLIS\n\n I'll take care of that.\n\n NORMA\n\n It's all taken care of. It's\n\n all paid for.\n\n GILLIS\n\n I'm used to paying my own bills.\n\n NORMA\n\n You proud boy, why didn't you tell\n\n me you were having difficulties.\n\n GILLIS\n\n Okay. We'll deduct it from my\n\n salary.\n\n NORMA\n\n Now, now, don't let's be small\n\n about such matters. We won't\n\n keep books.\n\n (To Max)\n\n Go on, unpack Mr. Gillis' things.\n\n GILLIS\n\n Unpack nothing. I didn't say\n\n I was staying.\n\n NORMA\n\n (Her glasses off again)\n\n Suppose you make up your mind.\n\n Do you want this job or don't you?\n\n DISSOLVE TO:\n\n B-8 BIG ROOM, NORMA DESMOND'S\n\n HOUSE - (DAY) GILLIS' VOICE\n\n Gillis sits at an impro- So I let him unpack my\n\n vised table, his typewriter things. I wanted the\n\n in front of him, working dough, and I wanted to\n\n hard at the manuscript. get out of there as\n\n Pencils, shears and a quickly as possible.\n\n paste-pot at hand. I thought if I really\n\n got going I could toss\n\n Facing him at some dis- it off in a couple or\n\n tance sits Norma,dressed weeks. But it wasn't\n\n in another version of her so simple, getting some\n\n favorite lounging pajamas, coherence into that wild,\n\n the cigaette contraption scrambled melodrama\n\n on her finger. She is she'd concocted. What\n\n autographing large photo- made it tougher was that\n\n graphs of herself and put- she was around all the\n\n ting them in envelopes. time -- hovering over\n\n me, afraid I'd do injury\n\n to that precious brain-\n\n child of hers.\n\n Gillis takes two or three pages from Norma's hand-\n\n written script, crosses them out and puts them to\n\n one side.\n\n Norma rises, crosses towards Gillis, looks over his\n\n shoulder.\n\n NORMA\n\n What's that?\n\n GILLIS\n\n Just a scene I cut out.\n\n NORMA\n\n What scene?\n\n GILLIS\n\n The one where you go to the slave\n\n market. You can cut right to the\n\n scene where John the Baptist -\n\n NORMA\n\n Cut away from me?\n\n GILLIS\n\n Honestly, it's a little old hat.\n\n They don't want that any more.\n\n NORMA\n\n They don't? Then why do they still\n\n write me fan letters every day.\n\n Why do they beg me for my photo-\n\n graphs? Because they want to see\n\n me, me, me! Norma Desmond.\n\n GILLIS\n\n (Resigned)\n\n Okay.\n\n He pulls the page from his typewriter. As he does\n\n so he glances over towards Norma.\n\n GILLIS' VOICE\n\n On the table in front I didn't argue with her.\n\n of her are the photo- You don't yell at a\n\n graphs which she is sign- sleepwalker-- he may fall\n\n ing. On the long table and break his neck.That's\n\n in the living room is a it -- she was still\n\n gallery of photographs sleepwalking along the\n\n in various frames -- all giddy heights of a lost\n\n Norma Desmond. On the career --plain crazy\n\n piano more photographs. when it came to that one\n\n Above the piano an oil subject: her celluloid\n\n portrait of her. On the self, the great Norma\n\n highboy beside him still Desmond. How could She\n\n more photographs. breathe in that house,\n\n so crowded with Norma\n\n DISSOLVE TO: Desmonds? More Norma\n\n Desmond and still more\n\n Norma Desmond.\n\n B-9 THE BIG ROOM - (NIGHT)\n\n GILLIS' VOICE\n\n Shooting towards the big It wasn't all work - of\n\n Gold Rush painting. Max, course. Two or three\n\n white gloves and all, times a week Max would\n\n steps into the shot, shoves haul up that enormous oil\n\n the painting up towards painting that had been\n\n the ceiling,revealing a presented to her by some\n\n motion picture screen. Nevada Chamber of Com-\n\n Max exits. merce, and we'd see a\n\n movie,right in her\n\n living room.\n\n B-1O NORMA AND GILLIS\n\n GILLIS' VOICE\n\n They sit on a couch,facing "So much nicer than going\n\n the screen. On a table in out," she'd say. The\n\n front of them are champagne, plain fact was that she\n\n cigarettes and coffee. was afraid of that world\n\n Above their heads are the outside. Afraid it\n\n typical openings for a pro- would remind her that\n\n jector. The lights go off. time had passed.\n\n From the opening above\n\n their heads shoots the wide\n\n beam of light.\n\n B-11 MAX, IN THE PROJECTION They were silent movies,\n\n BOOTH BEHIND THE ROOM and Max would run the\n\n projection machine, which\n\n The light of the machine was just as well -- it\n\n flickering over his face, kept him from giving us\n\n which is frozen, a somber an accompaniment on\n\n enigma. that wheezing organ.\n\n B-12 NORMA AND GILLIS\n\n She'd sit very close to\n\n watching the screen. me, and she'd smell of\n\n Gillis looks down and sees tuberoses, which is not\n\n that Norma's hand is clasp- my favorite perfume, not\n\n ing his ann tight. He by a long shot. Sometines\n\n doesn't like it much but as we watched, she'd c\n\n he can't do anything about lutch my arm or my hand\n\n it. However. when she for forgetting she was my\n\n a second lets go his arm employer becoming just a\n\n to pick up a glass of fan, excited about that\n\n champagne, he gently with- actress up there on the\n\n draws his arm, leans away screen....I guess I don't\n\n from her and crosses his have to tell you who the\n\n arms to discourage any star was. They were\n\n resumption of her approach. always her pictures --\n\n Norma puts the glass down that's all she wanted\n\n doesn't find his arn, but to see.\n\n is not aware of any signifi-\n\n cance in his maneuver. They\n\n both watch the screen.\n\n B-13 THE OTHER END OF THE BIG ROOM. WITH THE SCREEN\n\n On it flickers a famous scene from one of Norma's old\n\n silent pictures. It is not to be a funny scene. It\n\n is old-fashioned, but shows her incredible beauty\n\n and the screen presence which made her the great star\n\n of her day.\n\n B-14 NORMA AND GILLIS ON THE COUCH\n\n NORMA\n\n Still wonderful, isn't it? And\n\n no dialogue. We didn't need\n\n dialogue. We had faces. There\n\n just aren't any faces like that\n\n any more. Well, maybe one --\n\n Garbo.\n\n In a sudden flareup she jumps to her feet and stands\n\n in the flickering beam of light.\n\n NORMA\n\n Those idiot producers! Those\n\n imbeciles! Haven't they got any\n\n eyes? Have they forgotten what\n\n a star looks like? I'll show them.\n\n I'll be up there again. So help me!\n\n DISSOLVE TO:\n\n B-15 THE BIG ROOM - (NIGHT)\n\n It is apparently empty. GILLIS' VOICE\n\n The elaborate lamps Sometimes there'd be a\n\n make pools of light. little bridge game in the\n\n house, at a twentieth-of-\n\n THE CAMERA PULLS BACK a cent a point. I'd get\n\n AND PANS to reveal a half her winnings. Once\n\n card table around they ran up to seventy\n\n which sit Norma and cents, which was about\n\n three friends - three the only cash money I\n\n actors of her period. ever got. The others\n\n They sit erect and play around the table would\n\n with grim seriousness. be actor friends - dim\n\n figures you may still\n\n Beside Norma sits remember from the silent\n\n Gillis, kibitzing on a days. I used to think of\n\n game which bores him them as her Wax Works.\n\n extremely. An ashtray\n\n on the card table is\n\n full and Norma holds\n\n it out for Gillis to\n\n take away. He crosses\n\n the room to the fire-\n\n place. but his eyes\n\n fall on the entrance\n\n door and he stops.\n\n B-16 THE ENTRANCE HALL - (FROM GILLIS' POINT OF VIEW)\n\n Max stands in the open door. Outside are the two\n\n men who came to the apartment for Gillis' car.\n\n B-17 GILLIS\n\n He steps back so that he cannot be seen from the\n\n door. A second later Max appears, looking for him.\n\n MAX\n\n (Quietly)\n\n Some men are here. They asked\n\n for you.\n\n GILLIS\n\n I'm not here.\n\n MAX\n\n That's what I told them.\n\n GILLIS\n\n Good.\n\n MAX\n\n They found your car in the\n\n garage. They are going to tow\n\n it away.\n\n Gillis doesn't know what to do. From offstage\n\n comes:\n\n NORMA'S VOICE\n\n The ashtray, Joe dear! Can we\n\n have the ashtray?\n\n Gillis dumps the cigarette butts into the cold fire-\n\n place, crosses to the bridge table, puts the\n\n ashtray down, leans over and speaks into Norma's ear.\n\n GILLIS\n\n I want to talk to you for a\n\n minute.\n\n NORMA\n\n Not now, my dear. I'm playing\n\n three no trump.\n\n GILLIS\n\n They've come for my car.\n\n NORMA\n\n Please. Now I've forgotten how\n\n many spades are out.\n\n GILLIS\n\n I need some money right now.\n\n NORMA\n\n Can't you wait till I'm dummy?\n\n 3.22.49 GILLIS\n\n No.\n\n NORMA\n\n (Angry by now)\n\n Please!\n\n Gillis stands frustrated, hideously embarrassed\n\n by the stares of the waxworks. He turns away\n\n and hurries to the door.\n\n B-18 ENTRANCE DOOR TO THE HOUSE\n\n It is half open. Gillis comes into the shot\n\n and, taking cover, looks out.\n\n B-19 COURTYARD (FROM GILLIS' ANGLE)\n\n The men from the finance company are cranking up\n\n the car. Max stands watching silently. When they\n\n finish the cranking job, the men climb into the\n\n front seat of the truck.\n\n B-2O GILLIS - AT THE DOOR\n\n Over the shot the SOUND of the truck being started\n\n and the cars moving away. Gillis moves out into\n\n the courtyard and stands staring after the car.\n\n From the house comes Norma.\n\n NORMA\n\n Now what is it? Where's the\n\n fire?\n\n GILLIS\n\n I've lost my car.\n\n NORMA\n\n Oh...and I thought it was a\n\n matter of life and death.\n\n GILLIS\n\n It is to me. That's why I came\n\n to this house. That's why I took\n\n this job -- ghost writing!\n\n NORMA\n\n Now you're being silly. We don't\n\n need two cars. We have a car. And\n\n not one of thuse cheap new things\n\n made of chromium and spit. An\n\n Isotta-Fraschini. Have you ever\n\n heard of Isotta-Fraschinis? All\n\n hand-made. Cost me twenty-eight\n\n thousand dollars.\n\n THE CAMERA HAS PANNED over to the garage and FOCUSES\n\n on the dirty Isotta-Fraschini on its blocks.\n\n DISSOLVE TO:\n\n B-21 NORMA'S ISOTTA-FRASCHINI\n\n DRIVING IN THE HILLS\n\n ABOVE SUNSET (DAY)\n\n Max is at the wheel, GILLIS' VOICE\n\n dressed as usual except So Max got that old bus\n\n for a chauffeurfs cap. down off its blocks and\n\n polished it up. She'd\n\n take me for rides in the\n\n B-22 INSIDE THE CAR hills above Sunset.\n\n Gillis sits beside Norma, The whole thing was up-\n\n who is wearing a smart holstered in leopard\n\n tailleur and her eternal skin, and had one of\n\n sun glasses. Gillis those car phones, all\n\n wears his sport jacket- gold-plated.\n\n flannel trousers-moccasin\n\n combinatIon.\n\n He sits uncomfortably. Norma is studying him.\n\n NORMA\n\n That's a dreadful shirt you're\n\n wearing.\n\n GILLIS\n\n What's wrong with It?\n\n NORMA\n\n Nothing, if you work in a fill-\n\n ing station. And I'm getting\n\n rather bored with that sport\n\n jacket, and those same baggy\n\n pants.\n\n (She picks up\n\n the car phone)\n\n Max, what's a good men's shop\n\n in town? The very best...\n\n Well, go there !\n\n GILLIS\n\n I don't need any clothes, and\n\n I certainly don't want you buy-\n\n ing them for --\n\n NORMA\n\n Why begrudge me a little fun?\n\n I just want you to look nice,\n\n my stray little boy.\n\n By this time Max has made a U-turn.\n\n QUICK DISSOLVE TO:\n\n B-23 INT. MEN'S DEPARTMENT, AN ELEGANT WILSHIRE STORE\n\n Gillis stands in front of a full-length triple mirror,\n\n surrounded by a couple of salesmen and the tailor, who\n\n is busily working out alterations.\n\n Gillis wears a double-breasted gray flannel coat with\n\n chalk stripes. His trousers belong to another suit\n\n of glen plaid. Norma is running the show.\n\n NORMA\n\n There's nothing like gray flannel\n\n with a chalk stripe.\n\n (she points at\n\n the trousers)\n\n This one single-breasted, of course.\n\n (to another salesman)\n\n Now we need a topcoat. Let's see\n\n what you have in camel's hair.\n\n The salesman leaves.\n\n NORMA\n\n How about some evening clothes?\n\n GILLIS\n\n I don't need a tuxedo.\n\n NORMA\n\n Of course you do. A tuxedo and\n\n tails.\n\n GILLIS\n\n Tails. That's ridiculous.\n\n NORMA\n\n You'll need them for parties.\n\n You'll need them for New Year's\n\n Eve.\n\n (to a salesman)\n\n Where are your evening clothes?\n\n SALESMAN\n\n This way, Madame.\n\n He leads her off. The other salesman arrives with a\n\n selection of topcoats.\n\n SALESMAN\n\n Here are some camel hairs, but\n\n I'd like you just to feel this\n\n one. It's Vicuna. Of course,\n\n it's a little more expensive.\n\n GILLIS\n\n A camel's hair will do.\n\n SALESMAN\n\n (With an insulting\n\n inflection)\n\n As long as the lady is paying\n\n for it, why not take the Vicuna?\n\n DISSOLVE:\n\n END OF SEQUENCE "B"\n\n SEQUENCE "C"\n\n DISSOLVE IN:\n\n C-1 LONG SHOT DESMOND HOUSE\n\n A day in December. Rain.\n\n QUICK DISSOLVE TO:\n\n C-2 INT. ROOM OVER GARAGE\n\n Water is drizzling from GILLIS' VOICE\n\n two or three spots in the The last week in December\n\n ceiling into pans and the rains came -- a great\n\n bowls set to catch it, big package of rain.\n\n one bowl right on the Over-sized, like every-\n\n bed. The room is almost thing else in California.\n\n emptied of Gillis' be-\n\n longings by now. Max It came right through\n\n is carrying out a hand- the old roof of my room\n\n full of new suits on above the garage. She\n\n hangers. He has a had Max move me to the\n\n dressing gown over his main house. I didn't\n\n shoulder. Gillis holds much like the idea -- the\n\n a stack of shirts, his only time I could have\n\n typewriter, and some to myself was in that\n\n manuscript. He surveys room -- but it was better\n\n the room for the last than sleeping in a rain-\n\n time, to see whether coat and galoshes.\n\n he's forgotten any-\n\n thing. He has. He\n\n puts down the typewriter\n\n and picks up from under\n\n the bed a pair of very\n\n smart red leather bedroom\n\n slippers. He tucks them\n\n under his arm, picks up\n\n the typewriter and leaves.\n\n QUICK DISSOLVE TO:\n\n C-3 A BEDROOM IN TIiE MAIN HOUSE\n\n It is obviously a man's room -- heavy Spanish\n\n furniture -- one wall nothing but a closet with\n\n shelves and drawers for shirts and shoes. Max is\n\n hanging up the suits. Gillis throws the shirts on\n\n a big chair, tosses the slippers at the foot of the\n\n bed, places the typewriter and manuscript on a desk\n\n at the window.\n\n GILLIS\n\n Whose room was this?\n\n MAX\n\n It was the room of the husband.\n\n Or of the husbands, I should say.\n\n Madame has been married three\n\n times.\n\n Slightly embarrassed, Gillis picks up his toilet\n\n kit with razor, toothbrushes, soap, etc., and starts\n\n towards the bathroom, pausing en route at a rain-\n\n splattered window.\n\n GILLIS\n\n I guess this is the one you\n\n can see Catalina from. Only\n\n this isn't the day.\n\n He proceeds towards the half-opened door leading\n\n to the bathroom. Something strikes his attention\n\n and he stops. As in the door to the room above\n\n the garage, this lock, too, has been gouged out.\n\n GILLIS\n\n Hey, what's this with the\n\n door? There isn't any lock.\n\n MAX\n\n There are no locks anywhere\n\n in this house.\n\n He points to the entrance door of the room, and to\n\n another door.\n\n GILLIS\n\n How come?\n\n MAX\n\n The doctor suggested it.\n\n GILLIS\n\n What doctor?\n\n MAX\n\n Madame's doctor. She has moments\n\n of melancholy. There have been\n\n some suicide attempts.\n\n GILLIS\n\n Uh-huh?\n\n MAX\n\n We have to be very careful. No\n\n sleeping pills, no razor blades.\n\n We shut off the gas in her bed-\n\n room.\n\n GILLIS\n\n Why? Her career? She got enough\n\n out of it. She's not forgotten.\n\n She still gets those fan letters.\n\n MAX\n\n I wouldn't look too closely at the\n\n postmarks.\n\n GILLIS\n\n You send them. Is that it, Max?\n\n MAX\n\n I'd better press your evening\n\n clothes, sir. You have not for-\n\n gotten Madame's New Year's party.\n\n GILLIS\n\n No, I haven't. I suppose all\n\n the waxworks are coming?\n\n MAX\n\n I don't know, sir. Madame made\n\n the arrangements.\n\n Max leaves. Gillis comes out of the bathroom, picks\n\n up his shirts, goes over to a closet, opens it. As\n\n he does so one of the doors without a lock swings\n\n slightly open. Gillis looks through the half-open\n\n door and sees.\n\n C-4 NORMA DESMOND'S ROOM\n\n It is empty. The rainy GILLIS' VOICE\n\n day does nothing to There it was again - that\n\n help its gloom. room of hers, all satin and\n\n ruffles, and that bed like\n\n a gilded rowboat. The per-\n\n fect setting for a silent\n\n movie queen. Poor devil,\n\n still waving proudly to a\n\n parade which had long since\n\n passed her by.\n\n He pushes the door shut\n\n and walks back into the\n\n room.\n\n DISSOLVE TO:\n\n C-5 STAIRCASE OF DESMOND\n\n HOUSE (NIGHT)\n\n Gillis is coming down the GILLIS' VOICE\n\n stairs in his tailcoat It was at her New Year's\n\n adjusting the handkerchief party that I found out\n\n in his pocket. He obviously how she felt about me.\n\n feels a little uneasy in Maybe I'd been an idiot\n\n this outfit. From below not to have sensed it\n\n comes a tango of the Twen- was coming - that sad,\n\n ties. played by a small embarrassing revelation.\n\n orchestra. Gillis stops\n\n in the archway leading to\n\n the big room and looks\n\n around.\n\n C-6 THE BIG ROOM has been deco-\n\n rated for the occasion with\n\n laurel garlands. Dozens of\n\n candles in all the sconces\n\n and candelabra are ablaze.\n\n Their flickering flames are\n\n reflected in the waxed sur=\n\n face of the tile floor.\n\n There is a buffet, with\n\n buckets of champagne and\n\n caviar on ice. In one corner\n\n on a little platform banked\n\n with palms. a four-piece\n\n orchestra is playing.\n\n At the buffet are Max and Norma. She is drinking\n\n a glass of champagne. She is wearing a diamonte\n\n evening dress. very high style. with long black\n\n gloves and a headdress of paradise feathers. Her\n\n eyes fall on Gillis. She puts down the glass of\n\n champagne. picks up a gardenia boutonniere and\n\n moves toward him.\n\n NORMA\n\n Joe, you look absolutely\n\n divine. Turn around!\n\n GILLIS\n\n (Embarrassed}\n\n Please.\n\n NORMA\n\n Come on!\n\n Gillis makes a slow 36O-degree turn.\n\n NORMA\n\n Perfect. Wonderful shoulders.\n\n And I love that line.\n\n She indicates the V from his shoulders to his hips.\n\n GILLIS\n\n All padding. Don't let it fool\n\n you.\n\n NORMA\n\n Come here!\n\n She puts the gardenia on his lapel.