1. Lecture 1, 7/21/2005Natural Language Processing1 CS60057 Speech &Natural Language Processing Autumn 2007 Lecture 14b 24 August 2007

2. Lecture 1, 7/21/2005Natural Language Processing2 LING 180 SYMBSYS 138 Intro to Computer Speech and Language Processing Lecture 9: Machine Translation (I) November 7, 2006 Dan Jurafsky Thanks to Bonnie Dorr for some of these slides!!

3. Lecture 1, 7/21/2005Natural Language Processing3 Outline for MT Week  Intro and a little history  Language Similarities and Divergences  Three classic MT Approaches Transfer Interlingua Direct  Modern Statistical MT  Evaluation

4. Lecture 1, 7/21/2005Natural Language Processing4 What is MT?  Translating a text from one language to another automatically.

5. Lecture 1, 7/21/2005Natural Language Processing5 Machine Translation  dai yu zi zai chuang shang gan nian bao chai you ting jian chuang wai zhu shao xiang ye zhe shang, yu sheng xi li, qing han tou mu, bu jue you di xia lei lai.  Dai-yu alone on bed top think-of-with-gratitude Bao-chai again listen to window outside bamboo tip plantain leaf of on-top rain sound sigh drop clear cold penetrate curtain not feeling again fall down tears come  As she lay there alone, Dai-yu’s thoughts turned to Bao-chai… Then she listened to the insistent rustle of the rain on the bamboos and plantains outside her window. The coldness penetrated the curtains of her bed. Almost without noticing it she had begun to cry.

6. Lecture 1, 7/21/2005Natural Language Processing6 Machine Translation

7. Lecture 1, 7/21/2005Natural Language Processing7 Machine Translation  The Story of the Stone  =The Dream of the Red Chamber (Cao Xueqin 1792)  Issues: Word segmentation Sentence segmentation: 4 English sentences to 1 Chinese Grammatical differences  Chinese rarely marks tense:  As, turned to, had begun,  tou -> penetrated  Zero anaphora  No articles Stylistic and cultural differences  Bamboo tip plaintain leaf -> bamboos and plantains  Ma ‘curtain’ -> curtains of her bed  Rain sound sigh drop -> insistent rustle of the rain

8. Lecture 1, 7/21/2005Natural Language Processing8 Not just literature  Hansards: Canadian parliamentary proceeedings

9. Lecture 1, 7/21/2005Natural Language Processing9 What is MT not good for?  Really hard stuff Literature Natural spoken speech (meetings, court reporting)  Really important stuff Medical translation in hospitals, 911

10. Lecture 1, 7/21/2005Natural Language Processing10 What is MT good for?  Tasks for which a rough translation is fine Web pages, email  Tasks for which MT can be post-edited MT as first pass “Computer-aided human translation  Tasks in sublanguage domains where high-quality MT is possible FAHQT

11. Lecture 1, 7/21/2005Natural Language Processing11 Sublanguage domain  Weather forecasting “Cloudy with a chance of showers today and Thursday” “Low tonight 4”  Can be modeling completely enough to use raw MT output  Word classes and semantic features like MONTH, PLACE, DIRECTION, TIME POINT

12. Lecture 1, 7/21/2005Natural Language Processing12 MT History  1946 Booth and Weaver discuss MT at Rockefeller foundation in New York;  1947-48 idea of dictionary-based direct translation  1949 Weaver memorandum popularized idea  1952 all 18 MT researchers in world meet at MIT  1954 IBM/Georgetown Demo Russian-English MT  1955-65 lots of labs take up MT

13. Lecture 1, 7/21/2005Natural Language Processing13 History of MT: Pessimism  1959/1960: Bar-Hillel “Report on the state of MT in US and GB” Argued FAHQT too hard (semantic ambiguity, etc) Should work on semi-automatic instead of automatic His argument Little John was looking for his toy box. Finally, he found it. The box was in the pen. John was very happy. Only human knowledge let’s us know that ‘playpens’ are bigger than boxes, but ‘writing pens’ are smaller His claim: we would have to encode all of human knowledge

14. Lecture 1, 7/21/2005Natural Language Processing14 History of MT: Pessimism  The ALPAC report Headed by John R. Pierce of Bell Labs Conclusions:  Supply of human translators exceeds demand  All the Soviet literature is already being translated  MT has been a failure: all current MT work had to be post-edited  Sponsored evaluations which showed that intelligibility and informativeness was worse than human translations Results:  MT research suffered  Funding loss  Number of research labs declined  Association for Machine Translation and Computational Linguistics dropped MT from its name

