Shane Bergsma Johns Hopkins University

1. Better Together: Large Monolingual, Bilingual and Multimodal Corpora in Natural Language Processing Shane Bergsma Johns Hopkins University Fall, 2011

2. Research Vision Robust processing of human language requires knowledge beyond what’s in small manually-annotated data sets Many NLP successes exploit web-scale raw data: • Google Translate • IBM’s Watson • Things people use every day: • spelling correction, speech recognition, etc.

3. More data is better data [Banko & Brill, 2001] Grammar Correction Task @Microsoft

4. This Talk Derive lots of knowledge from web-scale data and apply to syntax, semantics, discourse: • Raw text on the web (Google N-grams) • Part 1: Non-referential pronouns • Bilingual text (words plus their translations) • Part 2: Parsing noun phrases • Visual data (labelled online images) • Part 3: Learning the meaning of words

5. Search Engines for NLP • Early web work: Use an Internet search engine to get data [Keller & Lapata, 2003]

6. Search Engines • Search Engines for NLP: some objections • Scientific: not reproducible, unreliable [Kilgarriff, 2007, “Googleology is bad science.”] • Practical: Too slow for millions of queries

7. N-grams • Google N-gram Data [Brants & Franz, 2006] • N words in sequence + their count on web • A compressed version of all the text on web • 24 GB zipped fits on your hard drive • Enables better features for a range of tasks [Bergsma et al. ACL 2008, IJCAI 2009, ACL 2010, etc.]

8. Part 1: Non-Referential Pronouns E.g. the word “it” in English • “You can make itin advance.” • referential (50-75%) • “You can makeitin Hollywood.” • non-referential (25-50%)

9. Non-Referential Pronouns • [Hirst, 1981]: detect non-referential pronouns, “lest precious hours be lost in bootless searches for textual referents.” • Most existing pronoun/coreference systems just ignore the problem • A common ambiguity: • “it” comprises 1% of English tokens

10. Non-Ref Detection as Classification • Input: s = “You can make itin advance” • Output: Is ita non-referential pronoun in s? Method: train a supervised classifier to make this decision on the basis of some features [Evans, 2001, Boyd et al. 2005, Müller 2006]

11. A Machine Learning Approach h(x) = w∙x(predict non-ref if h(x) > 0) • Typical ‘lexical’ features: binary indicators of context: x= (previous-word=make, next-word=in, previous-two-words=can+make, …) • Use training data to learn good values for the weights, w • Classifier learns, e.g., to give negative weight to PPs immediately preceding ‘it’ (e.g. … from it)

12. Better: Features from the Web [Bergsma, Lin, Goebel, ACL 2008] • Convert sentence to a context pattern: “make ____in advance” • Collect counts from the web: • “make it/themin advance” • 442vs. 449occurrences in Google N-gram Data • “makeit/themin Hollywood” • 3421 vs.0occurrences in Google N-gram Data

13. Applying the Web Counts • How wide should the patterns span? • We can use all that Google N-gram Data allows: You can make _ can make _ in make _ in advance _ in advance . • Five 5-grams, four 4-grams, three 3-grams and two bigrams • What fillers to use? (e.g. it, they/them, any NP?)

14. Web Count Features “it”: log-cnt(“You can make it in”) 5-grams log-cnt(“can make itin advance”) log-cnt(“make it in advance .”) ... log-cnt(“You can make it”) 4-grams log-cnt(“can make it in”) ... ... “them”: log-cnt(“You can make them in”) 5-grams ... ...