\n\n GILLIS\n\n You know, to me dressing up\n\n was always just putting on\n\n my dark blue suit.\n\n NORMA\n\n I don't like those studs they've\n\n sent. I want you to have pearls.\n\n Nice big pearls.\n\n GILLIS\n\n Now, I'm not going to wear ear-\n\n rings, I can tell you that.\n\n NORMA\n\n Cute. Let's have some drinks.\n\n She leads him over to the buffet.\n\n GILLIS\n\n Shouldn't we wait for the others?\n\n NORMA\n\n (Pointing at the floor)\n\n Careful, it's slippery. I\n\n had it waxed.\n\n They reach the buffet. Max is ready with two\n\n glasses of champagne. Norma hands Gillis a glass.\n\n NORMA\n\n Here's to us.\n\n They drink.\n\n NORMA\n\n You know, this floor used to\n\n be wood but I had it changed.\n\n Valentino said there is nothing\n\n like tiles for a tango.\n\n She opens her arms.\n\n GILLIS\n\n Not on the same floor with\n\n Valentino!\n\n NORMA\n\n Just follow me.\n\n They start to tango. After a moment --\n\n NORMA\n\n Don't bend back like that.\n\n GILLIS\n\n It's those feathers. They tickle.\n\n Norma pulls the paradise feathers from her hair\n\n and tosses them away.\n\n C-7 THE ORCHESTRA\n\n As they play the tango, the musicians eye the danc-\n\n ing couple, take in the situation, exchange glances\n\n and turn away with professional discretion.\n\n C-8 NORMA AND GILLIS, TANGOING\n\n Gillis glances at his wrist watch.\n\n GILLIS\n\n It's a quarter past ten. What\n\n time are they supposed to get\n\n here?\n\n NORMA\n\n Who?\n\n GILLIS\n\n The other guests?\n\n NORMA\n\n There are no other guests. We\n\n don't want to share this night\n\n with other people. This is for\n\n you and me.\n\n GILLIS\n\n I understand some rich guy bought\n\n up all the tickets for a perfor-\n\n mance at the Metropolitan and sat\n\n there listening to La Traviata,\n\n all by himself. He was afraid of\n\n catching cold.\n\n NORMA\n\n Hold me tighter.\n\n GILLIS\n\n Come midnight, how about blind-\n\n folding the orchestra and smash-\n\n ing champagne glasses on Max's\n\n head?\n\n NORMA\n\n You think this is all very funny.\n\n GILLIS\n\n A little.\n\n NORMA\n\n Is it funny that I'm in love\n\n with you?\n\n GILLIS\n\n What's that?\n\n NORMA\n\n I'm in love with you. Don't you\n\n know that? I've been in love\n\n with you all along.\n\n They dance on. Gillis is acutely embarrassed.\n\n THE CAMERA SLOWLY PULLS BACK, PANS past the faces\n\n of the musicians, who play on with a rather overe-\n\n mphasized lack of interest. Finally it winds up\n\n on Max, behind the buffet. He stands watching Gillis,\n\n a faint trace of pity in his eyes.\n\n DISSOLVE TO:\n\n C-9 NORMA'S FINGER, WITH THE\n\n CIGARETTE GADGET, as she GILLIS' VOICE\n\n inserts a cigarette. I'm sure a lot of you will\n\n laugh about this. Ridicu-\n\n lous situation, wasn't it?\n\n -- a woman almost twice my\n\n age ... It got to be about\n\n a quarter of eleven. I\n\n felt caught, like a cig-\n\n arette in the prongs of\n\n that contraption on her\n\n finger.\n\n PULL BACK TO:\n\n NORMA AND GILLIS sitting on a couch in front of the\n\n cavernous fireplace. Norma holds out her cigarette\n\n to Gillis, who lights it.\n\n NORMA.\n\n What a wonderful next year it's\n\n going to be. What fun we're going\n\n to have. I'II fill the pool for\n\n you. Or I'll open my house in\n\n Malibu, and you can have the whole\n\n ocean. Or I'll buy you a boat\n\n and we'll sail to Hawaii.\n\n GILLIS\n\n Stop it. You aren't going to buy\n\n me anything more.\n\n NORMA\n\n Don't be silly.\n\n (She reaches under a\n\n pillow of the couch\n\n and brings out a\n\n leather box)\n\n Here. I was going to give it to\n\n you at midniglht.\n\n Gillis opens the box. It contains a matched gold\n\n cigarette case and lighter.\n\n NORMA\n\n Read what's inside.\n\n Gillis snaps open the case. Engraved inside the\n\n cover is: TO JOE FROM NORMA, and two bars of\n\n music.\n\n GILLIS\n\n What are the notes?\n\n NORMA\n\n "Mad about the boy."\n\n GILLIS\n\n Norma, I can't take it. You've\n\n bought me enough.\n\n NORMA\n\n Shut up. I'm rich. I'm richer\n\n than all this new Hollywood trash.\n\n I've got a million dollars.\n\n GILLIS\n\n Keep it.\n\n NORMA\n\n I own three blocks downtown.\n\n I have oil in Bakersfield --\n\n pumping, pumping, pumping.\n\n What's it for but to buy us\n\n anything we want.\n\n GILLIS\n\n Cut out that us business.\n\n He rises.\n\n NORMA\n\n What's the matter with you?\n\n GILLIS\n\n What right do you have to take\n\n me for granted?\n\n NORMA\n\n What right? Do you want me to\n\n tell you?\n\n GILLIS\n\n Has it ever occurred that I may\n\n have a life of my own? That there\n\n may be some girl I'm crazy about?\n\n NORMA\n\n Who? Some car hop, or a dress\n\n extra?\n\n GILLIS\n\n Why not? What I'm trying to say\n\n is that I'm all wrong for you.\n\n You want a Valentino -- somebody\n\n with polo ponies -- a big shot --\n\n NORMA\n\n (Getting up slowly)\n\n What you're trying to say is\n\n that you don't want me to love\n\n you. Is that it?\n\n Gillis doesn't answer. Norma slaps his face and\n\n rushes from the room and upstairs.\n\n Gillis stands paralyzed, the slap burning his cheek.\n\n C-1O THE TOP OF THE STAIRCASE AND CORRIDOR\n\n Norma rushes up the last few steps, down the corridor\n\n and into her bedroom, banging the door. MOVE THE\n\n CAMERA toward the closed door, centering on the\n\n gouged-out lock.\n\n C-11 GILLIS, IN THE BIG ROOM\n\n He still stands motionless. He glances around fur-\n\n tively, to see if his humiliation has been observed.\n\n C-12 THE ORCHESTRA\n\n The musicians are playing away. They have turned\n\n their eyes away from Gillis rather too ostentatious-\n\n ly for comfort.\n\n C-13 GILLIS\n\n His eyes move over toward\n\n C-14 MAX\n\n He is subtler than the musicians. He appears very\n\n busy at the buffet, putting empty bottles and used\n\n glasses on a tray. He walks across the room with\n\n them.\n\n C-15 GILLIS\n\n He starts slowly out. As he does so his long gold\n\n key chain catches on a carved ornament of the sofa\n\n and holds him for a second of additional embarrass-\n\n ment. He yanks it loose and walks with as much\n\n nonchalance as he can muster to\n\n C-16 THE HALL\n\n Crossing towards the coat closet, Gillis throws a\n\n look upstairs. Then he pulls the Vicuna coat from\n\n its hangar and slips into it as he crosses to the\n\n entrance door. He opens the door on the darkness\n\n of the courtyard.\n\n C-17 EXT. DESMOND HOUSE \n\n (NIGHT - RAIN)\n\n Gillis shuts the door. GILLIS'VOICE\n\n He takes a few steps I didn't know where I was\n\n forward, then stands going. I just had to get\n\n for a while breathing out of there. I had to be\n\n deep. The rain is with people my own age. I\n\n balm to that cheek had to hear somebody laugh\n\n where the slap still a again. I thought of Artie\n\n burns. He walks for- Green. There was bound to\n\n ward with a great be a New Year's shindig\n\n sense of relief. going on in his apartment\n\n down on Las Palmas -- the\n\n hock shop set -- not a job\n\n C-18 DRIVEWAY LEADING TO in the room. but lots of\n\n \tfun on the cuff.\n\n Gillis walks to the\n\n street, which is dark\n\n and empty. He starts\n\n down Sunset in an\n\n Easterly direction.\n\n A car passes. He\n\n tries to thumb a\n\n ride, without success.\n\n However, the second\n\n car, a florist's\n\n delivery wagon, stops.\n\n Gillis jumps in and the\n\n car drives off.\n\n DISSOLVE TO:\n\n C-19 ARTIE GREEN'S APARTMENT\n\n It is the most modest one-room affair, jam packed\n\n with young people flowing over into the miniature\n\n bathroom and the microscopic kitchenette. The only\n\n drink being served is punch from a pressed-glass\n\n bowl -- but everybody is having a hell of a time.\n\n Most of the men are in slacks and sweaters, and only\n\n a few of the girls in something that vaguely suggests\n\n party dress.\n\n Abe Burroughs sits at a small, guest-festooned piano\n\n and sings Tokio Rose. By the door, a group of young\n\n men and girls respond to the song by sing1ng Rinso\n\n White or Dentyne Chewing Gum or something similar,\n\n in the manner of a Bach choral. Artie Green, a dark\n\n haired, pleasant-looking guy in his late twenties,\n\n is conducting with the ladle from the punch bowl.\n\n The door behind some of the singers is pushed open,\n\n jostling them out of their places. In comes Gillis,\n\n his hair and face wet, the collar of his Vicuna coat\n\n turned up. Artie stops conducting, but the commer-\n\n cial goes right on.\n\n ARTIE\n\n Well, what do you know ! Joe\n\n Gillis !\n\n GILLIS\n\n Hi, Artie.\n\n ARTIE\n\n Where have you been keeping that\n\n gorgeous face of yours?\n\n GILLIS\n\n In a deep freeze.\n\n ARTIE\n\n I almost reported you to the Bureau\n\n of Missing Persons.\n\n (To the company)\n\n Fans, you all know Joe Gillis, the\n\n well-known screen writer, opium\n\n smuggler and Black Dahlia suspect.\n\n Gillis greets some of the kids by name as he and\n\n Artie push their way into the room.\n\n ARTIE\n\n Give me your coat.\n\n GILLIS\n\n Let it ride for a while.\n\n ARTIE\n\n You're going to stay, aren't you?\n\n GILLIS\n\n That was the general idea.\n\n ARTIE\n\n Come on.\n\n Artie starts peeling the coat off Gillis. Its\n\n texture takes his breath away.\n\n ARTIE\n\n What is this - mink?\n\n He has taken the coat. He looks at Gillis standing\n\n there in tails.\n\n ARTIE\n\n Judas E. Priest, who did you\n\n borrow that from? Adolphe\n\n Menjou?\n\n GILLIS\n\n Close, but no cigar.\n\n Gillis stands embarrassed While Artie rolls up the\n\n Vicuna coat and tucks it above the books on a book-\n\n shelf.\n\n ARTIE\n\n Say, you're not really smuggling\n\n opium these days, are you?\n\n GILLIS\n\n Where's the bar?\n\n The two make their way toward the punch bowl. It's\n\n a little like running the gauntlet for Gillis. There\n\n are whistles and 'stares of astonishlnent at his tails.\n\n When they reach the punch bowl, Artie picks up a\n\n half-filled glass and fills it.\n\n GILLIS\n\n Good party.\n\n ARTIE\n\n The greatest. They call me the Elsa\n\n Maxwell of the assistant directors.\n\n (To some guests who are\n\n dipping their empty cups\n\n into the punch bowl)\n\n Hey, easy on the punch bowl. Budget\n\n only calls for three drinks per extra.\n\n Fake the rest.\n\n GILLIS\n\n Listen, Artie, can I stick around\n\n here for a while?\n\n ARTIE\n\n Sure, this'll go on all night.\n\n GILLIS\n\n I mean, could you put me up for\n\n a couple of weeks?\n\n ARTIE\n\n It just so happens we have a\n\n vacancy on the couch.\n\n GILLIS\n\n I'll take it.\n\n ARTIE\n\n I'll have the bell-hop take care\n\n of your luggage.\n\n He runs his finger across the decollete back of a\n\n girl standing in a group next them.\n\n ARTIE\n\n Just register here.\n\n The girl turns around. She is Betty Schaefer.\n\n BETTY\n\n Hello, Mr. Gillis.\n\n ARTIE\n\n You know each other?\n\n Gillis looks at her a little puzzled.\n\n BETTY\n\n Let me help you. Betty Schaeter,\n\n Sheldrake's office.\n\n GILLIS\n\n Sure. Bases Loaded.\n\n ARTIE\n\n Wait a minute. This is the woman\n\n I love. What's going on? Who\n\n was loaded?\n\n GILLIS\n\n Don't worry. She's just a fan\n\n for my literary output.\n\n BETTY\n\n (to Artie)\n\n Hurt feelings department.\n\n GILLIS\n\n About that luggage. Where's\n\n the phone?\n\n ARTIE\n\n Over by the Rainbow Room.\n\n Gillis squeezes his way through groups of people\n\n to the telephone, which is next to an open door\n\n leading to the bathroom. The phone is busy. A\n\n girl sits listening to it, giggling wildly. Another\n\n girl beside her is laughing too. They are apparently\n\n sharing a conversation with some man on the other end\n\n of the wire. The telephone passes from hand to hand.\n\n Gillis watches impatiently, then\n\n GILLIS\n\n When youlre through with that\n\n thing, can I have it?\n\n The girl just nods, going on with her chattering.\n\n Gillis stands waiting, and Betty Schaefer comes up\n\n with his glass.\n\n BETTY\n\n You forgot this.\n\n GILLIS\n\n Thanks.\n\n BETTY\n\n I've been hoping to run into you.\n\n GILLIS\n\n What for? To recover that knife\n\n you stuck in my back?\n\n BETTY\n\n I felt a little guilty, so I got\n\n out some of your old stories.\n\n GILLIS\n\n Why, you sweet kid.\n\n BETTY\n\n There's one called....Window...\n\n something with a window.\n\n GILLIS\n\n Dark Windows. How did you\n\n like it?\n\n BETTY\n\n I didn't.\n\n GILLIS\n\n Thank you.\n\n BETTY\n\n Except for about six pages.\n\n You've got a flashback there ...\n\n There is too much racket for her.\n\n BETTY\n\n Is there someplace we can talk?\n\n GILLIS\n\n How about the Rainbow Room?\n\n They squeeze their way towards the bathroom, past\n\n Artie.\n\n ARTIE\n\n I said you could have my couch.\n\n I didn't say you could have my\n\n girl.\n\n BETTY\n\n This is shop talk.\n\n She and Gillis go through the open door into\n\n C-20 ARTIE'S BATHROOM\n\n It's a little less noisy, although there are some\n\n guests there, chatting and having fun. Betty and\n\n Gillis sit down on the edge of the tub.\n\n GILLIS\n\n Now if I got you correctly, there\n\n was a short stretch of my fiction\n\n you found worthy of notice.\n\n BETTY\n\n The flashback in the courtroom,\n\n when she tells about being a\n\n school teacher.\n\n GILLIS\n\n I had a teacher like that once.\n\n BETTY\n\n Maybe that's why it's good.\n\n It's true, it's moving. Now\n\n why don't you use that character...\n\n GILLIS\n\n Who wants true? Who wants moving?\n\n BETTY\n\n Drop that attitude. Here's some-\n\n thing really worth while.\n\n GILLIS\n\n Want me to start right now?\n\n Maybe there's some paper around.\n\n BETTY\n\n I'm serious. I've got a few ideas.\n\n GILLIS\n\n I've got some ideas myself. One\n\n of them being this is New Year's\n\n Eve. How about living it up a\n\n little?\n\n BETTY\n\n As for instance?\n\n GILLIS\n\n Well....\n\n BETTY\n\n We could make some paper boats\n\n and have a regatta. Or should\n\n we just turn on the shower?\n\n GILLIS\n\n How about capturing the kitchen\n\n and barricading the door?\n\n BETTY\n\n Are you hungry?\n\n GILLIS\n\n Hungry? After twelve years in\n\n the Burmese jungle. I am starving,\n\n Lady Agatha -- starving for a\n\n white shoulder --\n\n BETTY\n\n Phillip, you're mad!\n\n One of the girls who was on the phone comes to\n\n the door.\n\n GIRL\n\n You can have the phone now.\n\n GILLIS\n\n (Paying no attention)\n\n Thirsting for the coolness of\n\n your lips -\n\n BETTY\n\n No, Phillip, no. We must be\n\n strong. You're still wearing\n\n the uniform of the Coldstream\n\n Guards! Furthermore, you can\n\n have the phone now.\n\n GILLIS\n\n O.K.\n\n (He gets up, starts\n\n out, turns)\n\n I find I'm terribly afraid of\n\n losing you.\n\n BETTY\n\n You won't.\n\n (She takes the glass\n\n out of his hand)\n\n I'll get us a refill of\n\n this awful stuff.\n\n GILLIS\n\n You'll be waiting for me?\n\n BETTY\n\n With a wildly beating heart.\n\n GILLIS\n\n Life can be beautiful!\n\n He leaves.\n\n C-21 THE MAIN ROOM\n\n Gillis squeezes himself through some guests to\n\n the phone. He has to stand in a cramped position,\n\n holding the instrument close to him as he dials\n\n a number.\n\n GILLIS\n\n Max? This is Mr. Gillis.\n\n I want you to do me a favor.\n\n C-22 NORMA DESMOND HOUSE\n\n Max is at the phone, in the lower hall.\n\n MAX\n\n I am sorry, Mr. Gillis.\n\n I cannot talk now.\n\n C-23 GILLIS ON THE PHONE\n\n GILLIS\n\n Yes you can. I want you to get\n\n my old suitcase and I want you\n\n to throw in my old clothes --\n\n the ones I came with, and my\n\n typewriter. I'll have somebody\n\n pick them up.\n\n C-24 MAX AT THE PHONE\n\n MAX\n\n I have no time to talk. The\n\n doctor is here.\n\n C-25 GILLIS ON THE PHONE\n\n GILLIS\n\n What doctor? What's going on?\n\n C-26 MAX AT THE PHONE\n\n MAX\n\n She got the razor from your\n\n room. She cut her wrists.\n\n Max hangs up, moves toward the staircase.\n\n C-27 GILLIS AT THE PHONE\n\n GILLIS\n\n Max ! Max !\n\n He hangs up the dead receiver, stands numb with\n\n shock. Betty elbows her way up to him, carrying\n\n the two punch glasses filled again.\n\n BETTY\n\n I just got the recipe: take\n\n two packages of cough drops,\n\n dissolve in one gallon of\n\n lukewarm grape juice --\n\n Gillis looks up at her. Without a word he pushes\n\n her aside so that she spills the drink. He makes\n\n his way through the guests to the Vicuna coat, pulls\n\n it from the shelf, some books tumbling with it, and\n\n rushes towards the door and out. Betty stands look-\n\n ing after him, completely bewildered.\n\n DISSOLVE TO:\n\n C-28 EXT. DESMOND HOUSE - (NIGHT, RAIN)\n\n The doctor's car is parked in the driveway. A taxi\n\n pulls up. Gillis, in his Vicuna coat now, jumps\n\n out, throws a couple of dollars to the rdriver and\n\n runs toward the house.\n\n C-28a DOORWAY, NORMA DESMOND HOUSE>\n\n Max is opening the door to let out the doctor, a\n\n professional looking man carrying a black bag.\n\n Gillis runs into the SHOT.\n\n GILLIS\n\n How is she?\n\n MAX\n\n She is upstairs.\n\n Gillis starts to push past Max. Max grabs his arm.\n\n MAX\n\n Be careful. Do not race up the\n\n stairs. The musicians must not\n\n know what has happened.\n\n Gillis goes into the house.\n\n C-29 ENRANCE HALL AND STAIRCASE\n\n Gillis crosses the hall and starts up the stairs.\n\n C-3O INT. NORMA DESMOND'S ROOM\n\n Only one alabaster lamp lights the big, cold room.\n\n On the bed lies Norma in her evening dress. She is\n\n white as a sheet. Her wrists are bandaged. Her eyes\n\n are wide open, staring at the ceiling. One of her\n\n shoes has halt slipped off her foot. The other is\n\n on. Gillis opens the door and stands there tor a\n\n second. Then he slowly moves to the toot of the bed.\n\n He takes the shoes from her feet and puts them on\n\n the floor.\n\n NORMA\n\n Go away.\n\n GILLIS\n\n What kind of a silly thing was\n\n that to do?\n\n NORMA\n\n To fall in love with you -- that\n\n was the idiotic thing.\n\n GILLIS\n\n It sure would have made attractive\n\n headlines: Great Star Kills Her-\n\n self for Unknown Writer.\n\n NORMA\n\n Great stars have great pride.\n\n She puts one bandaged forearm over her eyes, sobbing.\n\n Gillis walks slowly over to the mantelpiece, stands\n\n there for awhile.\n\n NORMA\n\n Go away. Go to that girl of yours.\n\n GILLIS\n\n Look, I was making that up because\n\n I thought the whole thing was a\n\n mistake. I didn't want to hurt you.