15. Lecture 1, 7/21/2005Natural Language Processing15 History of MT  1976 Meteo, weather forecasts from English to French  Systran (Babelfish) been used for 40 years  1970’s: European focus in MT; mainly ignored in US  1980’s ideas of using AI techniques in MT (KBMT, CMU)  1990’s Commercial MT systems Statistical MT Speech-to-speech translation

16. Lecture 1, 7/21/2005Natural Language Processing16 Language Similarities and Divergences  Some aspects of human language are universal or near- universal, others diverge greatly.  Typology: the study of systematic cross-linguistic similarities and differences  What are the dimensions along with human languages vary?

17. Lecture 1, 7/21/2005Natural Language Processing17 Morphological Variation  Isolating languages Cantonese, Vietnamese: each word generally has one morpheme  Vs. Polysynthetic languages Siberian Yupik (`Eskimo’): single word may have very many morphemes  Agglutinative languages Turkish: morphemes have clean boundaries  Vs. Fusion languages Russian: single affix may have many morphemes

18. Lecture 1, 7/21/2005Natural Language Processing18 Syntactic Variation  SVO (Subject-Verb-Object) languages English, German, French, Mandarin  SOV Languages Japanese, Hindi  VSO languages Irish, Classical Arabic  SVO lgs generally prepositions: to Yuriko  VSO lgs generally postpositions: Yuriko ni

19. Lecture 1, 7/21/2005Natural Language Processing19 Segmentation Variation  Not every writing system has word boundaries marked Chinese, Japanese, Thai, Vietnamese  Some languages tend to have sentences that are quite long, closer to English paragraphs than sentences: Modern Standard Arabic, Chinese

20. Lecture 1, 7/21/2005Natural Language Processing20 Inferential Load: cold vs. hot lgs  Some ‘cold’ languages require the hearer to do more “figuring out” of who the various actors in the various events are: Japanese, Chinese,  Other ‘hot’ languages are pretty explicit about saying who did what to whom. English

21. Lecture 1, 7/21/2005Natural Language Processing21 Inferential Load (2) All noun phrases in blue do not appear in Chinese text … But they are needed for a good translation

22. Lecture 1, 7/21/2005Natural Language Processing22 Lexical Divergences  Word to phrases: English “computer science” = French “informatique”  POS divergences Eng. ‘she likes/VERB to sing’ Ger. Sie singt gerne/ADV Eng ‘I’m hungry/ADJ Sp. ‘tengo hambre/NOUN

23. Lecture 1, 7/21/2005Natural Language Processing23 Lexical Divergences: Specificity  Grammatical constraints English has gender on pronouns, Mandarin not.  So translating “3rd person” from Chinese to English, need to figure out gender of the person!  Similarly from English “they” to French “ils/elles”  Semantic constraints English `brother’ Mandarin ‘gege’ (older) versus ‘didi’ (younger) English ‘wall’ German ‘Wand’ (inside) ‘Mauer’ (outside) German ‘Berg’ English ‘hill’ or ‘mountain’

24. Lecture 1, 7/21/2005Natural Language Processing24 Lexical Divergence: many-to-many

25. Lecture 1, 7/21/2005Natural Language Processing25 Lexical Divergence: lexical gaps  Japanese: no word for privacy  English: no word for Cantonese ‘haauseun’ or Japanese ‘oyakoko’ (something like `filial piety’)  English ‘cow’ versus ‘beef’, Cantonese ‘ngau’

26. Lecture 1, 7/21/2005Natural Language Processing26 Event-to-argument divergences  English The bottle floated out.  Spanish La botella salió flotando. The bottle exited floating  Verb-framed lg: mark direction of motion on verb Spanish, French, Arabic, Hebrew, Japanese, Tamil, Polynesian, Mayan, Bantu familiies  Satellite-framed lg: mark direction of motion on satellite Crawl out, float off, jump down, walk over to, run after Rest of Indo-European, Hungarian, Finnish, Chinese

27. Lecture 1, 7/21/2005Natural Language Processing27 Structural divergences  G: Wir treffen uns am Mittwoch  E: We’ll meet on Wednesday