15. A Machine Learning Approach Revisited h(x) = w∙x(predict non-ref if h(x) > 0) • Typical features: binary indicators of context: x= (previous-word=make, next-word=in, previous-two-words=can+make, …) • New features: real-valued counts in web text: x= (log-cnt(“make it in advance”), log-cnt(“make them in advance”, log-cnt(“make * in advance”), …) • Key conclusion: classifiers with web features are robust on new domains! [Bergsma, Pitler, Lin, ACL 2010]

16. NADA [Bergsma & Yarowsky, DAARC 2011] • Non-Anaphoric Detection Algorithm: • a system for identifying non-referential pronouns • Works on raw sentences; no parsing/tagging of input needed • Classifies ‘it’ in up to 20,000 sentences/second • It works well when used out-of-domain • Because it’s got those Web count features http://code.google.com/p/nada-nonref-pronoun-detector/

17. Using web counts works great… but is it practical? All N - grams in the Google N - gram corpus 93 GB Extract N - grams of length - 4 only 33 GB Extract N - grams containing it, they, them only 500 MB Lower - case, truncate tokens to four 189 MB characters, replace special tokens (e.g. named entities, pronouns, digits) with symbols, etc. Encode tokens (6 bytes) and values (2 bytes), 44 MB store only changes from previous line gzip resulting file 33 MB

18. NADA versus Other Systems

19. Part 1: Conclusion • N-gram data better than search engines • Classifiers with N-gram counts are very effective, particularly on new domains • But we needed a large corpus of manually- annotated data to learn how to use the counts • We’ll see now how bilingual data can provide the supervision (for some problems)

20. Part 2: Coordination Ambiguity in NPs • [dairy andmeat]production • [sustainability]and [meat production] yes: [dairy production] in (1) no: [sustainability production] in (2) • new semantic features from raw web text and a new approach to using bilingual data as soft supervision [Bergsma, Yarowsky & Church, ACL 2011]

21. Coordination Ambiguity • Words whose POS tags match pattern: [DT|PRP$] (N.*|J.*) and [DT|PRP$] (N.*|J.*) N.* • Output: Decide if one NP or two • Resolving Coordination is classic hard problem • Treebank doesn’t annotate NP-internal structure • Modern parsers thus do very poorly on these decisions (78% Minipar, 79% for C&C parser) • For training/evaluation, we patched Treebank with Vadas & Curran ’07 NP annotations

22. One Noun Phrase or Two:A Machine Learning Approach Input: “dairy and meatproduction”→ features: x x= (…,first-noun=dairy, … second-noun=meat,… first+second-noun=dairy+meat, …) h(x) = w∙x(predict one NP if h(x) > 0) • Set w via training on annotated training data using some machine learning algorithm

23. Leveraging Web-Derived Knowledge [dairy andmeat]production • If there is only one NP, then it is implicitly talking about “dairy production” • Do we see this phrase occurring a lot on the web? [Yes] sustainability and [meat production] • If there is only one NP, then it is implicitly talking about “sustainability production” • Do we see this phrase occurring a lot on the web? [No] • Classifier has features for these counts • But the web can gives us more!

24. Features for Explicit Paraphrases ❶ and❷ ❸ ❶ and❷ ❸ New paraphrases extending ideas in [Nakov & Hearst, 2005]

25. Human-Annotated Data (small) Training Examples motor and heating fuels freedom and security agenda conservation and good management Google N-gram Data Feature Vectors x1, x2, x3, x4 Machine Learning Raw Data (HUGE) Classifier: h(x)

26. Using Bilingual Data • Bilingual data: a rich source of paraphrases dairy and meatproductionproducciónláctea y cárnica • Build a classifier which uses bilingual features • Applicable when we know the translation of the NP

27. Bilingual “Paraphrase” Features ❶ and❷ ❸ ❶ and❷ ❸

28. Bilingual “Paraphrase” Features ❶ and❷ ❸ ❶ and❷ ❸

29. Human-Annotated Data (small) Training Examples motor and heating fuels freedom and security agenda conservation and good management Translation Data Feature Vectors x1, x2, x3, x4 Machine Learning Bilingual Data (medium) Classifier: h(xb)