\n\n You've been good to me. You're the\n\n only person in this stinking town\n\n that has been good to me.\n\n NORMA\n\n Why don't you just say thank you\n\n and go, go, go --\n\n GILLIS\n\n Not until you promise to act like\n\n a sensible human being.\n\n NORMA\n\n I'll do it again, I'll do it again,\n\n I'll do it again!\n\n Gillis stands looking at her helplessly.\n\n C-31 LIVING ROOM, THE DESMOND HOUSE\n\n The candles burned down, the orchestra playing to\n\n the emptiness. The orchestra leader looks at his\n\n watch, rises, silences the orchestra, then starts\n\n them in on Auld Lang Syne.\n\n C-32 INT. NORMA'S ROOM\n\n Gillis still stands. Norma lies on the bed, arms\n\n over her eyes, sobbing.\n\n GILLIS\n\n Happy New Year.\n\n Norma continues to sob. Gillis goes to the bed,\n\n puts his arms on her shoulders and turns her around.\n\n GILLIS\n\n Happy New Year.\n\n Norma looks at him, tears in her eyes. Slowly she\n\n enfolds him in her bandaged arms.\n\n NORMA\n\n Happy New Year. darling.\n\n She kisses him.\n\n DISSOLVE\n\n END OF SEQUENCE "C"\n\n SEQUENCE "D"\n\n DISSOLVE IN ON:\n\n D-1 INT. HALLWAY, NORMA GILLIS' VOICE\n\n DESMOND'S HOUSE (DAY) Around the middle of May\n\n some incidents happened\n\n The telephone is heard which I think I should tell\n\n ringing. Max comes from you about.\n\n living room to the phone,\n\n picks it up.\n\n MAX\n\n Hello ... Yes?\n\n D-1a BETTY SCHAEFER, AT THE PHONE ON HER DESK IN THE\n\n READERS' DEPARTMENT\n\n BETTY\n\n Is this Crestview 5-1733? ... I'm\n\n sorry to bother you again, but I've\n\n confirmed the number. I must speak\n\n to Mr. Gillis.\n\n D-1b MAX, AT THE PHONE\n\n MAX\n\n He is not here.\n\n D-1c BETTY ON THE PHONE\n\n BETTY\n\n Where can I reach him? Maybe\n\n somebody else in the house could\n\n tell me.\n\n D-1d MAX ON THE PHONE\n\n MAX\n\n Nobody here can give you any\n\n information. You will please\n\n not call again.\n\n He hangs up. From off comes:\n\n NORMA'S VOICE\n\n Who was it, Max? What is it?\n\n D-1e PATIO, NORMA'S HOUSE\n\n It is a sunny day. The garden is in somewhat better\n\n shape. The old house looks less unkept. The pool\n\n is filled. Norma sits on a wicker chaise longue, her\n\n face shielded by an enormous straw hat, her eyes by\n\n dark glasses. Gillis, in bathing trunks, is on a\n\n rubber mattress in the pool. Max comes to the\n\n entrance door.\n\n MAX\n\n Nothing, Madame. Somebody Inqu-\n\n iring about a stray dog. We must\n\n have a number very similar to the\n\n pound.\n\n He starts to turn back.\n\n NORMA\n\n Wait a minute. I want you to get\n\n out the car. You're going to\n\n take the script over to Paramount\n\n and deliver it to Mr. De Mille in\n\n person.\n\n MAX\n\n Yes, Madame.\n\n He goes into the house.\n\n GILLIS\n\n (climbing out\n\n of the water)\n\n You're really going to send it\n\n to De Mille?\n\n NORMA\n\n This is the right day.\n\n She indicates a typewritten letter she is holding.\n\n NORMA (Cont'd)\n\n The chart from my astrologer.\n\n She read deMille's horoscope.\n\n She read mine.\n\n GILLIS\n\n Did she read the script?\n\n NORMA\n\n DeMille is Leo. I'm Scorpio.\n\n Mars has been transmitting\n\n Jupiter for weeks. Today is\n\n the day of greatest conjuction.\n\n Now turn around. Let me dry\n\n you.\n\n She puts the towel around his sholders and starts\n\n drying him.\n\n GILLIS\n\n I hope you realize, Norma,\n\n that scripts don't sell on\n\n astrologers' charts.\n\n NORMA\n\n I'm not just selling the script.\n\n I'm selling me. DeMille always\n\n said I was his greatest star.\n\n GILLIS\n\n When did he say it, Norma?\n\n NORMA\n\n So he said it quite a few years\n\n ago. So what? I never looked\n\n better in my life. Do you know\n\n why? Because I've never been as\n\n happy in my life.\n\n She kisses him.\n\n DISSOLVE TO:\n\n D-2 INT. THE ISOTTA, DRIVING\n\n DOWN SUNSET ABOUT 8:30\n\n IN THE EVENING GILLIS' VOICE\n\n A few evenings later we\n\n Max is driving. In the were going to the house of\n\n tonneau sit Norma, in a one of the waxworks for\n\n chinchilla wrap, and some bridge. She'd taught\n\n Gillis in his tuxedo. me how to play bridge by\n\n Norma is rummaging then, just as she'd taught\n\n through her evening me some fancy tango steps,\n\n bag. She finds a and what wine to drink\n\n cigarette case, opens with what fish.\n\n it. It is empty.\n\n NORMA\n\n That idiot. He forgot to fill\n\n my cigarette case.\n\n GILLIS\n\n (Proffering his case)\n\n Have one of mine.\n\n NORMA\n\n They're awful. They make me cough.\n\n GILLIS\n\n (Pushing open the glass\n\n partition, to Max)\n\n Pull up at the drugstore, will\n\n you, Max.\n\n (To Norma)\n\n I'll get you some.\n\n NORMA\n\n You're a darling.\n\n She takes a dollar bill from her purse and gives it\n\n to him.\n\n D-3 EXT. SCHWAB'S DRUGSTORE\n\n The car drives up and Gillis hurries into the store.\n\n D-4 INT. SCHWAB'S DRUGSTORE\n\n Business is still rather lively. There are about a\n\n dozen shoppers, and the soda counter is half filled.\n\n Gillis enters and steps to the tobacco counter.\n\n GILLIS\n\n (To the salesgirl)\n\n Give me a pack of those Turkish\n\n cigarettes -- Melachrinos.\n\n The girl opens the glass showcase to locate the fancy\n\n brand. From OFF comes\n\n ARTIE'S VOICE\n\n Stick 'em up, Gillis, or I'll\n\n let you have it!\n\n Gillis turns.\n\n D-5 AT THE SODA FOUNTAIN\n\n Artie Green and Betty Schaefer sit having a sandwich\n\n and a milk shake. With his forefinger and a sound\n\n effect, Artie riddles Gillis' body. Gillis walks\n\n INTO THE SHOT.\n\n GILLIS\n\n Hello, Artie. Good evening,\n\n Miss Schaefer.\n\n BETTY\n\n (Excitedly)\n\n You don't know how glad I am\n\n to see youl\n\n ARTIE\n\n Walking out on the mob. What's\n\n the big idea?\n\n GILLIS\n\n I'm sorry about New Year's. Would\n\n you believe me if I said I had\n\n to be with a sick friend?\n\n ARTIE\n\n Someone in the formal set, no\n\n doubt, with a ten-carat kidney\n\n stone.\n\n BETTY\n\n Stop it, Artie, will you?\n\n (To Gillis)\n\n Where have you been keeping your-\n\n self? I've got the most wonderful\n\n news for you.\n\n GILLIS\n\n I haven't been keeping myself at\n\n all. Not lately.\n\n BETTY\n\n I called your agent. I called the\n\n Screen Writers Guild. Finally your\n\n old apartment gave me some Crestview\n\n number. There was always somebody\n\n with an accent growling at me. You\n\n were not there. You were not to be\n\n spoken to. They never heard of you.\n\n GILLIS\n\n Is that so? What's the wonderful\n\n news?\n\n BETTY\n\n Sheldrake likes that angle about\n\n the teacher.\n\n GILLIS\n\n What teacher?\n\n BETTY\n\n Dark Windows. I got him all\n\n hopped up about it.\n\n GILLIS\n\n You did?\n\n BETTY\n\n He thinks it could be made into\n\n something.\n\n GILLIS\n\n Into what? A lampshade?\n\n BETTY\n\n Into something for Barbara Stan-\n\n wyck. They have a commitment with\n\n Barbara Stanwyck.\n\n ARTIE\n\n Unless you'd rather have Sarah\n\n Bernhardt.\n\n BETTY\n\n This is on the level. Sheldrake\n\n really went for it.\n\n GILLIS\n\n O.K. Where's the cash?\n\n BETTY\n\n Where's the story? I bluffed it\n\n out with a few notions of my own.\n\n It's really just a springboard.\n\n It needs work.\n\n GILLIS\n\n I was afraid of that.\n\n BETTY\n\n I've got twenty pages of notes.\n\n I've got a pretty good character\n\n for the man.\n\n ARTIE\n\n Could you write in plenty of back-\n\n ground action, so they'll need an\n\n extra assistant director?\n\n BETTY\n\n Shut up, Artie.\n\n (To Gillis)\n\n Now if we could sit down for two\n\n weeks and get a story.\n\n GILLIS\n\n Sorry, Miss Schaefer, but I've\n\n given up writing on spec.\n\n BETTY\n\n I tell you this is half sold.\n\n GILLIS\n\n As a matter of fact. I've given\n\n up writing altogether.\n\n Max has appeared in the door.\n\n MAX\n\n Mr. Gillis, if you please.\n\n GILLIS\n\n Right with you.\n\n Max leaves.\n\n ARTIE\n\n The accent! I've got it: this guy\n\n is in the pay of a foreign government.\n\n Get those studs. Get those cuff-links.\n\n GILLIS\n\n I've got to run along. Thanks any-\n\n way for your interest in my career.\n\n BETTY\n\n It's not your career -- it's mine.\n\n I kind of hoped to get in on this\n\n deal. I don't want to be a reader\n\n all my life. I want to write.\n\n GILLIS\n\n Sorry if I crossed you up.\n\n BETTY\n\n You sure have.\n\n GILLIS\n\n So long.\n\n He leaves.\n\n ARTIE\n\n (Patting her hand)\n\n Babe, it's like that producer says:\n\n In life, you've got to take the\n\n bitter with the sour.\n\n D-6 THE ISOTTA, PARKED OUTSIDE\n\n Gillis comes from Schwab's, gets into the car.\n\n Max takes off.\n\n NORMA\n\n What on earth, darling? It took\n\n you hours.\n\n GILLIS\n\n I ran into some people I knew.\n\n NORMA\n\n Where are my cigarettes?\n\n GILLIS\n\n Where are your...?\n\n He realizes he's forgotten them, takes the dollar\n\n and hands it back to her.\n\n GILLIS\n\n Norma, you're smoking too much.\n\n DISSOLVE TO:\n\n D-7 LIVING ROOM, NORMA\n\n DESMOND'S HOUSE \n\n (EARLY AFTERNOON)\n\n Start on a tiny GILLIS' VOICE\n\n parasol being Whenever she suspected I\n\n twirled...Norma was getting bored, she\n\n peeks out from one would put on a live show\n\n side of the parasol, for me: the Norma Desmond\n\n a bandanna tied Follies. Her first number\n\n around her head with was always the Mack Sennett\n\n a rabbit's-ear bow. Bathing Beauty.\n\n She bats her eyes,\n\n winks roguishly.\n\n THE CAMERA PULLS BACK to reveal that Norma's black\n\n pyjama trousers are rolled up over her knees and her\n\n black stockings rolled down below them. The whole\n\n effect approximates a Mack Sennett bathing costume\n\n pretty effectively. She points at a leather pour.\n\n NORMA\n\n This is a rock.\n\n She climbs on it, pantomimes timidity, an attempted\n\n dive, then jumps off.\n\n Gillis lolls on a couch, watching the performance,\n\n very bored.\n\n NORMA\n\n I can still see myself in the\n\n line: Bebe Daniels, Marie Prevost,\n\n Mabel Normand ... Mabel was always\n\n stepping on my feet ...What's the\n\n matter with you, darling? Why are\n\n you so glum?\n\n GILLIS\n\n (Lighting a cigarette\n\n with a match)\n\n Nothing is the matter. I'm having\n\n a great time. Show me some more.\n\n NORMA\n\n (Taking the match)\n\n All right. Give me this. I need\n\n it for a moustache. Now close\n\n your eyes.\n\n She runs out of the GILLIS' VOICE\n\n picture. Gillis has Something was the matter,\n\n closed his eyes. all right. I was thinking\n\n THE CAMERA MOVES to about that girl of Artie's,\n\n his face. that Miss Schaefer. She\n\n was so like all us writers\n\n when we first hit Holly-\n\n wood -- itching with am-\n\n bition, panting to get\n\n your names up there:\n\n Screenplay by. Original\n\n Story by. Hmph! Audiences\n\n don't know somebody sits\n\n down and writes a picture.\n\n They think the actors make\n\n it up as they go along.\n\n NORMA'S VOICE\n\n Open your eyes.\n\n Gillis opens his eyes.\n\n Norma has equipped herselr with a derby hat, a cane,\n\n and blacked in a small moustache. She goes into a\n\n little Chaplin routine. While she is doing it, the\n\n telephone rings. After a moment Max comes to the\n\n living room door.\n\n MAX\n\n Madame is wanted on the telephone.\n\n NORMA\n\n You know better than to interrupt me.\n\n MAX\n\n Paramount is calling.\n\n NORMA\n\n Who?\n\n MAX\n\n Paramount studios.\n\n NORMA\n\n (To Gillis)\n\n Now, now do you belive me? I told\n\n you deMille would jump at it.\n\n MAX\n\n It is not Mr. deMille in person.\n\n It is someone by the name or Gordon\n\n Cole. He says it's very important.\n\n NORMA\n\n Certainly it's important. It's\n\n important enough for Mr. deMille\n\n to call me personally. The idea\n\n of having an assistant call me!\n\n MAX\n\n I myself was surprised at Mr. de\n\n Mille's manners.\n\n NORMA\n\n Say that I'm busy, and hang up.\n\n MAX\n\n Very good, Madam.\n\n He bows and exits.\n\n NORMA\n\n How do you like that? We've\n\n made twelve pictures together.\n\n His greatest successes.\n\n GILLIS\n\n Maybe deMille is shooting.\n\n NORMA\n\n I know that trick! He wants to\n\n belittle me. He's trying to get\n\n my price down. I've waited\n\n twenty years for this call. Now\n\n Mr. deMille can wait till I'm\n\n good and ready.\n\n DISSOLVE TO:\n\n D-8 NORMA, IN THE TONNEAU\n\n OF THE LIMOUSINE,\n\n DRIVING DOWN MELROSE\n\n She is in full makeup, GILLIS' VOICE\n\n with a veil, a daring About three days later she\n\n hat, a suit so stunning was good and ready. In-\n\n only she would venture credible as it may seem,\n\n to wear it. THE CAMERA there had been some more\n\n PULLS BACK. Beside her of those calls from\n\n sits Gillis in the glen Paramount. So she put on\n\n plaid suit. Max is about half a pound of\n\n driving. makeup, fixed it up with\n\n a veil, and set forth to\n\n see deMille in person.\n\n Norma is examining her face in the mirror of her\n\n vanity. Max, while driving, sees her in the rear\n\n view mirror.\n\n MAX\n\n If you will pardon me, Madame.\n\n The shadow over the left eye\n\n is not quite balanced.\n\n NORMA\n\n Thank you, Max.\n\n With a handkerchief, she corrects it.\n\n D-9 MAIN GATE, EXT. PARAMOUNT STUDIO\n\n The car drives down Bronson and stops smack in front\n\n of the iron gate. A young policeman is talking to\n\n an extra; an old policeman sits reading a newspaper.\n\n Max sounds the horn impatiently.\n\n YOUNG POLICEMAN\n\n Hold that noise!\n\n MAX\n\n To see Mr. de Mille. Open the gate.\n\n YOUNG POLICEMAN\n\n Mr. deMille is shooting. You\n\n got an appointment?\n\n MAX\n\n No appointment is necessary. I\n\n am bringing Norma Desmond.\n\n YOUNG POLICEMAN\n\n Norma Who?\n\n Norma has rolled down the window on her side. She\n\n calls to the old policeman.\n\n NORMA\n\n Jonesy! Come here, Jonesy!\n\n OLD POLICEMAN\n\n Yeah?\n\n (He comes forward slowly)\n\n Why, if it isn't Miss Desmond!\n\n How have you been, Miss Desmond?\n\n NORMA\n\n Fine, Jonesy. Now open that gate.\n\n OLD POLICEMAN\n\n Sure, Miss Desmond.\n\n (To the young policeman}\n\n Come on, Mac.\n\n YOUNG POLICEMAN\n\n They can't drive on the lot\n\n without a pass.\n\n OLD POLICEMAN\n\n Miss Desmond can. Come on.\n\n They fling open the gate.\n\n OLD POLICEMAN\n\n (As the car drives through)\n\n Stage eighteen, Miss Desmond.\n\n NORMA\n\n Thank you, Jonesy. And teach\n\n your friend some manners. Tell\n\n him without me he wouldn't have\n\n any job, because without me there\n\n wouldn't be any Paramount Studio.\n\n (To Max)\n\n Go on.\n\n They drive through the gates. The old policeman\n\n goes to wall phone beside the gate, dials a number.\n\n OLD POLICEMAN\n\n (Into phone)\n\n Norma Desmond coming in to\n\n see Mr. deMille.\n\n D-10 STAGE 18\n\n A scene from SAMPSON AND DELILAH is being rehearsed\n\n in the background. The usual turbulent activity\n\n surrounds it: extras. makeup men, grips,\n\n assistants, etc., etc. In the dim foreground a\n\n stage hand is answering a stand telephone. He\n\n puts down the phone and moves (CAMERA WITH HIM)\n\n to a second assistant.\n\n STAGE HAND\n\n Norma Desmond is coming to see\n\n Mr. deMille.\n\n The second assistant walks (CAMERA WITH HIM)\n\n to the first assistant.\n\n 2nd ASSISTANT\n\n Norma Desmond coming in to\n\n see Mr. deMille.\n\n The first assistant (CAMERA WITH HIM) hurries\n\n to the set. Sitting with his back toward us\n\n is C.B. himself. He is rehearsing a scene with\n\n Hedy Lamarr.\n\n 1ST ASSISTANT\n\n Norma Desmond is coming in to\n\n see you, Mr. deMille.\n\n C. B. turns his head.\n\n DEMILLE\n\n Norma Desmond?\n\n lst ASSISTANT\n\n She must be a million years old.\n\n DEMILLE\n\n I hate to think where that puts\n\n me. I could be her father.\n\n 1ST ASSISTANT\n\n I'm terribly sorry, Mr. de Mille.\n\n By this time de Mille is on his feet.\n\n DEMILLE\n\n It must be about that appalling\n\n script of hers. What can I say\n\n to her? What can I say?\n\n 1ST ASSISTANT\n\n I can tell her you're all tied\n\n up in the projection room. I\n\n can give her the brush ...\n\n DEMILLE\n\n Listen, thirty million fans\n\n have given her the brush.\n\n Isn't that enough?\n\n 1ST ASSISTANT\n\n I didn't mean to --\n\n DEMILLE\n\n Of course you didn't. You didn't\n\n know Norma Desmond as a plucky\n\n little girl of seventeen, with\n\n more courage and wit and heart\n\n than ever came together in one\n\n youngster.\n\n 1ST ASSISTANT\n\n I hear she was a terror to\n\n work with.\n\n DEMILLE\n\n She got to be. A dozen press\n\n agents working overtime can\n\n do terrible things to the human\n\n spirit.\n\n (to the set)\n\n Hold everything.\n\n He leaves, accompanied by his entourage.\n\n D-11 EXT. STAGE 18\n\n Norma's limousine drives up. Max dismounts\n\n and opens the door.\n\n NORMA\n\n (taking Gillis's hand)\n\n Don't you want to come along,\n\n darling?\n\n GILLIS\n\n I don't think so. It's your\n\n script. It's your show.\n\n Good luck.\n\n NORMA\n\n Thank you, darling.\n\n She presses his hand against her cheek, descends\n\n from the car and walks toward -\n\n D-12 THE DOOR OF STAGE 18\n\n The first assistant is holding it open. In the door-\n\n way stands Mr. deMille. Seeing Norma, he stretches\n\n out his arms.\n\n DE MILLE\n\n Hello, young fellow.\n\n NORMA\n\n Hello, Mr. deMille.\n\n She has reached him. They embrace.\n\n NORMA\n\n Last time I saw you was someplace\n\n very gay. I remember waving to you.\n\n I was dancing on a table.\n\n DE MILLE\n\n Lots of people were. Lindbergh had\n\n just landed in Paris. Come on in.\n\n He leads her into\n\n D-13 STAGE 18\n\n During the ensuing dialogue, Mr. deMille walks Norma\n\n towards the set.\n\n DE MILLE\n\n Norma, I want to apologize for\n\n not calling you.\n\n NORMA\n\n You'd better. I'm very angry.\n\n DE MILLE\n\n I'm pretty busy, as you can see...\n\n NORMA\n\n That's no excuse. You read the\n\n script, didn't you?\n\n DE MILLE\n\n Yes, I did.\n\n NORMA\n\n Then you could have picked up the\n\n phone yourself instead of leaving\n\n it to one of your assistants.\n\n DE MILLE\n\n What assistant?\n\n NORMA\n\n Don't play innocent. Somebody\n\n named Gordon Cole.\n\n DE MILLE\n\n Gordon Cole?\n\n NORMA\n\n And if you hadn't been pretty\n\n darned interested in that script,\n\n he wouldn't have tried to get\n\n me on the phone ten times.\n\n DE MILLE\n\n Gordon Cole... Look, Norma,\n\n I'm in the middle of a rehearsal.