28. Lecture 1, 7/21/2005Natural Language Processing28 Head Swapping  E: X swim across Y  S: X crucar Y nadando  E: I like to eat  G: Ich esse gern  E: I’d prefer vanilla  G: Mir wäre Vanille lieber

29. Lecture 1, 7/21/2005Natural Language Processing29 Thematic divergence  Y me gusto  I like Y  G: Mir fällt der Termin ein  E: I forget the date

30. Lecture 1, 7/21/2005Natural Language Processing30 Divergence counts from Bonnie Dorr  32% of sentences in UN Spanish/English Corpus (5K) Categorial X tener hambre Y have hunger 98% Conflational X dar pu ñ aladas a Z X stab Z 83% Structural X entrar en Y X enter Y 35% Head Swapping X cruzar Y nadando X swim across Y 8% Thematic X gustar a Y Y likes X 6%

31. Lecture 1, 7/21/2005Natural Language Processing31 MT on the web  Babelfish: http://babelfish.altavista.com/  Google: http://www.google.com/search?hl=en&lr=&client=safa ri&rls=en&q="1+taza+de+jugo"+%28zumo%29+de+n aranja+5+cucharadas+de+azucar+morena&btnG=Se arch http://www.google.com/search?hl=en&lr=&client=safa ri&rls=en&q="1+taza+de+jugo"+%28zumo%29+de+n aranja+5+cucharadas+de+azucar+morena&btnG=Se arch

32. Lecture 1, 7/21/2005Natural Language Processing32 3 methods for MT  Direct  Transfer  Interlingua

33. Lecture 1, 7/21/2005Natural Language Processing33 Three MT Approaches: Direct, Transfer, Interlingual

34. Lecture 1, 7/21/2005Natural Language Processing34 Direct Translation  Proceed word-by-word through text  Translating each word  No intermediate structures except morphology  Knowledge is in the form of Huge bilingual dictionary word-to-word translation information  After word translation, can do simple reordering Adjective ordering English -> French/Spanish

35. Lecture 1, 7/21/2005Natural Language Processing35 Direct MT Dictionary entry

36. Lecture 1, 7/21/2005Natural Language Processing36 Direct MT

37. Lecture 1, 7/21/2005Natural Language Processing37 Problems with direct MT  German  Chinese

38. Lecture 1, 7/21/2005Natural Language Processing38 The Transfer Model  Idea: apply contrastive knowledge, i.e., knowledge about the difference between two languages  Steps: Analysis: Syntactically parse Source language Transfer: Rules to turn this parse into parse for Target language Generation: Generate Target sentence from parse tree

39. Lecture 1, 7/21/2005Natural Language Processing39 English to French  Generally English: Adjective Noun French: Noun Adjective Note: not always true  Route mauvaise ‘bad road, badly-paved road’  Mauvaise route ‘wrong road’)  But is a reasonable first approximation Rule:

40. Lecture 1, 7/21/2005Natural Language Processing40 Transfer rules

41. Lecture 1, 7/21/2005Natural Language Processing41 Lexical transfer  Transfer-based systems also need lexical transfer rules  Bilingual dictionary (like for direct MT)  English home:  German nach Hause (going home) Heim (home game) Heimat (homeland, home country) zu Hause (at home)  Can list “at home zu Hause”  Or do Word Sense Disambiguation

42. Lecture 1, 7/21/2005Natural Language Processing42 Systran: combining direct and transfer  Analysis Morphological analysis, POS tagging Chunking of NPs, PPs, phrases Shallow dependency parsing  Transfer Translation of idioms Word sense disambiguation Assigning prepositions based on governing verbs  Synthesis Apply rich bilingual dictionary Deal with reordering Morphological generation

43. Lecture 1, 7/21/2005Natural Language Processing43 Transfer: some problems  N 2 sets of transfer rules!  Grammar and lexicon full of language-specific stuff  Hard to build, hard to maintain

44. Lecture 1, 7/21/2005Natural Language Processing44 Interlingua  Intuition: Instead of lg-lg knowledge rules, use the meaning of the sentence to help  Steps: 1) translate source sentence into meaning representation 2) generate target sentence from meaning.