30. Training Examples + Features from Google Data h(xm) coal and steel money coal and steel money coal and steel money North and South Carolina North and South Carolina North and South Carolina rocket and mortar attacks rocket and mortar attacks rocket and mortar attacks pollution and transport safety pollution and transport safety pollution and transport safety business and computer science business and computer science business and computer science insurrection and regime change insurrection and regime change insurrection and regime change the environment and air transport the environment and air transport the environment and air transport the Bosporus and Dardanelles straits the Bosporus and Dardanelles straits the Bosporus and Dardanelles straits h(xb) Bitext Examples Training Examples + Features from Translation Data

31. Training Examples + Features from Google Data h(xm) North and South Carolina North and South Carolina North and South Carolina pollution and transport safety pollution and transport safety pollution and transport safety business and computer science business and computer science insurrection and regime change insurrection and regime change insurrection and regime change the environment and air transport the environment and air transport the Bosporus and Dardanelles straits the Bosporus and Dardanelles straits business and computer science the environment and air transport the Bosporus and Dardanelles straits h(xb)1 Training Examples + Features from Translation Data coal and steel money rocket and mortar attacks

32. Training Examples + Features from Google Data business and computer science the Bosporus and Dardanelles straits the environment and air transport h(xm)1 North and South Carolina North and South Carolina pollution and transport safety pollution and transport safety insurrection and regime change insurrection and regime change h(xb)1 Co-Training: [Yarowsky’95], [Blum & Mitchell’98] Training Examples + Features from Translation Data coal and steel money rocket and mortar attacks

33. Error rate (%) of co-trained classifiers h(xb)i h(xm)i

34. Error rate (%) on Penn Treebank (PTB) unsupervised h(xm)N 800 PTB training examples 800 PTB training examples 2 training examples

35. Part 2: Conclusion • Knowledge from large-scale monolingual corpora is crucial for parsing noun phrases • New paraphrase features • New way to use bilingual data as soft supervision to guide the use of monolingual features

36. Part 3: Using visual data to learn the meaning of words • Large volumes of visual data also reveal meaning (semantics), but in language-universal way • Humans label their images as they post them online, providing the word-meaning link • There’s lots of images to work with [from Facebook’s Twitter feed]

37. English Web Images Spanish Web Images cockatoo vela turtle cacatúa candle tortuga [Bergsma and Van Durme, IJCAI 2011]

38. Linking bilingual words by web-based visual similarity Step 1: Retrieve online images via Google Image Search (in each lang.), 20 images for each word • Google competitive with “hand-prepared datasets” [Fergus et al., 2005]

39. Step 2: Create Image Feature Vectors Color histogram features

40. Step 2: Create Image Feature Vectors SIFTkeypoint features Using David Lowe’s software [Lowe, 2004]

41. Step 3: Compute an Aggregate Similarity for Two Words Vector Cosine Similarity 0.33 0.55 Avg. over all English images Best match for one English image 0.19 0.46

42. Output: Ranking of Foreign Translations by Aggregate Visual Similarities • Lots of details in the paper: • Finding a class of words where this works (physical objects) • Comparing visual similarity to string similarity (cognate finder)

43. Task #2: Lexical Semantics from Images • Selectional Preference: • Is noun X a plausible object for verb Y? Can you eat “migas”? Can you eat “carillon”? Can you eat “mamey”? • [Bergsma and Goebel, RANLP 2011]

44. Conclusion • Robust NLP needs to look beyond human-annotated data to exploit large corpora • Size matters: • Many NLP systems trained on 1 million words • We use: • billions of words in bitexts • trillions of words of monolingual text • online images: hundreds of billions (⨯1000 words each  a 100 trillion words!)

45. Questions + Thanks • Gold sponsors: • Platinum sponsors (collaborators): • Kenneth Church (Johns Hopkins), Randy Goebel (Alberta), Dekang Lin (Google), Emily Pitler (Penn), Benjamin Van Durme (Johns Hopkins) and David Yarowsky (Johns Hopkins)