\n\n (Indicating his\n\n own chair)\n\n Make yourself comfortable.\n\n He walks onto the set, accompanied by his assistants.\n\n DE MILLE\n\n (Sotto voce, to his\n\n first assistant)\n\n Get me Gordon Cole on the phone.\n\n Meanwhile, Norma starts to sit, sees the name\n\n MISS LAMARR on the chair and with a look of\n\n distaste changes and sits on the one marked\n\n C.B. DE MILLE. From somewhere comes\n\n A VOICE\n\n Hey, Miss Desmond! Miss Desmond!\n\n She looks around her.\n\n VOICE\n\n Up here!\n\n Norma looks up at the scaffolding.\n\n On the scaffold stands one of the electricians,\n\n next to his light.\n\n ELECTRICIAN\n\n It's met It's Hog-eyel\n\n Norma waves at him.\n\n NORMA\n\n Hello.\n\n Hog-eye points his light at her.\n\n HOG-EYE\n\n Let's get a look at you.\n\n The beam of the lamp moves toward Norma. It hits\n\n her. She sits bathed in light. A couple of old\n\n costume extras recognize her.\n\n EXTRAS\n\n Say, it's Norma! Norma Desmond!\n\n They rush over and start wringing her hand. Into\n\n the shot comes a middle-aged hairdresser.\n\n HAIRDRESSER\n\n Hello, Miss Desmond. It's Bessie.\n\n Some elderly electricians and stagehands move in.\n\n D-14 ANOTHER PART OF THE STAGE\n\n The first assistant brings the portable phone to\n\n deMille. DeMille lifts the receiver.\n\n DE MILLE\n\n Hello.\n\n D-15 GORDON COLE'S OFFICE IN THE PROPERTY DEPARTMENT,\n\n GORDON COLE ON THE PHONE.\n\n COLE\n\n Prop Department. Gordon Cole speaking.\n\n D-16 DE MILLE ON THE PHONE\n\n DE MILLE\n\n Cole, this is C. B. deMille. Have\n\n you been calling Norma Desmond?...\n\n What's it about?\n\n D-17 GORDON COLE, ON THE PHONE\n\n COLE\n\n It's that car of hers -- an old\n\n Isotta-Fraschini. Her chauffeur\n\n drove it on the lot the other day.\n\n It looks just right for the Crosby\n\n picture. We want to rent it for a\n\n couple of weeks.\n\n D-18 DE MILLE ON THE PHONE\n\n DE MILLE\n\n (Troubled)\n\n Oh. Well, thank you.\n\n He hangs up, walks back towards Norma. (CAMERA\n\n WITH HIM).\n\n Norma stills sits in the shaft of light, surrounded\n\n by about a dozen people who have come up to pay court.\n\n DeMille gestures up to Hog-eye and the light shifts\n\n away. The people about Norma disperse slowly with\n\n various ad-libs.\n\n DE MILLE\n\n Well, Norma ...\n\n (He sits down next to her)\n\n I got hold of Gordon Cole.\n\n Norma hasn't heard a word.\n\n NORMA\n\n Did you see them? Did you see\n\n how they came?\n\n DE MILLE\n\n You know, crazy things happen in\n\n this business. I hope you haven't\n\n lost your sense of humor ...\n\n Suddenly he realizes that she is crying. She takes\n\n the handkerchief from his pocket and puts it over her\n\n eyes.\n\n DEMILLE\n\n What's the matter, Norma?\n\n NORMA\n\n Nothing. I just didn't realize\n\n what it would be like to come back\n\n to the old studio. I had no idea\n\n how I'd missed it.\n\n DEMILLE\n\n We've missed you too, dear.\n\n NORMA\n\n We'll be working again, won't we, Chief?\n\n We'll make our greatest picture.\n\n DEMILLE\n\n That's what I want to talk to you about.\n\n NORMA\n\n It's a good script, isn't it?\n\n DEMILLE\n\n It's got a lot of good things. Of\n\n course, it would be an expensive picture...\n\n NORMA\n\n I don't care about the money.\n\n I just want to work again. You\n\n don't know what it means to know\n\n that you want me.\n\n DEMILLE\n\n Nothing would thrill me more --\n\n if it were possible.\n\n NORMA\n\n But remember, darling -- I don't\n\n work before ten in the morning,\n\n and never after 4:30 in the afternoon.\n\n The first assistant comes up.\n\n 1ST ASSISTANT\n\n We're ready with the shot, Mr. deMille.\n\n DEMILLE\n\n You'll pardon me, Norma? Why\n\n don't you just sit and watch?\n\n (He steps onto the set)\n\n O.K. Here we go.\n\n 1ST ASSISTANT\n\n Roll 'em.\n\n DEMILLE\n\n Action!\n\n The scene starts.\n\n D-19 THE ISOTTA, PARKED OUTSIDE STAGE 18\n\n Max stands talking to Gillis, who is seated in the\n\n car.\n\n MAX\n\n (Pointing to the row\n\n of offices in the\n\n building opposite)\n\n You see those offices there, Mr.\n\n Gillis? They used to be her\n\n dressing room, The whole row.\n\n GILLIS\n\n That didn't leave much for Wallace\n\n Reid.\n\n MAX\n\n He had a great big bungalow on\n\n wheels. I had the upstairs. See\n\n where it says 'Readers' Department'?\n\n I remember my walls were covered\n\n with black patent leather...\n\n The words "Readers' Department" have registered on\n\n Gillis' mind. He gets out of the car.\n\n GILLIS\n\n I'll be with you in a minute.\n\n He crosses the street towards the green staircase\n\n leading to the second floor.\n\n Meanwhile, two prop men walking down the street\n\n come into the SHOT.\n\n 1ST PROP MAN\n\n Hey, that's the comic car Cole\n\n was talking about!\n\n (To Max)\n\n Do you mind if we look inside?\n\n MAX\n\n Go away. Go away.\n\n D-2O CUBICLE IN THE READERS' DEPARTMENT\n\n Behind the desk sits Betty, typing the synopsis of\n\n a novel, a half-eaten apple marking her place. The\n\n door behind her opens and Gillis enters.\n\n GILLIS\n\n Just so you don't think I'm a\n\n complete swine -- if there's\n\n anything in Dark Windows you\n\n can use, take it. It's all\n\n yours.\n\n BETTY\n\n Well, for heaven's sake!\n\n She moves the book and the apple aside and points at\n\n the free space on the desk.\n\n BETTY\n\n Have a chair.\n\n Gillis sits on the desk.\n\n GILLIS\n\n I mean it. It's no good to me\n\n anyway. Help yourself.\n\n BETTY\n\n Why should you do that?\n\n GILLIS\n\n If you get a hundred thousand for\n\n it, you buy me a box of chocolate\n\n creams. If you get an Oscar, I\n\n get the left foot.\n\n BETTY\n\n You know, I'd take you up on that\n\n in a minute. I'm just not good\n\n enough to do it all by myself.\n\n GILLIS\n\n What about all those ideas you had?\n\n BETTY\n\n See if they make sense. To begin\n\n with, I think you should throw out\n\n all that psychological stuff --\n\n exploring a killer's sick mind.\n\n GILLIS\n\n Psychopaths sell like hotcakes.\n\n BETTY\n\n This story is about teachers --\n\n their threadbare lives, their\n\n struggles. Here are people doing\n\n the most important job in the\n\n world, and they have to wprry\n\n about getting enough money to\n\n re-sole their shoes. To me it\n\n can be as exciting as any chase,\n\n any gunplay.\n\n GILLIS\n\n Check.\n\n BETTY\n\n Now I see her teaching day classes\n\n while he teaches night school. The\n\n first time they meet ...\n\n From below comes the SOUND of the Isotta's horn.\n\n GILLIS\n\n Look, if you don't mind, I haven't\n\n got time to listen to the whole\n\n plot ...\n\n BETTY\n\n I'll make it short.\n\n GILLIS\n\n Sorry. It's your baby now.\n\n BETTY\n\n I'm not good enough to write it\n\n alone. We'll have to do it together.\n\n GILLIS\n\n I'm all tied up. I can't.\n\n BETTY\n\n Couldn't we work in the evenings?\n\n Six o'clock in the morning? This\n\n next month I'm completely at your\n\n disposal. Artie is out of town.\n\n GILLIS\n\n What has Artie to do with it.\n\n BETTY\n\n We're engaged.\n\n GILLIS\n\n Good for you. You've got yourself\n\n the best guy in town.\n\n BETTY\n\n I think so. They're on location\n\n in Arizona, shooting a Western.\n\n I'm free every evening, every week-\n\n end. If you want, we could work at\n\n your place.\n\n GILLIS\n\n It's just impossible.\n\n BETTY\n\n Nobody can be that busy.\n\n There is another honk: from down below.\n\n GILLIS\n\n Look, Betty, It can't be done.\n\n It's out.\n\n BETTY\n\n You're tough, all right.\n\n GILLIS\n\n You're on your own. Stop being\n\n chicken-hearted and write that story.\n\n BETTY\n\n Honest to goodness, I hate you.\n\n GILLIS\n\n (Turning 1n the open door)\n\n And don't make it too dreary. How\n\n about this for a situation: she\n\n teaches daytimes. He teaches at\n\n night. Right? They don't even know\n\n each other, but they share the same\n\n room. It's cheaper that way. As a\n\n matter of fact, they sleep in the\n\n same bed -- in shifts, of oourse.\n\n BETTY\n\n Are you kidding? Because I think\n\n it's good.\n\n GILLIS\n\n So do I.\n\n BETTY\n\n Came on back. Let me show you\n\n where it fits in.\n\n She reaches in a drawer for her notes on Dark\n\n Windows.\n\n GILLIS\n\n (At the door)\n\n So long.\n\n Betty picks up the apple and is about to throw it\n\n after him.\n\n BETTY\n\n Oh, you --\n\n GILLIS\n\n And here's a title: AN APPLE FOR\n\n THE TEACHER.\n\n He ducks out quiokly, slamming the door behind him.\n\n Betty looks after him, then angrlly hurls the\n\n apple into the wastebasket.\n\n D-21 STAIRCASE OUTSIDE READERS' DEPARTMENT\n\n Max is rush1ng up the stairs toward the descending\n\n Gillis.\n\n GILLIS\n\n What's the matter, Max?\n\n MAX\n\n I just found out why all those tele-\n\n phone calls. It is not Miss Desmond\n\n they want. It is the car they want\n\n to rent.\n\n GILLIS\n\n What?\n\n Max has seen something off.\n\n MAX\n\n Ssh...\n\n With his head he indicates\n\n D-22 ENTRANCE TO STAGE 18\n\n The first assistant has opened the door. DeMille\n\n is showing Norma out.\n\n DE MILLE\n\n Goodbye, young fellow. We'll see\n\n what we can do.\n\n NORMA\n\n (embracing him)\n\n I'm not worried. Everything will\n\n be fine. The old team together.\n\n Nothing can stop us.\n\n She turns and walks out of the shot. De Mille\n\n stands for a second watching her, then turns to\n\n his assistant.\n\n DE MILLE\n\n Get Gordon Cole. Tell him to forget\n\n about her car. He can find another\n\n old car. I'll buy him five old cars,\n\n if necessary.\n\n 1ST ASSISTANT\n\n Yes, Mr. De Mille.\n\n They turn back into Stage 18.\n\n D-23 THE ISOTTA\n\n Gillis seated in the rear. Max is helping Norma\n\n in and putting the robe over her.\n\n GILLIS\n\n (Apprehensively)\n\n How did it go?\n\n NORMA\n\n It couldn't have gone better.\n\n It's practically set. Of course,\n\n he has to finish this picture\n\n first, but mine will be his next.\n\n There is an exchange of looks between Max and Gillis.\n\n GILLIS\n\n He must be quite a guy.\n\n NORMA\n\n He'a a shrewd old fox. He can\n\n smell box office. Only I'm going\n\n to outfox him a litt1e. This isn't\n\n going to be C. B. deMille's Salome.\n\n It's going to be Norma Desmond's\n\n Salome, a Norma Desmond Production,\n\n starring Norma Desmond...Home, Max.\n\n MAX\n\n Yes, Miss Desmond.\n\n As he says the words, he and Gillis exchange a glance\n\n in the rear view mirror.\n\n SLOW DISSOLVE:\n\n END OF SEQUENCE "D"\n\n SEQUENCE "E"\n\n DISSOLVE IN ON:\n\n E-1 CLOSEUP OF NORMA'S FACE\n\n GILLIS' VOICE\n\n Absolutely no makeup. A After that, an army of\n\n hand with a strong small beauty experts invaded\n\n flashlight comes into the her house on Sunset\n\n picture. The beam of the Boulevard. She went\n\n flashlight travels over the through a merciless\n\n face, exploring it merci- series of treatments,\n\n lessly. While the light is massages, sweat cabinets,\n\n still on it, two pairs of mud baths, ice compres-\n\n creamed hands come into the ses, electric devices.\n\n shot and start to massage it. She lived on vegetable\n\n juices and went to bed\n\n DISSOLVE TO: at nine. She was deter-\n\n mined to be ready --\n\n ready for those cameras\n\n E-2 A SHORT MONTAGE of various that would never turn.\n\n beauty treatments applied\n\n to Norma.\n\n DISSOLVE TO:\n\n E-3 NORMA BEFORE THE MIRROR\n\n IN HER BEDROOM\n\n It is nine o'clock in the evening. She is in night\n\n gown and negligee and has put triangular patches on\n\n the saddle of her nose and at the outer corner of\n\n each eye. She is rubbing lotion on her hands.\n\n She gets up and crosses to the door of Gillis' room\n\n and opens it a crack.\n\n NORMA\n\n Joe darling, are you there?\n\n E-4 GILLIS' ROOM\n\n It is dark except for a lamp over the chaise longue.\n\n Gillis lies on it, fully clothed, reading a book.\n\n GILLIS\n\n Yes, Norma.\n\n Through the slit in the door there is a suggestion\n\n of Norma.\n\n NORMA\n\n Don't turn around. Keep your\n\n eyes on the book.\n\n GILLIS\n\n Yes, Norma.\n\n Norma pushes the door open and comes in.\n\n NORMA\n\n I just came to say good night.\n\n I don't want you to see me --\n\n I'm not very attractive.\n\n GILLIS\n\n Good night.\n\n NORMA\n\n I've lost half a pound since\n\n Tuesday.\n\n GILLIS\n\n Good.\n\n NORMA\n\n I was a little worried about the\n\n line of my throat. This woman\n\n has done wonders with it.\n\n GILLIS\n\n Good.\n\n NORMA\n\n You'd better get to bed yourself.\n\n GILLIS\n\n I think I'll read a little.\n\n NORMA\n\n You went out last night, didn't\n\n you, Joe?\n\n GILLIS\n\n Why do you say that?\n\n NORMA\n\n I just happen to know it. I had\n\n a nightmare and I screamed for\n\n you. You weren't here. Where\n\n were you?\n\n GILLIS\n\n I went for a walk.\n\n NORMA\n\n No you didn't. You took the\n\n car.\n\n GILLIS\n\n All right, I drove to the beach.\n\n Norma, you don't want me to feel\n\n I'm locked up in this house?\n\n NORMA\n\n Of course not, Joe. It's just\n\n that I don't want to be left alone.\n\n Not now, while I'm under this\n\n terrible strain. My nerves are\n\n being torn apart. All I ask is\n\n for you to be a little patient and a\n\n little kind.\n\n GILLIS\n\n I haven't done anything, Norma.\n\n NORMA\n\n Of course you haven't. I wouldn't\n\n let you.\n\n She bends and kisses the top of his head.\n\n NORMA\n\n Good night, my darling.\n\n She goes into her room, shutting the door behind her.\n\n Gillis puts his book down and looks at her door.\n\n E-5 THE DOOR TO NORMA'S ROOM\n\n The light can be seen through the gouged-out\n\n keyhole. It goes out.\n\n DISSOLVE TO:\n\n E-6 UPPER LANDING STAIRWAY\n\n AND HALL BELOW (NIGHT) GILLIS' VOICE\n\n Gillis, with his coat on by Yes, I was playing hooky\n\n now, comes cautiously to\n\n the upper railing and looks every evening along in\n\n down into the lighted hall\n\n below. there. It made me think I\n\n Max is just extinguishing of when I was twelve and\n\n the lights. Max exits in,\n\n the direction of the liv- used to sneak out on the\n\n ing room.\n\n folks to see a gangster\n\n After a moment Gillis starts\n\n silently down the stairs. picture. This time it\n\n wasn't to see a picture,\n\n E-7 LIVING ROOM\n\n it was to try and write\n\n (Lighted only by the last\n\n flicker of a fire on the one. That story of mine\n\n hearth). Max is putting a\n\n fire screen in front of Betty Schaerer had dug\n\n the fire. He hears some\n\n steps and the creak or the up kept going through\n\n main door being opened.\n\n He looks out and sees my head like a dozen\n\n locomotives...\n\n E-7a THE MAIN DOOR\n\n Gillis, in the moonlit porch,\n\n is closing the main door\n\n behind him.\n\n E-8 LIVING ROOM\n\n Max looks after Gillis, his\n\n face enigmatic as ever.\n\n DISSOLVE TO:\n\n E-9 GARAGE AND DRIVEWAY\n\n (MOONLIGHT)\n\n Gillis comes into the shot,\n\n gets into the Isotta, drives\n\n it out or the garage and down\n\n the driveway to Sunset, as\n\n quietly as possible.\n\n DISSOLVE TO:\n\n E-10 READERS' OFFICE BUILDING\n\n PARAMOUNT (NIGHT)\n\n Start on a LONG SHOT. THE GILLIS' VOICE\n\n BOOM MOVES FORWARD to the only So we'd started\n\n two lights. They are the door working on it, the\n\n and window of Betty Schaefer's two of us. Nights,\n\n cubicle. Betty sits at the when the studio was\n\n desk, typing. Gillis, his deserted, up in her\n\n coat off, his shirt-sleeves little cubby-hole\n\n rolled up, j.s pacing the floor, of an office.\n\n discussing the construction of\n\n a sentence. The discussion at\n\n a stalemate, Betty suggests\n\n some coffee. Gillis agrees.\n\n From the electric plate on the\n\n shelf beside her, Betty takes\n\n a glass coffee machine. Gillis\n\n seats himself in her chair\n\n and starts typing.\n\n Betty opens the door and comes out on the balcony to\n\n fill the coffee machine from the water cooler stand-\n\n ing beside the door.\n\n BETTY\n\n I got the funniest letter from\n\n Artie. It's rained every day\n\n since they got to Arizona. They\n\n re-wrote the whole picture for\n\n rain and shot half of it. Now\n\n the sun is out. Nobody knows\n\n when they'll get back.\n\n She moves back into the room.\n\n GILLIS\n\n Good.\n\n BETTY\n\n What's good about it? I miss\n\n him something fierce.\n\n GILLIS\n\n I mean this is good dialogue\n\n along in here. It'll play.\n\n BETTY\n\n It will?\n\n GILLIS\n\n Sure. Especially with lots\n\n of music underneath, drowning\n\n it out.\n\n BETTY\n\n Don't you sometimes hate yourself?\n\n GILLIS\n\n Constantly. No, in all serious-\n\n ness, it's really good. It's\n\n fun writing again. I'm happy\n\n here, honest I am.\n\n He resumes typing. Betty puts the water on. She\n\n picks up a pack of cigarettes on the desk, finds it's\n\n empty and throws it away, sees Gillis' open gold\n\n cigarette case and lighter on the table by the couch.\n\n Betty reaches for a cigarette. The inscription en-\n\n graved inside the case catches her eye. It reads:\n\n MAD ABOUT THE BOY --\n\n Norma\n\n BETTY\n\n Who's Norma?\n\n GILLIS\n\n Who's who?\n\n BETTY\n\n I'm sorry. I don't usually\n\n read private cigarette cases.\n\n GILLIS\n\n Oh, that. It's from a friend\n\n of mine. A middle-aged lady,\n\n very foolish and very generous.\n\n BETTY\n\n I'll say. This is solid gold.\n\n GILLIS\n\n I gave her some advice on an\n\n idiotic script.\n\n BETTY\n\n It's that old familiar story,\n\n you help a timid little soul\n\n across a crowded street. She\n\n turns out to be a multimillionaire\n\n and leaves you all her money.\n\n GILLIS\n\n That's the trouble with you\n\n readers. You know all the plots.\n\n Now suppose you proof-read page\n\n ten while the water boils.\n\n DISSILVE TO:\n\n E-11 AN EMPTY STREET AT THE GILLIS' VOICE\n\n PARAMOUNT STUDIO (NIGHT) Sometimes when we got\n\n stuck we'd make a\n\n Gillis and Betty are walking litte tour of the\n\n down it. From a stage where drowsing lot, not talk-\n\n they are erecting a new set ing much, just wandering\n\n comes a great shaft of light. down alleys between the\n\n They stop at an apple-vending sound stages, or through\n\n machine in the foreground,buy the sets they were get-\n\n themselves a couple of apples ting ready for the next\n\n and walk on. day's shooting. As a\n\n matter of fact, it was\n\n DISSOLVE TO: on one of those walks\n\n when she first told me\n\n about her nose ...\n\n E-12 PARAMOUNT'S NEW YORK STREET (NIGHT)\n\n Betty and Gillis are walking down it, THE CAMERA\n\n AHEAD OF THEM.