45. Lecture 1, 7/21/2005Natural Language Processing45 Interlingua for Mary did not slap the green witch

46. Lecture 1, 7/21/2005Natural Language Processing46 Interlingua  Idea is that some of the MT work that we need to do is part of other NLP tasks  E.g., disambiguating E:book S:‘libro’ from E:book S:‘reservar’  So we could have concepts like BOOKVOLUME and RESERVE and solve this problem once for each language

47. Lecture 1, 7/21/2005Natural Language Processing47 Direct MT: pros and cons (Bonnie Dorr)  Pros Fast Simple Cheap No translation rules hidden in lexicon  Cons Unreliable Not powerful Rule proliferation Requires lots of context Major restructuring after lexical substitution

48. Lecture 1, 7/21/2005Natural Language Processing48 Interlingual MT: pros and cons (B. Dorr)  Pros Avoids the N 2 problem Easier to write rules  Cons: Semantics is HARD Useful information lost (paraphrase)

49. Lecture 1, 7/21/2005Natural Language Processing49 The impossibility of translation  Hebrew “adonoi roi” for a culture without sheep or shepherds Something fluent and understandable, but not faithful:  “The Lord will look after me” Something faithful, but not fluent and nautral  “The Lord is for me like somebody who looks after animals with cotton-like hair”

50. Lecture 1, 7/21/2005Natural Language Processing50 What makes a good translation  Translators often talk about two factors we want to maximize:  Faithfulness or fidelity How close is the meaning of the translation to the meaning of the original (Even better: does the translation cause the reader to draw the same inferences as the original would have)  Fluency or naturalness How natural the translation is, just considering its fluency in the target language

51. Lecture 1, 7/21/2005Natural Language Processing51 Statistical MT: Faithfulness and Fluency formalized!  Best-translation of a source sentence S:  Developed by researchers who were originally in speech recognition at IBM  Called the IBM model

52. Lecture 1, 7/21/2005Natural Language Processing52 The IBM model  Hmm, those two factors might look familiar…  Yup, it’s Bayes rule:

53. Lecture 1, 7/21/2005Natural Language Processing53 More formally  Assume we are translating from a foreign language sentence F to an English sentence E: F = f 1, f 2, f 3,…, f m  We want to find the best English sentence E-hat = e 1, e 2, e 3,…, e n E-hat = argmax E P(E|F) = argmax E P(F|E)P(E)/P(F) = argmax E P(F|E)P(E) Translation ModelLanguage Model

54. Lecture 1, 7/21/2005Natural Language Processing54 The noisy channel model for MT

55. Lecture 1, 7/21/2005Natural Language Processing55 Fluency: P(T)  How to measure that this sentence That car was almost crash onto me  is less fluent than this one: That car almost hit me.  Answer: language models (N-grams!) For example P(hit|almost) > P(was|almost)  But can use any other more sophisticated model of grammar  Advantage: this is monolingual knowledge!

56. Lecture 1, 7/21/2005Natural Language Processing56 Faithfulness: P(S|T)  French: ça me plait [that me pleases]  English: that pleases me - most fluent I like it I’ll take that one  How to quantify this?  Intuition: degree to which words in one sentence are plausible translations of words in other sentence Product of probabilities that each word in target sentence would generate each word in source sentence.

57. Lecture 1, 7/21/2005Natural Language Processing57 Faithfulness P(S|T)  Need to know, for every target language word, probability of it mapping to every source language word.  How do we learn these probabilities?  Parallel texts! Lots of times we have two texts that are translations of each other If we knew which word in Source Text mapped to each word in Target Text, we could just count!

58. Lecture 1, 7/21/2005Natural Language Processing58 Faithfulness P(S|T)  Sentence alignment: Figuring out which source language sentence maps to which target language sentence  Word alignment Figuring out which source language word maps to which target language word

59. Lecture 1, 7/21/2005Natural Language Processing59 Big Point about Faithfulness and Fluency  Job of the faithfulness model P(S|T) is just to model “bag of words”; which words come from say English to Spanish.  P(S|T) doesn’t have to worry about internal facts about Spanish word order: that’s the job of P(T)  P(T) can do Bag generation: put the following words in order (from Kevin Knight) have programming a seen never I language better -actual the hashing is since not collision-free usually the is less perfectly the of somewhat capacity table

60. Lecture 1, 7/21/2005Natural Language Processing60 P(T) and bag generation: the answer  “Usually the actual capacity of the table is somewhat less, since the hashing is not collision-free”  How about: loves Mary John

61. Lecture 1, 7/21/2005Natural Language Processing61 Summary  Intro and a little history  Language Similarities and Divergences  Three classic MT Approaches Transfer Interlingua Direct  Modern Statistical MT  Evaluation