\n\n BETTY\n\n Look at this street. All card-\n\n board, all hollow, all phoney.\n\n All done with mirrors. I like\n\n it better than any street in the\n\n world. Maybe because I used to\n\n play here when I was a kid.\n\n GILLIS\n\n What were you -- a child actress?\n\n BETTY\n\n I was born just two blocks from\n\n this studio. Right on Lemon Grove\n\n Avenue. Father was head elec-\n\n trician here till he died. Mother\n\n still works in Wardrobe.\n\n GILLIS\n\n Second generation, huh?\n\n BETTY\n\n Third. Grandma did stunt work\n\n for Pearl White. I come from a\n\n picture family. Naturally they\n\n took it for granted I was to become\n\n a great star. So I had ten years of\n\n dramatic lessons, diction, dancing.\n\n Then the studio made a test. Well,\n\n they didn't like my nose -- it slanted\n\n this way a little. I went to a doctor\n\n and had it fixed. They made more\n\n tests, and they were crazy about my\n\n nose -- only they didn't like my acting.\n\n GILLIS\n\n (Examining her nose\n\n by the flame of his\n\n lighter)\n\n Nice job.\n\n BETTY\n\n Should be. It cost three hundred\n\n dollars.\n\n GILLIS\n\n Saddest thing I ever heard.\n\n BETTY\n\n Not at all. It taught me a little\n\n sense. I got me a job in the mail\n\n room, worked up to the Stenographic.\n\n Now I'm a reader...\n\n GILLIS\n\n Come clean, Betty. At night you\n\n weep for those lost closeups, those\n\n gala openings...\n\n BETTY\n\n Not once. What's wrong with being\n\n on the other side of the cameras?\n\n It's really more fun.\n\n GILLIS\n\n Three cheers for Betty Schaefer!\n\n I will now kiss that nose of yours.\n\n BETTY\n\n If you please.\n\n Gillis kisses her nose. As he stands there, his\n\n face close to hers -\n\n GILLIS\n\n May I say you smell real special.\n\n BETTY\n\n It must be my new shampoo.\n\n GILLIS\n\n That's no shampoo. It'smore like\n\n a pile of freehly laundred hand-\n\n kerchiefs, like a brand new auto-\n\n mobile. How old are you anyway?\n\n BETTY\n\n Twenty-two.\n\n GILLIS\n\n That's it -- there's nothing like\n\n being twenty-two. Now may I suggest\n\n that if we're ever to finish this\n\n story you keep at least two feet\n\n away from me ... Now back to the\n\n typewriter.\n\n They start walking in the direction of the office.\n\n DISSOLVE TO:\n\n E-13 THE GARAGE\n\n Gillis gets out. From the seat next him he takes a\n\n batch of script, folds it and puts it in his pocket.\n\n He suddenly becomes aware that he is watched, turns.\n\n Max stands in the moonlight, evidently waiting for\n\n him.\n\n GILLIS\n\n What is it, Max? Want to wash\n\n the car, or are you doing a little\n\n spying in your off hours?\n\n MAX\n\n You must be very careful as you\n\n cross the patio. Madame may be\n\n watching.\n\n GILLIS\n\n How about my going up the kitchen\n\n stairs and undressing in the dark.\n\n Will that do it?\n\n MAX\n\n I'm not inquiring where Mr.\n\n Gillis goes every night...\n\n GILLIS\n\n Why don't you? I'm writing a\n\n script and I'm dying to finish\n\n it, no matter what.\n\n MAX\n\n It's just that I'm very worried\n\n about Madame.\n\n GILLIS\n\n Sure you are. And we're not help-\n\n ing her any, feeding her lies and\n\n more lies. Getting herself ready\n\n for a pioture ... What happens when\n\n she finds out?\n\n MAX\n\n She never will. That is my job.\n\n It has been for a long time. You\n\n must understand I discovered her\n\n when she was eighteen. I made her\n\n a star. I cannot let her be destroyed.\n\n GILLIS\n\n You made her a star?\n\n MAX\n\n I directed all her early pictures.\n\n There were three young directors\n\n who showed promise in those days:\n\n D.W. Grirrith, C.B. deMille, and\n\n Max von Mayerling.\n\n GILLIS\n\n And she's turned you into a\n\n servant.\n\n MAX\n\n It was I who asked to come back,\n\n humiliating as it may seem. I\n\n could have gone on witn my career,\n\n only I found everything unendur-\n\n able arter she divorced me. You\n\n see, I was her rirst husband.\n\n DISSOLVE TO:\n\n E-14 NORMA DESMOND'S BEDROOM\n\n One lamp lit. Norma, in a white negligee, with the\n\n patches on her face, is pacing up and down -- a\n\n small, tormented, pitiable woman. Finally she opens\n\n the door to:\n\n E-15 GILLIS' ROOM (MOONLIGHT)\n\n Gillis lies in bed asleep, Norma in the doorway.\n\n NORMA\n\n You're here, Joe ... When did\n\n you come home? Where were you?\n\n Is it a woman? I know it's a\n\n woman ... Who is she? Oh Joe,\n\n why can't I ask you? I must know,\n\n I must!\n\n Her eyes fall on Gillis' coat, which hangs over a\n\n chair. In a pocket is part of the script. Norma\n\n takes it out, looks at it. She can't see it in the\n\n moonlight. She hurries with it into:\n\n E-16 NORMA'S BEDROOM\n\n Carrying the script Norma goes to the lamp and looks\n\n at it. On the first page she sees something which\n\n confirms all her suspicionso It reads:\n\n UNTITLED LOVE STORY\n\n by\n\n Joseph C. Gilliss\n\n and\n\n Betty Schaefer\n\n DISSOLVE:\n\n E-17 BETTY'S CUBICLE (NIGHT)\n\n Betty is typing. Gillis sits on the couch, proof-\n\n reading a scene. Betty stops typing and Gillis\n\n becomes aware of her eyes fixed on him.\n\n GILLIS\n\n Hey, what's the matter...\n\n Betty, wake up!\n\n (He whistles and\n\n catches her attention)\n\n Why are you staring at me like that?\n\n BETTY\n\n Was I? I'm sorry.\n\n GILLIS\n\n What's wrong with you tonight?\n\n What is it, Betty?\n\n BETTY\n\n Something came up. I don't want\n\n to talk about it.\n\n GILLIS\n\n Why not?\n\n BETTY\n\n I just don't.\n\n GILLIS\n\n What is it you've heard. Come\n\n on, let's have it.\n\n Betty gets up.\n\n GILLIS\n\n Is it about me?\n\n Betty doesn't answer, walks out on\n\n E-18 THE BALCONY\n\n She leans against a post, crying. Gillis comes out\n\n after her.\n\n GILLIS\n\n Betty, there's no use running\n\n out on it. Let's face it, what-\n\n ever it is.\n\n BETTY\n\n It's nothing. I got a telegram\n\n from Artie.\n\n GILLIS\n\n From Artie. What's wrong?\n\n BETTY\n\n He wants me to come on to Arizona.\n\n He says it only oosts two dollars\n\n to get married there. It would\n\n kind of save us a honeymoon.\n\n GILLIS\n\n Why don't you? We can finish the\n\n script by Thursday.\n\n Betty stands crying silently.\n\n GILLIS\n\n Stop crying. You're getting\n\n married. That's what you've\n\n always wanted.\n\n BETTY\n\n I don't want it now.\n\n GILLIS\n\n Why not? Don't you love Artie?\n\n BETTY\n\n Of course I love him. I always\n\n will. I'm just not in love\n\n with him any more.\n\n GILLIS\n\n What happened?\n\n BETTY\n\n You did.\n\n There is a moment's pause before he takes her in\n\n his arms. THE CAMERA MOVES AWAY.\n\n DISSOLVE TO:\n\n E-19 HALL AND STAIRCASE GILLIS' VOICE\n\n DESMOND HOME- (NIGHT) It wasn' t until I got\n\n back to that peculiar\n\n Gillis enters, closes prison of mine that I\n\n the door as quietly as started facing the facts.\n\n he can, and goes up There it was -- Betty\n\n the stairs. Schaefer's future right\n\n in the palm of my hand.\n\n E-20 GILLIS' ROOM Betty Schaefer engaged\n\n to Artie Green, as nice\n\n He enters and turns on the a guy as ever lived.\n\n light. He sinks down on And she was in love with\n\n the chaise longue,thinking. me. Me ! She was a fool\n\n His eyes wander to the not to sense that there\n\n door of Norma's room. was something phony in\n\n Through the gouged-out key- my set-up. And I was a\n\n hole he sees the light. heel not to have told\n\n her. But you just can't\n\n say those things to\n\n somebody you're crazy\n\n about. Maybe I'd never\n\n have to. Maybe I could\n\n get away with it, get\n\n away from Norma. Maybe\n\n I could wipe the whole\n\n nasty mess right out of\n\n my life...\n\n From Norma's room comes the sound of a telephone\n\n being dialled. Gillis enters the shot and stands\n\n listening.\n\n NORMA'S VOICE\n\n Is this Gladstone 0858?\n\n E-21 NORMA'S BEDROOM\n\n Norma lies in bed, dialing a number. She has the\n\n beauty patches at the corners of her eyes and over\n\n her nose.\n\n NORMA\n\n Can I speak to Miss Betty\n\n Schaefer? She must be home by\n\n now.\n\n E-22 A BEDROOM IN BETTY'S FLAT\n\n Connie, a girl of Betty's age with whom she shares\n\n the flat, is on the phone. Betty, in a dressing-\n\n gown, comes from the bathroom, toothbrush in hand.\n\n CONNIE\n\n (Hand over mouthpiece)\n\n Betty, here's that weird-sounding\n\n woman again.\n\n BETTY\n\n What is this anyway?\n\n (Taking the phone)\n\n This is Betty Schaefer.\n\n E-23 NORMA AT IHE PHONE\n\n NORMA\n\n Miss Schaefer, you must forgive\n\n me for calling you so late, but\n\n I really feel it's my duty. It's\n\n about Mr. Gillis. You do know Mr.\n\n Gillis? ...Exactly how much do you\n\n know about him? Do you know where\n\n he lives? Do you know how he lives?\n\n Do you know what he lives on?\n\n E-24 BETTY AT THE PHONE\n\n BETTY\n\n Who are you? What do you want?\n\n What business is it of yours\n\n anyway?\n\n E-25 NORMA ON THE PHONE\n\n NORMA\n\n Miss Schaefer, I'm trying to do\n\n you a favor. I'm trying to spare\n\n you a great deal of misery. Of\n\n course you may be too young to even\n\n suspect there are men of his sort...\n\n NORMA (Cont'd)\n\n I don't know what he's told you, but\n\n he does not live with relatives, nor\n\n with friends, in the usual sense of\n\n the word. Ask him ... Ask him again.\n\n During the latter part of her call, the doors from\n\n Gillis' room have been pushed open and Gillis has\n\n walked towards her. Suddenly Norma senses his pre-\n\n sence and turns around. The telephone freezes in her\n\n hand. She tries to hang it up. Very calmly Gillis\n\n takes the receiver from her hand.\n\n GILLIS\n\n (Into phone)\n\n That's right, Betty, ask me again.\n\n This is Joe.\n\n E-26 BETTY ON THE PHONE\n\n BETTY\n\n Joe, where are you? What's this\n\n all about?\n\n E-27 GILLIS ON THE PHONE\n\n Norma beside him.\n\n GILLIS\n\n Or maybe it would be a better\n\n idea if you came over and saw it\n\n for yourself. The address is 10086\n\n .\n\n He hangs up. Norma looks up at him as he crosses to\n\n the other end of the room and stands staring at her.\n\n The silence becomes unbearable.\n\n NORMA\n\n Don't hate me, Joe. I did it because\n\n I need you. I need you as I never\n\n needed you. Look at me. Look at my\n\n hands, look at my face, look under my\n\n eyes. How can I go back to work if I'm\n\n wasting away under this torment? You\n\n don't know what I've been through these\n\n last weeks. I got myself a revolver.\n\n You don't believe me, but I did, I did!\n\n I stood in front of that mirror, only\n\n I couldn't make myself. It wouldn't be\n\n NORMA (Cont'd)\n\n fair to all those people who are\n\n waiting to see me back on the\n\n screen. I can't disappoint them.\n\n Only, if I'm to work, I need\n\n sleep, I need quiet, I need you!\n\n Don't just stand there hating\n\n me! Shout at me, strike me!\n\n But don't hate me, Joe. Don't\n\n you hear me, Joe?\n\n GILLIS\n\n Yes, I hear you. And I wish you'd\n\n keep still so I can hear the doorbell\n\n when she rings it.\n\n E-28 BETTY AND CONNIE, DRIVING IN A SMALL COUPE DOWN\n\n (NIGHT)\n\n E-29 INT. COUPE\n\n Connie is looking at the house numbers.\n\n CONNIE\n\n Here's ten thousand seventy-nine,\n\n Betty. It must be over there.\n\n Betty turns the car into the driveway of Norma's\n\n place, stops at the entrance steps. Betty gets out.\n\n CONNIE\n\n Betty, let me come along with\n\n you. Please.\n\n BETTY\n\n No, I'll be all right.\n\n She shuts the door of the car and goes up the steps.\n\n E-30 NORMA'S BEDROOM\n\n Norma lies on the bed. Gillis sits in a far corner\n\n of the room, motionless.\n\n NORMA\n\n (In a whimpering monotone)\n\n I love you, Joe. I love you, Joe.\n\n I love you, Joe. I love you, Joe.\n\n There is the sound of footsteps below and the ringing\n\n of a doorbell. Gillis rises.\n\n NORMA\n\n What are you going to do, Joe?\n\n Without a word, he leaves the room. Norma raises\n\n herself on the bed, reaching for a black negligee\n\n lying at the foot of it. As she does so, she dis-\n\n lodges her pillow a little, revealing a revolver\n\n hidden beneath it.\n\n E-31 DOWNSTAIRS HALL, THE DESMOND HOUSE (DARK)\n\n Max crosses the hall, putting on his alpaca jacket.\n\n He turns on the lights. Outside stands Betty.\n\n From the staircase comes -\n\n GILLIS' VOICE\n\n It's all right, Max. I'll take it.\n\n MAX\n\n Yes, sir.\n\n He stands back as Gillis opens the door.\n\n GILLIS\n\n Hello, Betty.\n\n BETTY\n\n (On the threshold)\n\n I don't know why I'm so scared,\n\n Joe. Is it something awful?\n\n GILLIS\n\n Come on in, Betty,\n\n Betty enters. As he leads her into the living room,\n\n Gillis puts his arm around her shoulders.\n\n GILLIS\n\n Ever been in one of these old\n\n Hollywood palazzos? That's from\n\n when they were making eighteen thou-\n\n sand a week, and no taxes. Careful\n\n of these tiles, they're slippery.\n\n Valentino used to dance here.\n\n BETTY\n\n This is where you live?\n\n GILLIS\n\n You bet.\n\n BETTY\n\n Whose house is it?\n\n They have reached\n\n E-32 THE LIVING ROOM\n\n Gillis leads Betty in.\n\n GILLIS\n\n Hers.\n\n BETTY\n\n Whose?\n\n GILLIS\n\n Just look around. There's a lot\n\n of her spread about. If you don't\n\n remember the face, you must have\n\n heard the name of Norma Desmond.\n\n BETTY\n\n That was Norma Desmond on the phone?\n\n GILLIS\n\n Want something to drink? There's\n\n always champagne on ice, and plenty\n\n of caviar.\n\n BETTY\n\n Why did she call me?\n\n GILLIS\n\n Jealous. Ever see so much junk?\n\n She had the ceiling brought from\n\n Portugal. Look at this.\n\n He pulls the rope, showing the projection screen\n\n under the picture.\n\n GILLIS\n\n Her own movie theatre.\n\n BETTY\n\n I didn't come here to see a house.\n\n What about Norma Desmond?\n\n GILLIS\n\n I'm trying to tell you. This is\n\n an enormous place. Eight master\n\n bedrooms. A sunken tub in every\n\n bathroom. There's a bowling alley\n\n in the cellar. It's lonely here,\n\n so she got herself a companion.\n\n A very simple set-up: An older\n\n woman who is well-to-do. A younger\n\n man who is not doing too well ...\n\n Can you figure it out yourself?\n\n BETTY\n\n No.\n\n GILLIS\n\n All right. I'll give you a few\n\n more clues.\n\n BETTY\n\n No, no! I haven't heard any of\n\n this. I never got those telephone\n\n calls. I've never been in this\n\n house ... Get your things together.\n\n Let's get out of here.\n\n GILLIS\n\n All my things? All the eighteen\n\n suits, all the custom-made shoes and\n\n the eighteen dozen shirts, and the\n\n cuff-links and the platinum key-\n\n chains, and the cigarette cases?\n\n BETTY\n\n Come on, Joe.\n\n GILLIS\n\n Come on where? Back to a one-room\n\n apartment that I can't pay for?\n\n Back to a story that may sell and\n\n very possibly will not?\n\n BETTY\n\n If you love me, Joe.\n\n GILLIS\n\n Look, sweetie -- be practical.\n\n l've got a good thing here.\n\n A long-term contract with no options.\n\n I like it that way. Maybe it's not\n\n very admirable. Well, you and Artie\n\n can be admirable.\n\n BETTY\n\n Joe, I can't look at you any more.\n\n GILLIS\n\n Nobody asked you to.\n\n Betty turns from him, to hide the fact that she is\n\n crying.\n\n GILLIS\n\n All right, baby. This way out.\n\n He leads her in the direction of the door.\n\n E-33 UPPER LANDING, DESMOND HOUSE\n\n Sitting crouched behind the balustrade is Norma,\n\n peering down into\n\n E-34 THE LOWER HALL\n\n Betty and Gillis have reached the entrance door.\n\n Gillis opens it.\n\n GILLIS\n\n Good luck to you, Betty. You can\n\n finish that story on the way to\n\n Arizona. When you and Artie get\n\n back, if the two of you ever feel\n\n like a swim, here's the pool ...\n\n He switches on the light.\n\n E-35 THE PATIO\n\n The lights go on in the pool, which shines brilliant-\n\n ly in the dark garden.\n\n E-36 BETTY\n\n She doesn't even look. Her eyes filled with tears,\n\n she runs down the entrance porch toward her car.\n\n E-37 THE ENTRANCE HALL\n\n Gillis looks after her, closes the door. From the\n\n upper landing comes the sound of soft sobbing. He\n\n looks up.\n\n E-38 NORMA, ON THE UPPER LANDING\n\n Gillis ascends the stairs.\n\n NORMA\n\n Thank you, Joe -- thank you, Joe.\n\n She tries to take his hand to kiss it as he passes.\n\n He doesn't stop. Norma catches his coat. Gillis\n\n moves right on into his room. Norma lies on the\n\n floor looking after him. She crawls toward a con-\n\n sole, pulls herself up by it, starts towards Gillis'\n\n door, passes a mirror, realizes how she looks, moves\n\n back to the mirror and takes the patches off her\n\n face and does a hasty job of removing the cream with\n\n her handkerchief, readjusts her expression to a poor\n\n travesty of a smile and goes to the door of Gillis'\n\n room.\n\n NORMA\n\n May I come in? I've stopped cry-\n\n ing. I'm all right again. Joe,\n\n tell me you're not cross -- tell\n\n me everything is just as it was,\n\n Joe.\n\n She opens the door.\n\n E-39 GILLIS' ROOM\n\n In the foreground, open on the bed, is a half-packed\n\n suitcase, Gillis just putting some of his old shirts\n\n in. Norma stands staring, speechless, for a second.\n\n Gillis moves out of the shot towards the closets.\n\n NORMA\n\n What are you doing, Joe? What\n\n are you doing? You're not leaving\n\n me?\n\n GILLIS\n\n Yes, I am, Norma.\n\n NORMA\n\n No, you're not.\n\n (Calling)\n\n Max! Max!\n\n GILLIS\n\n Max is a good idea. He can help\n\n with my luggage.\n\n (He gestures in the\n\n direction of the closet)\n\n Thanks for letting me wear the\n\n handsome wardrobe. And thanks\n\n for the use of all the trinkets.\n\n He takes the cigarette case and throws it on the\n\n chaise longue. Then he throws the lighter, the\n\n wrist watch, the platinum key-chain and the tie clip.\n\n GILLIS\n\n (Indicating the bureau)\n\n The rest of the jewelry is in the\n\n top drawer.\n\n NORMA\n\n It's yours, Joe. I gave it to\n\n you.\n\n GILLIS\n\n And I'd take it in a second, Norma --\n\n only it's a little too dressy for\n\n sitting behind the copy desk in\n\n Dayton, Ohio.\n\n NORMA\n\n These are nothing. You can have\n\n anything you want if you'll only\n\n stay. What is it you want --\n\n money?\n\n GILLIS\n\n Norma, you'd be throwing it away.\n\n I don't qualify for the job, not any\n\n more.\n\n NORMA\n\n You can't do this! Max! Max!\n\n ... I can't face life without you,\n\n and I'm not afraid to die, you\n\n know.\n\n GILLIS\n\n That's between you and yourself,\n\n Norma.\n\n NORMA\n\n You think I made that up about\n\n the gun...\n\n She rushes into her room. Gillis closes the suitcase\n\n calmly, notices that he is still wearing some cuff-\n\n links Norma gave him, takes them off.\n\n Norma reappears in the door, carrying the revolver.\n\n NORMA\n\n See, you didn't believe me!..\n\n Now I suppose you don't think I\n\n have the courage!\n\n GILLIS\n\n Oh. sure -- if it would make a\n\n good scene.\n\n NORMA\n\n You don't care. do you? But\n\n hundreds of thousands of people\n\n will carel\n\n GILLIS\n\n Wake up, Norma. You'd be killing\n\n yourself to an empty house. The\n\n audience left twenty years ago.\n\n Now face it.\n\n During the preceding. Max has entered. He stands\n\n listening, paralyzed.\n\n NORMA\n\n That's a lie! They still want me!\n\n GILLIS\n\n No, they don't.\n\n NORMA\n\n What about the studio?\n\n What about De Mille?\n\n GILLIS\n\n He was trying to spare your feelings.\n\n The studio wanted to rent your car.\n\n NORMA\n\n Wanted what?\n\n GILLIS\n\n De Mille didn't have the heart\n\n to tell you. None of us has had\n\n the heart.\n\n NORMA\n\n That's a lie! They want me, they\n\n want me! I get letters every day!\n\n GILLIS\n\n You tell her, Max. Come on, do\n\n her that favor. Tell her there\n\n isn't going to be any picture --\n\n there aren't any fan letters,\n\n except the ones you write yourself.\n\n NORMA\n\n That isn't true! Max?\n\n MAX\n\n Madame is the greatest star of\n\n them all... I will take Mr.\n\n Gillis' bags.\n\n He leaves.\n\n NORMA\n\n You heard him. I'm a star!\n\n GILLIS\n\n Norma, grow up. You're a woman\n\n of fifty. There's nothing tragic\n\n about being fifty - not unless\n\n you try to be twenty-five.\n\n NORMA\n\n I'm the greatest star of them\n\n all.\n\n GILLIS\n\n Goodbye. Norma.\n\n NORMA\n\n No one leaves a star. That\n\n makes one a star.\n\n Gillis picks up the typewriter and leaves.\n\n NORMA\n\n You're not leaving me!\n\n E-40 STAIRCASE\n\n Gillis descending with the typewriter.\n\n NORMA'S VOICE\n\n Joe! ...Joe!\n\n There is the SOUND OF A SHOT. The glass of the front\n\n door is shattered. Gillis at the door opens it and\n\n walks out, without looking back.\n\n Down the staircase rushes Norma. a disordered wild-\n\n ness in the way she moves.\n\n NORMA\n\n You're not leaving me!\n\n She hurries after Gillis.\n\n E-41 PATIO (NIGHT)\n\n Dark except for lights from the house and the\n\n luminousness of the lit pool.\n\n Gillis is crossing the patio towards the garage. He\n\n is carrying the typewriter. He doesn't accelerate\n\n his step, although he has heard the shot. Behind\n\n him Norma comes from the lighted house.\n\n NORMA\n\n You're not leaving me!\n\n She shoots twice in rapid succession. Gillis drops\n\n the typewriter. The shots have swung him around. He\n\n is now facing Norma. She shoots him. This shot\n\n hits him in the belly. He doubles up, instinctively\n\n backs away from her, plummets into the lit pool.\n\n Up the stone steps from the garage rushes Max.\n\n He sees the situation, hurries towards Norma, who\n\n stands exultant in the strange light from the pool.\n\n NORMA\n\n Stars are ageless, aren't they?\n\n DISSOLVE TO:\n\n E-42 THE PATIO\n\n Dawn is breaking. At the edge of the pool\n\n stand policemen, detectives and police photographers.\n\n Motorcycle policemen are holding off the mob which\n\n is trying to storm the house.\n\n A lietuenant from the Homicide Bureau leaves the\n\n crowd around the pool and goes into\n\n E-43 THE LOWER HALL, DESMOND HOUSE\n\n It is filled with a pandemonium of police officers,\n\n newspaper people, etc. who are kept from the upper\n\n floor by two policemen at the head of the stairs.\n\n The lieutenant from the Homicide Bureau goes\n\n through the crowd to the telephone at the foot of\n\n the stairs, picks up the phone and dials.\n\n LIEUTENANT\n\n Coroner's office? ... I want to\n\n speak to the Coroner ... Who's\n\n on this phone?\n\n E-44 THE WHITE TELEPHONE IN NORMA'S BEDROOM\n\n Standing talking into it is Hedda Hopper.\n\n MISS HOPPER\n\n I am! Now get off, this is more\n\n important ... Times City Desk?\n\n Hedda Hopper speaking. I'm talking\n\n from the bedroom of Norma Desmond.\n\n Don't bother with a rewrite man, take\n\n this direct. Ready? -- As day breaks\n\n over the murder house, Norma Desmond,\n\n famed star of yesteryear, is in\n\n a state of complete mental shock ...\n\n THE CAMERA PANS TO ANOTHER PART OF THE BEDROOM, where\n\n Norma sits at a mirror, staring at herself blankly.\n\n Firing questions at her are the Captain of the Holmby\n\n Hills Division and the L.A. Homicide Squad. Max\n\n stands by faithfully.\n\n HOLMBY HILLS CAPTAIN\n\n You do not deny having killed\n\n this man, Miss Desmond?\n\n HEAD OF HOMICIDE\n\n Did you intend to kill him?\n\n Just answer me that.\n\n HOLMBY HILLS CAPTAIN\n\n Was it a sudden quarrel? Had there\n\n been any trouble between you before?\n\n HEAD OF HOMICIDE\n\n If it was a quarrel, how come you\n\n had the gun right there?\n\n HOLMBY HILLS CAPTAIN\n\n This guy -- where did you meet him\n\n for the first time? Where did he\n\n come from? Who is he?\n\n HEAD OF HOMICIDE\n\n Did he have a wife? Did he had a\n\n girl friend? Did you know them?\n\n HOLMBY HILLS CAPTAIN\n\n Had he been trying to blackmail you?\n\n E-45 PATIO - (DAWN) GILLIS' VOICE\n\n The body of Gillis Well, this is where you came.\n\n being fished from Here's that pool again,the one\n\n the pool, put on a I always wanted. They must have\n\n stretcher, covered photographed me a hundred times.\n\n with an army blanket.Then they got a couple of prun-\n\n Two men from the ing hooks from the garden and\n\n Coroner's office fished me out ever so gently.\n\n carry it towards Funny how gentle people get with\n\n the Coroner's you once you're dead. They\n\n hearse, CAMERA beached me, like a harpooned\n\n PANNING with them. baby whale, and started to check\n\n the damage, just for the record\n\n ... By this time the whole joint\n\n was jumping -- cops,reporters,\n\n neighbors, passersby -- as much\n\n hoopdedoo as we get in Los\n\n Angeles when they open a Super\n\n Market. Even the newsreel guys\n\n came roaring in. Here was an\n\n item everybody could have some\n\n fun with, the heartless so-and-\n\n so's. What would they do to her?\n\n Even if she got away with it in\n\n court- crime of passion - tempo-\n\n rary insanity - those headlines\n\n would kill her: Forgotten Star\n\n a Slayer--Aging Actress--\n\n Yesterday's Glamour Queen...\n\n E-46 NORMA'S BEDROOM\n\n The interrogators are still firing questions at Norma\n\n who sits lifeless, staring at herself. Max watches.\n\n HEAD OF HOMICIDE\n\n Did the deceased ever threaten you?\n\n Were you in fear of bodily injury?\n\n HOLMBY HILLS CAPTAIN\n\n Did you hate him? Had you ever thought\n\n of doing something like this before?\n\n HEAD OF HOMICIDE\n\n Was theft involved? Did you catch\n\n him trying to steal something, or\n\n find he had stolen something?\n\n A police lieutenant has entered, goes to the Head of\n\n Homicide.\n\n LIEUTENANT\n\n The newsreel guys have arrived with\n\n the cameras.\n\n HEAD OF HOMICIDE\n\n Tell them to go fly a kite. This\n\n is no time for cameras.\n\n A word has pierced the mists that surround Norma.\n\n NORMA\n\n Cameras? ...What is it, Max?\n\n MAX\n\n The cameras have arrived, Madame.\n\n NORMA\n\n They have? Thank you, Max. Tell\n\n Mr. DeMille I will be on the set\n\n at once.\n\n Max flashes a look at the Head of Homicide.\n\n HEAD OF HOMICIDE\n\n What is this?\n\n MAX\n\n Please ...\n\n HOLMBY HILLS CAPTAIN\n\n (sotto voce, to Head of Homicide)\n\n Well, it's one way to get her down stairs.\n\n HEAD OF HOMICIDE\n\n Okay. And let's have the car right\n\n outside.\n\n 7-1 NORMA\n\n You will pardon me, gentlemen.\n\n I have to get ready for my scene.\n\n She takes a comb and runs it through her hair, then\n\n starts applying some wild makeup.\n\n E-47 STAIRCASE AND LOWER HALL\n\n Max makes his way down the stairs through the crowd\n\n of newsmen to the newsreel cameras, which are being\n\n set up in the hall below.\n\n MAX\n\n Is everything set up, gentlemen?\n\n Are the lights ready?\n\n From the stairway comes a murnur. They look up.\n\n Norma has emerged from the bedroom and comes to the\n\n head of the stairs. There are golden spangles in\n\n her hair and in her hand she carries a golden scarf.\n\n The police clear a path for her to descend. Press\n\n cameras flash at her every step.\n\n Max stands at the cameras.\n\n MAX\n\n Is everything set up, gentlemen?\n\n CAMERAMAN\n\n Just about.\n\n The portable lights flare up and illuminate the\n\n staircase.\n\n MAX\n\n Are the lights ready?\n\n 2ND CAMERA MAN\n\n All set.\n\n MAX\n\n Quiet, everybody! Lights!\n\n Are you ready, Norma?\n\n NORMA\n\n (From the top of the\n\n stairs)\n\n What is the scene? Where am I?\n\n MAX\n\n This is the staircase of the palace.\n\n NORMA\n\n Oh, yes, yes. They're below,\n\n waiting for the Princess ...\n\n I'm ready.\n\n MAX\n\n All right.\n\n (To cameramen)\n\n Camera!\n\n (To Norma)\n\n Action!\n\n Norma arranges the golden GILLIS' VOICE\n\n scarf ebout her and proudy So they were grinding\n\n starts to descend the stair- after all, those cam-\n\n case. The cameras grind. eras. Life, which can\n\n Everyone watches in awe. be strangely merciful,\n\n had taken pity on Norma\n\n Desmond. The dream she\n\n had clung to so des-\n\n perately had enfolded\n\n her...\n\n At the foot of the stairs Norma stops, moved.\n\n NORMA\n\n I can't go on with the scene.\n\n I'm too happy. Do you mind,\n\n Mr. DeMille, if I say a few words?\n\n Thank you. I just want to tell\n\n you how happy I am to be back in\n\n the studio making a picture again.\n\n You don't know how much I've missed\n\n all of you. And I promise you\n\n I'll never desert you again, because\n\n after "Salome" we'll make another\n\n picture, and another and another.\n\n You see, this is my life. It always\n\n will be. There's nothing else -\n\n just us and the cameras and those\n\n wonderful people out there in the\n\n dark... All right, Mr. DeMille,\n\n I'm ready for my closeup.\n\n FADE OUT.\n\n THE END\n\n |
| 18 | The Most Dangerous Game--Richard Connell.txt | af2b1960-5ca | literature | The Most Dangerous Game\n\nby Richard Connell (1893-1949)\n\n"OFF THERE to the right--somewhere--is a large island," said Whitney." It's rather a mystery--"\n\n"What island is it?" Rainsford asked.\n\n"The old charts call it 'Ship-Trap Island,"' Whitney replied." A suggestive name, isn't it? Sailors have a curious dread of the place. I don't know why. Some superstition--"\n\n"Can't see it," remarked Rainsford, trying to peer through the dank tropical night that was palpable as it pressed its thick warm blackness in upon the yacht.\n\n"You've good eyes," said Whitney, with a laugh," and I've seen you pick off a moose moving in the brown fall bush at four hundred yards, but even you can't see four miles or so through a moonless Caribbean night."\n\n"Nor four yards," admitted Rainsford. "Ugh! It's like moist black velvet."\n\n"It will be light enough in Rio," promised Whitney. "We should make it in a few days. I hope the jaguar guns have come from Purdey's. We should have some good hunting up the Amazon. Great sport, hunting."\n\n"The best sport in the world," agreed Rainsford.\n\n"For the hunter," amended Whitney. "Not for the jaguar."\n\n"Don't talk rot, Whitney," said Rainsford. "You're a big-game hunter, not a philosopher. Who cares how a jaguar feels?"\n\n"Perhaps the jaguar does," observed Whitney.\n\n"Bah! They've no understanding."\n\n"Even so, I rather think they understand one thing--fear. The fear of pain and the fear of death."\n\n"Nonsense," laughed Rainsford. "This hot weather is making you soft, Whitney. Be a realist. The world is made up of two classes--the hunters and the huntees. Luckily, you and I are hunters. Do you think we've passed that island yet?"\n\n"I can't tell in the dark. I hope so."\n\n"Why? " asked Rainsford.\n\n"The place has a reputation--a bad one."\n\n"Cannibals?" suggested Rainsford.\n\n"Hardly. Even cannibals wouldn't live in such a God-forsaken place. But it's gotten into sailor lore, somehow. Didn't you notice that the crew's nerves seemed a bit jumpy today?"\n\n"They were a bit strange, now you mention it. Even Captain Nielsen--"\n\n"Yes, even that tough-minded old Swede, who'd go up to the devil himself and ask him for a light. Those fishy blue eyes held a look I never saw there before. All I could get out of him was 'This place has an evil name among seafaring men, sir.' Then he said to me, very gravely, 'Don't you feel anything?'--as if the air about us was actually poisonous. Now, you mustn't laugh when I tell you this--I did feel something like a sudden chill.\n\n"There was no breeze. The sea was as flat as a plate-glass window. We were drawing near the island then. What I felt was a--a mental chill; a sort of sudden dread."\n\n"Pure imagination," said Rainsford.\n\n"One superstitious sailor can taint the whole ship's company with his fear."\n\n"Maybe. But sometimes I think sailors have an extra sense that tells them when they are in danger. Sometimes I think evil is a tangible thing--with wave lengths, just as sound and light have. An evil place can, so to speak, broadcast vibrations of evil. Anyhow, I'm glad we're getting out of this zone. Well, I think I'll turn in now, Rainsford."\n\n"I'm not sleepy," said Rainsford. "I'm going to smoke another pipe up on the afterdeck."\n\n"Good night, then, Rainsford. See you at breakfast."\n\n"Right. Good night, Whitney."\n\nThere was no sound in the night as Rainsford sat there but the muffled throb of the engine that drove the yacht swiftly through the darkness, and the swish and ripple of the wash of the propeller.\n\nRainsford, reclining in a steamer chair, indolently puffed on his favorite brier. The sensuous drowsiness of the night was on him." It's so dark," he thought, "that I could sleep without closing my eyes; the night would be my eyelids--"\n\nAn abrupt sound startled him. Off to the right he heard it, and his ears, expert in such matters, could not be mistaken. Again he heard the sound, and again. Somewhere, off in the blackness, someone had fired a gun three times.\n\nRainsford sprang up and moved quickly to the rail, mystified. He strained his eyes in the direction from which the reports had come, but it was like trying to see through a blanket. He leaped upon the rail and balanced himself there, to get greater elevation; his pipe, striking a rope, was knocked from his mouth. He lunged for it; a short, hoarse cry came from his lips as he realized he had reached too far and had lost his balance. The cry was pinched off short as the blood-warm waters of the Caribbean Sea dosed over his head.\n\nHe struggled up to the surface and tried to cry out, but the wash from the speeding yacht slapped him in the face and the salt water in his open mouth made him gag and strangle. Desperately he struck out with strong strokes after the receding lights of the yacht, but he stopped before he had swum fifty feet. A certain coolheadedness had come to him; it was not the first time he had been in a tight place. There was a chance that his cries could be heard by someone aboard the yacht, but that chance was slender and grew more slender as the yacht raced on. He wrestled himself out of his clothes and shouted with all his power. The lights of the yacht became faint and ever-vanishing fireflies; then they were blotted out entirely by the night.\n\nRainsford remembered the shots. They had come from the right, and doggedly he swam in that direction, swimming with slow, deliberate strokes, conserving his strength. For a seemingly endless time he fought the sea. He began to count his strokes; he could do possibly a hundred more and then--\n\nRainsford heard a sound. It came out of the darkness, a high screaming sound, the sound of an animal in an extremity of anguish and terror.\n\nHe did not recognize the animal that made the sound; he did not try to; with fresh vitality he swam toward the sound. He heard it again; then it was cut short by another noise, crisp, staccato.\n\n"Pistol shot," muttered Rainsford, swimming on.\n\nTen minutes of determined effort brought another sound to his ears--the most welcome he had ever heard--the muttering and growling of the sea breaking on a rocky shore. He was almost on the rocks before he saw them; on a night less calm he would have been shattered against them. With his remaining strength he dragged himself from the swirling waters. Jagged crags appeared to jut up into the opaqueness; he forced himself upward, hand over hand. Gasping, his hands raw, he reached a flat place at the top. Dense jungle came down to the very edge of the cliffs. What perils that tangle of trees and underbrush might hold for him did not concern Rainsford just then. All he knew was that he was safe from his enemy, the sea, and that utter weariness was on him. He flung himself down at the jungle edge and tumbled headlong into the deepest sleep of his life.\n\nWhen he opened his eyes he knew from the position of the sun that it was late in the afternoon. Sleep had given him new vigor; a sharp hunger was picking at him. He looked about him, almost cheerfully.\n\n"Where there are pistol shots, there are men. Where there are men, there is food," he thought. But what kind of men, he wondered, in so forbidding a place? An unbroken front of snarled and ragged jungle fringed the shore.\n\nHe saw no sign of a trail through the closely knit web of weeds and trees; it was easier to go along the shore, and Rainsford floundered along by the water. Not far from where he landed, he stopped.\n\nSome wounded thing--by the evidence, a large animal--had thrashed about in the underbrush; the jungle weeds were crushed down and the moss was lacerated; one patch of weeds was stained crimson. A small, glittering object not far away caught Rainsford's eye and he picked it up. It was an empty cartridge.\n\n"A twenty-two," he remarked. "That's odd. It must have been a fairly large animal too. The hunter had his nerve with him to tackle it with a light gun. It's clear that the brute put up a fight. I suppose the first three shots I heard was when the hunter flushed his quarry and wounded it. The last shot was when he trailed it here and finished it."\n\nHe examined the ground closely and found what he had hoped to find--the print of hunting boots. They pointed along the cliff in the direction he had been going. Eagerly he hurried along, now slipping on a rotten log or a loose stone, but making headway; night was beginning to settle down on the island.\n\nBleak darkness was blacking out the sea and jungle when Rainsford sighted the lights. He came upon them as he turned a crook in the coast line; and his first thought was that be had come upon a village, for there were many lights. But as he forged along he saw to his great astonishment that all the lights were in one enormous building--a lofty structure with pointed towers plunging upward into the gloom. His eyes made out the shadowy outlines of a palatial chateau; it was set on a high bluff, and on three sides of it cliffs dived down to where the sea licked greedy lips in the shadows.\n\n"Mirage," thought Rainsford. But it was no mirage, he found, when he opened the tall spiked iron gate. The stone steps were real enough; the massive door with a leering gargoyle for a knocker was real enough; yet above it all hung an air of unreality.\n\nHe lifted the knocker, and it creaked up stiffly, as if it had never before been used. He let it fall, and it startled him with its booming loudness. He thought he heard steps within; the door remained closed. Again Rainsford lifted the heavy knocker, and let it fall. The door opened then--opened as suddenly as if it were on a spring--and Rainsford stood blinking in the river of glaring gold light that poured out. The first thing Rainsford's eyes discerned was the largest man Rainsford had ever seen--a gigantic creature, solidly made and black bearded to the waist. In his hand the man held a long-barreled revolver, and he was pointing it straight at Rainsford's heart.\n\nOut of the snarl of beard two small eyes regarded Rainsford.\n\n"Don't be alarmed," said Rainsford, with a smile which he hoped was disarming. "I'm no robber. I fell off a yacht. My name is Sanger Rainsford of New York City."\n\nThe menacing look in the eyes did not change. The revolver pointing as rigidly as if the giant were a statue. He gave no sign that he understood Rainsford's words, or that he had even heard them. He was dressed in uniform--a black uniform trimmed with gray astrakhan.\n\n"I'm Sanger Rainsford of New York," Rainsford began again. "I fell off a yacht. I am hungry."\n\nThe man's only answer was to raise with his thumb the hammer of his revolver. Then Rainsford saw the man's free hand go to his forehead in a military salute, and he saw him click his heels together and stand at attention. Another man was coming down the broad marble steps, an erect, slender man in evening clothes. He advanced to Rainsford and held out his hand.\n\nIn a cultivated voice marked by a slight accent that gave it added precision and deliberateness, he said, "It is a very great pleasure and honor to welcome Mr. Sanger Rainsford, the celebrated hunter, to my home."\n\nAutomatically Rainsford shook the man's hand.\n\n"I've read your book about hunting snow leopards in Tibet, you see," explained the man. "I am General Zaroff."\n\nRainsford's first impression was that the man was singularly handsome; his second was that there was an original, almost bizarre quality about the general's face. He was a tall man past middle age, for his hair was a vivid white; but his thick eyebrows and pointed military mustache were as black as the night from which Rainsford had come. His eyes, too, were black and very bright. He had high cheekbones, a sharpcut nose, a spare, dark face--the face of a man used to giving orders, the face of an aristocrat. Turning to the giant in uniform, the general made a sign. The giant put away his pistol, saluted, withdrew.\n\n"Ivan is an incredibly strong fellow," remarked the general, "but he has the misfortune to be deaf and dumb. A simple fellow, but, I'm afraid, like all his race, a bit of a savage."\n\n"Is he Russian?"\n\n"He is a Cossack," said the general, and his smile showed red lips and pointed teeth. "So am I."\n\n"Come," he said, "we shouldn't be chatting here. We can talk later. Now you want clothes, food, rest. You shall have them. This is a most-restful spot."\n\nIvan had reappeared, and the general spoke to him with lips that moved but gave forth no sound.\n\n"Follow Ivan, if you please, Mr. Rainsford," said the general. "I was about to have my dinner when you came. I'll wait for you. You'll find that my clothes will fit you, I think."\n\nIt was to a huge, beam-ceilinged bedroom with a canopied bed big enough for six men that Rainsford followed the silent giant. Ivan laid out an evening suit, and Rainsford, as he put it on, noticed that it came from a London tailor who ordinarily cut and sewed for none below the rank of duke.\n\nThe dining room to which Ivan conducted him was in many ways remarkable. There was a medieval magnificence about it; it suggested a baronial hall of feudal times with its oaken panels, its high ceiling, its vast refectory tables where twoscore men could sit down to eat. About the hall were mounted heads of many animals--lions, tigers, elephants, moose, bears; larger or more perfect specimens Rainsford had never seen. At the great table the general was sitting, alone.\n\n"You'll have a cocktail, Mr. Rainsford," he suggested. The cocktail was surpassingly good; and, Rainsford noted, the table apointments were of the finest--the linen, the crystal, the silver, the china.\n\nThey were eating borsch, the rich, red soup with whipped cream so dear to Russian palates. Half apologetically General Zaroff said, "We do our best to preserve the amenities of civilization here. Please forgive any lapses. We are well off the beaten track, you know. Do you think the champagne has suffered from its long ocean trip?"\n\n"Not in the least," declared Rainsford. He was finding the general a most thoughtful and affable host, a true cosmopolite. But there was one small trait of the general's that made Rainsford uncomfortable. Whenever he looked up from his plate he found the general studying him, appraising him narrowly.\n\n"Perhaps," said General Zaroff, "you were surprised that I recognized your name. You see, I read all books on hunting published in English, French, and Russian. I have but one passion in my life, Mr. Rainsford, and it is the hunt."\n\n"You have some wonderful heads here," said Rainsford as he ate a particularly well-cooked filet mignon. " That Cape buffalo is the largest I ever saw."\n\n"Oh, that fellow. Yes, he was a monster."\n\n"Did he charge you?"\n\n"Hurled me against a tree," said the general. "Fractured my skull. But I got the brute."\n\n"I've always thought," said Rainsford, "that the Cape buffalo is the most dangerous of all big game."\n\nFor a moment the general did not reply; he was smiling his curious red-lipped smile. Then he said slowly, "No. You are wrong, sir. The Cape buffalo is not the most dangerous big game." He sipped his wine. "Here in my preserve on this island," he said in the same slow tone, "I hunt more dangerous game."\n\nRainsford expressed his surprise. "Is there big game on this island?"\n\nThe general nodded. "The biggest."\n\n"Really?"\n\n"Oh, it isn't here naturally, of course. I have to stock the island."\n\n"What have you imported, general?" Rainsford asked. "Tigers?"\n\nThe general smiled. "No," he said. "Hunting tigers ceased to interest me some years ago. I exhausted their possibilities, you see. No thrill left in tigers, no real danger. I live for danger, Mr. Rainsford."\n\nThe general took from his pocket a gold cigarette case and offered his guest a long black cigarette with a silver tip; it was perfumed and gave off a smell like incense.\n\n"We will have some capital hunting, you and I," said the general. "I shall be most glad to have your society."\n\n"But what game--" began Rainsford.\n\n"I'll tell you," said the general. "You will be amused, I know. I think I may say, in all modesty, that I have done a rare thing. I have invented a new sensation. May I pour you another glass of port?"\n\n"Thank you, general."\n\nThe general filled both glasses, and said, "God makes some men poets. Some He makes kings, some beggars. Me He made a hunter. My hand was made for the trigger, my father said. He was a very rich man with a quarter of a million acres in the Crimea, and he was an ardent sportsman. When I was only five years old he gave me a little gun, specially made in Moscow for me, to shoot sparrows with. When I shot some of his prize turkeys with it, he did not punish me; he complimented me on my marksmanship. I killed my first bear in the Caucasus when I was ten. My whole life has been one prolonged hunt. I went into the army--it was expected of noblemen's sons--and for a time commanded a division of Cossack cavalry, but my real interest was always the hunt. I have hunted every kind of game in every land. It would be impossible for me to tell you how many animals I have killed."\n\nThe general puffed at his cigarette.\n\n"After the debacle in Russia I left the country, for it was imprudent for an officer of the Czar to stay there. Many noble Russians lost everything. I, luckily, had invested heavily in American securities, so I shall never have to open a tearoom in Monte Carlo or drive a taxi in Paris. Naturally, I continued to hunt--grizzlies in your Rockies, crocodiles in the Ganges, rhinoceroses in East Africa. It was in Africa that the Cape buffalo hit me and laid me up for six months. As soon as I recovered I started for the Amazon to hunt jaguars, for I had heard they were unusually cunning. They weren't." The Cossack sighed. "They were no match at all for a hunter with his wits about him, and a high-powered rifle. I was bitterly disappointed. I was lying in my tent with a splitting headache one night when a terrible thought pushed its way into my mind. Hunting was beginning to bore me! And hunting, remember, had been my life. I have heard that in America businessmen often go to pieces when they give up the business that has been their life."\n\n"Yes, that's so," said Rainsford.\n\nThe general smiled. "I had no wish to go to pieces," he said. "I must do something. Now, mine is an analytical mind, Mr. Rainsford. Doubtless that is why I enjoy the problems of the chase."\n\n"No doubt, General Zaroff."\n\n"So," continued the general, "I asked myself why the hunt no longer fascinated me. You are much younger than I am, Mr. Rainsford, and have not hunted as much, but you perhaps can guess the answer."\n\n"What was it?"\n\n"Simply this: hunting had ceased to be what you call 'a sporting proposition.' It had become too easy. I always got my quarry. Always. There is no greater bore than perfection."\n\nThe general lit a fresh cigarette.\n\n"No animal had a chance with me any more. That is no boast; it is a mathematical certainty. The animal had nothing but his legs and his instinct. Instinct is no match for reason. When I thought of this it was a tragic moment for me, I can tell you."\n\nRainsford leaned across the table, absorbed in what his host was saying.\n\n"It came to me as an inspiration what I must do," the general went on.\n\n"And that was?"\n\nThe general smiled the quiet smile of one who has faced an obstacle and surmounted it with success. "I had to invent a new animal to hunt," he said.\n\n"A new animal? You're joking."\n\n"Not at all," said the general. "I never joke about hunting. I needed a new animal. I found one. So I bought this island built this house, and here I do my hunting. The island is perfect for my purposes--there are jungles with a maze of traits in them, hills, swamps--"\n\n"But the animal, General Zaroff?"\n\n"Oh," said the general, "it supplies me with the most exciting hunting in the world. No other hunting compares with it for an instant. Every day I hunt, and I never grow bored now, for I have a quarry with which I can match my wits."\n\nRainsford's bewilderment showed in his face.\n\n"I wanted the ideal animal to hunt," explained the general. "So I said, 'What are the attributes of an ideal quarry?' And the answer was, of course, 'It must have courage, cunning, and, above all, it must be able to reason."'\n\n"But no animal can reason," objected Rainsford.\n\n"My dear fellow," said the general, "there is one that can."\n\n"But you can't mean--" gasped Rainsford.\n\n"And why not?"\n\n"I can't believe you are serious, General Zaroff. This is a grisly joke."\n\n"Why should I not be serious? I am speaking of hunting."\n\n"Hunting? Great Guns, General Zaroff, what you speak of is murder."\n\nThe general laughed with entire good nature. He regarded Rainsford quizzically. "I refuse to believe that so modern and civilized a young man as you seem to be harbors romantic ideas about the value of human life. Surely your experiences in the war--"\n\n"Did not make me condone cold-blooded murder," finished Rainsford stiffly.\n\nLaughter shook the general. "How extraordinarily droll you are!" he said. "One does not expect nowadays to find a young man of the educated class, even in America, with such a naïve, and, if I may say so, mid-Victorian point of view. It's like finding a snuffbox in a limousine. Ah, well, doubtless you had Puritan ancestors. So many Americans appear to have had. I'll wager you'll forget your notions when you go hunting with me. You've a genuine new thrill in store for you, Mr. Rainsford."\n\n"Thank you, I'm a hunter, not a murderer."\n\n"Dear me," said the general, quite unruffled, "again that unpleasant word. But I think I can show you that your scruples are quite ill founded."\n\n"Yes?"\n\n"Life is for the strong, to be lived by the strong, and, if needs be, taken by the strong. The weak of the world were put here to give the strong pleasure. I am strong. Why should I not use my gift? If I wish to hunt, why should I not? I hunt the scum of the earth: sailors from tramp ships--lassars, blacks, Chinese, whites, mongrels--a thoroughbred horse or hound is worth more than a score of them."\n\n"But they are men," said Rainsford hotly.\n\n"Precisely," said the general. "That is why I use them. It gives me pleasure. They can reason, after a fashion. So they are dangerous."\n\n"But where do you get them?"\n\nThe general's left eyelid fluttered down in a wink. "This island is called Ship Trap," he answered. "Sometimes an angry god of the high seas sends them to me. Sometimes, when Providence is not so kind, I help Providence a bit. Come to the window with me."\n\nRainsford went to the window and looked out toward the sea.\n\n"Watch! Out there!" exclaimed the general, pointing into the night. Rainsford's eyes saw only blackness, and then, as the general pressed a button, far out to sea Rainsford saw the flash of lights.\n\nThe general chuckled. "They indicate a channel," he said, "where there's none; giant rocks with razor edges crouch like a sea monster with wide-open jaws. They can crush a ship as easily as I crush this nut." He dropped a walnut on the hardwood floor and brought his heel grinding down on it. "Oh, yes," he said, casually, as if in answer to a question, "I have electricity. We try to be civilized here."\n\n"Civilized? And you shoot down men?"\n\nA trace of anger was in the general's black eyes, but it was there for but a second; and he said, in his most pleasant manner, "Dear me, what a righteous young man you are! I assure you I do not do the thing you suggest. That would be barbarous. I treat these visitors with every consideration. They get plenty of good food and exercise. They get into splendid physical condition. You shall see for yourself tomorrow."\n\n"What do you mean?"\n\n"We'll visit my training school," smiled the general. "It's in the cellar. I have about a dozen pupils down there now. They're from the Spanish bark San Lucar that had the bad luck to go on the rocks out there. A very inferior lot, I regret to say. Poor specimens and more accustomed to the deck than to the jungle." He raised his hand, and Ivan, who served as waiter, brought thick Turkish coffee. Rainsford, with an effort, held his tongue in check.\n\n"It's a game, you see," pursued the general blandly. "I suggest to one of them that we go hunting. I give him a supply of food and an excellent hunting knife. I give him three hours' start. I am to follow, armed only with a pistol of the smallest caliber and range. If my quarry eludes me for three whole days, he wins the game. If I find him "--the general smiled--" he loses."\n\n"Suppose he refuses to be hunted?"\n\n"Oh," said the general, "I give him his option, of course. He need not play that game if he doesn't wish to. If he does not wish to hunt, I turn him over to Ivan. Ivan once had the honor of serving as official knouter to the Great White Czar, and he has his own ideas of sport. Invariably, Mr. Rainsford, invariably they choose the hunt."\n\n"And if they win?"\n\nThe smile on the general's face widened. "To date I have not lost," he said. Then he added, hastily: "I don't wish you to think me a braggart, Mr. Rainsford. Many of them afford only the most elementary sort of problem. Occasionally I strike a tartar. One almost did win. I eventually had to use the dogs."\n\n"The dogs?"\n\n"This way, please. I'll show you."\n\nThe general steered Rainsford to a window. The lights from the windows sent a flickering illumination that made grotesque patterns on the courtyard below, and Rainsford could see moving about there a dozen or so huge black shapes; as they turned toward him, their eyes glittered greenly.\n\n"A rather good lot, I think," observed the general. "They are let out at seven every night. If anyone should try to get into my house--or out of it--something extremely regrettable would occur to him." He hummed a snatch of song from the Folies Bergere.\n\n"And now," said the general, "I want to show you my new collection of heads. Will you come with me to the library?"\n\n"I hope," said Rainsford, "that you will excuse me tonight, General Zaroff. I'm really not feeling well."\n\n"Ah, indeed?" the general inquired solicitously. "Well, I suppose that's only natural, after your long swim. You need a good, restful night's sleep. Tomorrow you'll feel like a new man, I'll wager. Then we'll hunt, eh? I've one rather promising prospect--" Rainsford was hurrying from the room.\n\n"Sorry you can't go with me tonight," called the general. "I expect rather fair sport--a big, strong, black. He looks resourceful--Well, good night, Mr. Rainsford; I hope you have a good night's rest."\n\nThe bed was good, and the pajamas of the softest silk, and he was tired in every fiber of his being, but nevertheless Rainsford could not quiet his brain with the opiate of sleep. He lay, eyes wide open. Once he thought he heard stealthy steps in the corridor outside his room. He sought to throw open the door; it would not open. He went to the window and looked out. His room was high up in one of the towers. The lights of the chateau were out now, and it was dark and silent; but there was a fragment of sallow moon, and by its wan light he could see, dimly, the courtyard. There, weaving in and out in the pattern of shadow, were black, noiseless forms; the hounds heard him at the window and looked up, expectantly, with their green eyes. Rainsford went back to the bed and lay down. By many methods he tried to put himself to sleep. He had achieved a doze when, just as morning began to come, he heard, far off in the jungle, the faint report of a pistol.\n\nGeneral Zaroff did not appear until luncheon. He was dressed faultlessly in the tweeds of a country squire. He was solicitous about the state of Rainsford's health.\n\n"As for me," sighed the general, "I do not feel so well. I am worried, Mr. Rainsford. Last night I detected traces of my old complaint."\n\nTo Rainsford's questioning glance the general said, "Ennui. Boredom."\n\nThen, taking a second helping of crâpes Suzette, the general explained: "The hunting was not good last night. The fellow lost his head. He made a straight trail that offered no problems at all. That's the trouble with these sailors; they have dull brains to begin with, and they do not know how to get about in the woods. They do excessively stupid and obvious things. It's most annoying. Will you have another glass of Chablis, Mr. Rainsford?"\n\n"General," said Rainsford firmly, "I wish to leave this island at once."\n\nThe general raised his thickets of eyebrows; he seemed hurt. "But, my dear fellow," the general protested, "you've only just come. You've had no hunting--"\n\n"I wish to go today," said Rainsford. He saw the dead black eyes of the general on him, studying him. General Zaroff's face suddenly brightened.\n\nHe filled Rainsford's glass with venerable Chablis from a dusty bottle.\n\n"Tonight," said the general, "we will hunt--you and I."\n\nRainsford shook his head. "No, general," he said. "I will not hunt."\n\nThe general shrugged his shoulders and delicately ate a hothouse grape. "As you wish, my friend," he said. "The choice rests entirely with you. But may I not venture to suggest that you will find my idea of sport more diverting than Ivan's?"\n\nHe nodded toward the corner to where the giant stood, scowling, his thick arms crossed on his hogshead of chest.\n\n"You don't mean--" cried Rainsford.\n\n"My dear fellow," said the general, "have I not told you I always mean what I say about hunting? This is really an inspiration. I drink to a foeman worthy of my steel--at last." The general raised his glass, but Rainsford sat staring at him.\n\n"You'll find this game worth playing," the general said enthusiastically." Your brain against mine. Your woodcraft against mine. Your strength and stamina against mine. Outdoor chess! And the stake is not without value, eh?"\n\n"And if I win--" began Rainsford huskily.\n\n"I'll cheerfully acknowledge myself defeat if I do not find you by midnight of the third day," said General Zaroff. "My sloop will place you on the mainland near a town." The general read what Rainsford was thinking.\n\n"Oh, you can trust me," said the Cossack. "I will give you my word as a gentleman and a sportsman. Of course you, in turn, must agree to say nothing of your visit here."\n\n"I'll agree to nothing of the kind," said Rainsford.\n\n"Oh," said the general, "in that case--But why discuss that now? Three days hence we can discuss it over a bottle of Veuve Cliquot, unless--"\n\nThe general sipped his wine.\n\nThen a businesslike air animated him. "Ivan," he said to Rainsford, "will supply you with hunting clothes, food, a knife. I suggest you wear moccasins; they leave a poorer trail. I suggest, too, that you avoid the big swamp in the southeast corner of the island. We call it Death Swamp. There's quicksand there. One foolish fellow tried it. The deplorable part of it was that Lazarus followed him. You can imagine my feelings, Mr. Rainsford. I loved Lazarus; he was the finest hound in my pack. Well, I must beg you to excuse me now. I always' take a siesta after lunch. You'll hardly have time for a nap, I fear. You'll want to start, no doubt. I shall not follow till dusk. Hunting at night is so much more exciting than by day, don't you think? Au revoir, Mr. Rainsford, au revoir." General Zaroff, with a deep, courtly bow, strolled from the room.\n\nFrom another door came Ivan. Under one arm he carried khaki hunting clothes, a haversack of food, a leather sheath containing a long-bladed hunting knife; his right hand rested on a cocked revolver thrust in the crimson sash about his waist.\n\nRainsford had fought his way through the bush for two hours. "I must keep my nerve. I must keep my nerve," he said through tight teeth.\n\nHe had not been entirely clearheaded when the chateau gates snapped shut behind him. His whole idea at first was to put distance between himself and General Zaroff; and, to this end, he had plunged along, spurred on by the sharp rowers of something very like panic. Now he had got a grip on himself, had stopped, and was taking stock of himself and the situation. He saw that straight flight was futile; inevitably it would bring him face to face with the sea. He was in a picture with a frame of water, and his operations, clearly, must take place within that frame.\n\n"I'll give him a trail to follow," muttered Rainsford, and he struck off from the rude path he had been following into the trackless wilderness. He executed a series of intricate loops; he doubled on his trail again and again, recalling all the lore of the fox hunt, and all the dodges of the fox. Night found him leg-weary, with hands and face lashed by the branches, on a thickly wooded ridge. He knew it would be insane to blunder on through the dark, even if he had the strength. His need for rest was imperative and he thought, "I have played the fox, now I must play the cat of the fable." A big tree with a thick trunk and outspread branches was near by, and, taking care to leave not the slightest mark, he climbed up into the crotch, and, stretching out on one of the broad limbs, after a fashion, rested. Rest brought him new confidence and almost a feeling of security. Even so zealous a hunter as General Zaroff could not trace him there, he told himself; only the devil himself could follow that complicated trail through the jungle after dark. But perhaps the general was a devil--\n\nAn apprehensive night crawled slowly by like a wounded snake and sleep did not visit Rainsford, although the silence of a dead world was on the jungle. Toward morning when a dingy gray was varnishing the sky, the cry of some startled bird focused Rainsford's attention in that direction. Something was coming through the bush, coming slowly, carefully, coming by the same winding way Rainsford had come. He flattened himself down on the limb and, through a screen of leaves almost as thick as tapestry, he watched. . . . That which was approaching was a man.\n\nIt was General Zaroff. He made his way along with his eyes fixed in utmost concentration on the ground before him. He paused, almost beneath the tree, dropped to his knees and studied the ground. Rainsford's impulse was to hurl himself down like a panther, but he saw that the general's right hand held something metallic--a small automatic pistol.\n\nThe hunter shook his head several times, as if he were puzzled. Then he straightened up and took from his case one of his black cigarettes; its pungent incenselike smoke floated up to Rainsford's nostrils.\n\nRainsford held his breath. The general's eyes had left the ground and were traveling inch by inch up the tree. Rainsford froze there, every muscle tensed for a spring. But the sharp eyes of the hunter stopped before they reached the limb where Rainsford lay; a smile spread over his brown face. Very deliberately he blew a smoke ring into the air; then he turned his back on the tree and walked carelessly away, back along the trail he had come. The swish of the underbrush against his hunting boots grew fainter and fainter.\n\nThe pent-up air burst hotly from Rainsford's lungs. His first thought made him feel sick and numb. The general could follow a trail through the woods at night; he could follow an extremely difficult trail; he must have uncanny powers; only by the merest chance had the Cossack failed to see his quarry.\n\nRainsford's second thought was even more terrible. It sent a shudder of cold horror through his whole being. Why had the general smiled? Why had he turned back?\n\nRainsford did not want to believe what his reason told him was true, but the truth was as evident as the sun that had by now pushed through the morning mists. The general was playing with him! The general was saving him for another day's sport! The Cossack was the cat; he was the mouse. Then it was that Rainsford knew the full meaning of terror.\n\n"I will not lose my nerve. I will not."\n\nHe slid down from the tree, and struck off again into the woods. His face was set and he forced the machinery of his mind to function. Three hundred yards from his hiding place he stopped where a huge dead tree leaned precariously on a smaller, living one. Throwing off his sack of food, Rainsford took his knife from its sheath and began to work with all his energy.\n\nThe job was finished at last, and he threw himself down behind a fallen log a hundred feet away. He did not have to wait long. The cat was coming again to play with the mouse.\n\nFollowing the trail with the sureness of a bloodhound came General Zaroff. Nothing escaped those searching black eyes, no crushed blade of grass, no bent twig, no mark, no matter how faint, in the moss. So intent was the Cossack on his stalking that he was upon the thing Rainsford had made before he saw it. His foot touched the protruding bough that was the trigger. Even as he touched it, the general sensed his danger and leaped back with the agility of an ape. But he was not quite quick enough; the dead tree, delicately adjusted to rest on the cut living one, crashed down and struck the general a glancing blow on the shoulder as it fell; but for his alertness, he must have been smashed beneath it. He staggered, but he did not fall; nor did he drop his revolver. He stood there, rubbing his injured shoulder, and Rainsford, with fear again gripping his heart, heard the general's mocking laugh ring through the jungle.\n\n"Rainsford," called the general, "if you are within sound of my voice, as I suppose you are, let me congratulate you. Not many men know how to make a Malay mancatcher. Luckily for me I, too, have hunted in Malacca. You are proving interesting, Mr. Rainsford. I am going now to have my wound dressed; it's only a slight one. But I shall be back. I shall be back."\n\nWhen the general, nursing his bruised shoulder, had gone, Rainsford took up his flight again. It was flight now, a desperate, hopeless flight, that carried him on for some hours. Dusk came, then darkness, and still he pressed on. The ground grew softer under his moccasins; the vegetation grew ranker, denser; insects bit him savagely.\n\nThen, as he stepped forward, his foot sank into the ooze. He tried to wrench it back, but the muck sucked viciously at his foot as if it were a giant leech. With a violent effort, he tore his feet loose. He knew where he was now. Death Swamp and its quicksand.\n\nHis hands were tight closed as if his nerve were something tangible that someone in the darkness was trying to tear from his grip. The softness of the earth had given him an idea. He stepped back from the quicksand a dozen feet or so and, like some huge prehistoric beaver, he began to dig.\n\nRainsford had dug himself in in France when a second's delay meant death. That had been a placid pastime compared to his digging now. The pit grew deeper; when it was above his shoulders, he climbed out and from some hard saplings cut stakes and sharpened them to a fine point. These stakes he planted in the bottom of the pit with the points sticking up. With flying fingers he wove a rough carpet of weeds and branches and with it he covered the mouth of the pit. Then, wet with sweat and aching with tiredness, he crouched behind the stump of a lightning-charred tree.\n\nHe knew his pursuer was coming; he heard the padding sound of feet on the soft earth, and the night breeze brought him the perfume of the general's cigarette. It seemed to Rainsford that the general was coming with unusual swiftness; he was not feeling his way along, foot by foot. Rainsford, crouching there, could not see the general, nor could he see the pit. He lived a year in a minute. Then he felt an impulse to cry aloud with joy, for he heard the sharp crackle of the breaking branches as the cover of the pit gave way; he heard the sharp scream of pain as the pointed stakes found their mark. He leaped up from his place of concealment. Then he cowered back. Three feet from the pit a man was standing, with an electric torch in his hand.\n\n"You've done well, Rainsford," the voice of the general called. "Your Burmese tiger pit has claimed one of my best dogs. Again you score. I think, Mr. Rainsford, I'll see what you can do against my whole pack. I'm going home for a rest now. Thank you for a most amusing evening."\n\nAt daybreak Rainsford, lying near the swamp, was awakened by a sound that made him know that he had new things to learn about fear. It was a distant sound, faint and wavering, but he knew it. It was the baying of a pack of hounds.\n\nRainsford knew he could do one of two things. He could stay where he was and wait. That was suicide. He could flee. That was postponing the inevitable. For a moment he stood there, thinking. An idea that held a wild chance came to him, and, tightening his belt, he headed away from the swamp.\n\nThe baying of the hounds drew nearer, then still nearer, nearer, ever nearer. On a ridge Rainsford climbed a tree. Down a watercourse, not a quarter of a mile away, he could see the bush moving. Straining his eyes, he saw the lean figure of General Zaroff; just ahead of him Rainsford made out another figure whose wide shoulders surged through the tall jungle weeds; it was the giant Ivan, and he seemed pulled forward by some unseen force; Rainsford knew that Ivan must be holding the pack in leash.\n\nThey would be on him any minute now. His mind worked frantically. He thought of a native trick he had learned in Uganda. He slid down the tree. He caught hold of a springy young sapling and to it he fastened his hunting knife, with the blade pointing down the trail; with a bit of wild grapevine he tied back the sapling. Then he ran for his life. The hounds raised their voices as they hit the fresh scent. Rainsford knew now how an animal at bay feels.\n\nHe had to stop to get his breath. The baying of the hounds stopped abruptly, and Rainsford's heart stopped too. They must have reached the knife.\n\nHe shinned excitedly up a tree and looked back. His pursuers had stopped. But the hope that was in Rainsford's brain when he climbed died, for he saw in the shallow valley that General Zaroff was still on his feet. But Ivan was not. The knife, driven by the recoil of the springing tree, had not wholly failed.\n\nRainsford had hardly tumbled to the ground when the pack took up the cry again.\n\n"Nerve, nerve, nerve!" he panted, as he dashed along. A blue gap showed between the trees dead ahead. Ever nearer drew the hounds. Rainsford forced himself on toward that gap. He reached it. It was the shore of the sea. Across a cove he could see the gloomy gray stone of the chateau. Twenty feet below him the sea rumbled and hissed. Rainsford hesitated. He heard the hounds. Then he leaped far out into the sea. . . .\n\nWhen the general and his pack reached the place by the sea, the Cossack stopped. For some minutes he stood regarding the blue-green expanse of water. He shrugged his shoulders. Then he sat down, took a drink of brandy from a silver flask, lit a cigarette, and hummed a bit from Madame Butterfly.\n\nGeneral Zaroff had an exceedingly good dinner in his great paneled dining hall that evening. With it he had a bottle of Pol Roger and half a bottle of Chambertin. Two slight annoyances kept him from perfect enjoyment. One was the thought that it would be difficult to replace Ivan; the other was that his quarry had escaped him; of course, the American hadn't played the game--so thought the general as he tasted his after-dinner liqueur. In his library he read, to soothe himself, from the works of Marcus Aurelius. At ten he went up to his bedroom. He was deliciously tired, he said to himself, as he locked himself in. There was a little moonlight, so, before turning on his light, he went to the window and looked down at the courtyard. He could see the great hounds, and he called, "Better luck another time," to them. Then he switched on the light.\n\nA man, who had been hiding in the curtains of the bed, was standing there.\n\n"Rainsford!" screamed the general. "How in God's name did you get here?"\n\n"Swam," said Rainsford. "I found it quicker than walking through the jungle."\n\nThe general sucked in his breath and smiled. "I congratulate you," he said. "You have won the game."\n\nRainsford did not smile. "I am still a beast at bay," he said, in a low, hoarse voice. "Get ready, General Zaroff."\n\nThe general made one of his deepest bows. "I see," he said. "Splendid! One of us is to furnish a repast for the hounds. The other will sleep in this very excellent bed. On guard, Rainsford."\n\n. . .\n\nHe had never slept in a better bed, Rainsford decided.\n\nBack to Top\n |