Unambiguous + Unlimited = Unsupervised Using the Web for Natural Language Processing Problems Marti Hearst School of Information, UC Berkeley Joint work with Preslav Nakov BYU CS Colloquium, Dec 6, 2007 This research supported in part by NSF DBI
Marti Hearst, BYU CS 2007 Natural Language Processing The ultimate goal: write programs that read and understand stories and conversations. This is too hard! Instead we tackle sub-problems. There have been notable successes lately: Machine translation is vastly improved Speech recognition is decent in limited circumstances Text categorization works with some accuracy
Marti Hearst, BYU CS 2007 How can a machine understand these differences? Get the cat with the gloves.
Marti Hearst, BYU CS 2007 How can a machine understand these differences? Get the sock from the cat with the gloves. Get the glove from the cat with the socks.
Marti Hearst, BYU CS 2007 How can a machine understand these differences? Decorate the cake with the frosting. Decorate the cake with the kids. Throw out the cake with the frosting. Throw out the cake with the kids.
Marti Hearst, BYU CS 2007 Why is this difficult? Same syntactic structure, different meanings. Natural language processing algorithms have to deal with the specifics of individual words. Enormous vocabulary sizes. The average English speaker’s vocabulary is around 50,000 words, Many of these can be combined with many others, And they mean different things when they do!
Marti Hearst, BYU CS 2007 How to tackle this problem? The field was stuck for quite some time. Hand-enter all semantic concepts and relations A new approach started around 1990 Get large text collections Compute statistics over the words in those collections There are many different algorithms.
Marti Hearst, BYU CS 2007 Size Matters Recent realization: bigger is better than smarter! Banko and Brill ’01: “Scaling to Very, Very Large Corpora for Natural Language Disambiguation”, ACL
Marti Hearst, BYU CS 2007 Example Problem Grammar checker example: Which word to use? Solution: use well-edited text and look at which words surround each use: I am in my third year as the principal of Anamosa High School. School-principal transfers caused some upset. This is a simple formulation of the quantum mechanical uncertainty principle. Power without principle is barren, but principle without power is futile. (Tony Blair)
Marti Hearst, BYU CS 2007 Using Very, Very Large Corpora Keep track of which words are the neighbors of each spelling in well-edited text, e.g.: Principal: “high school” Principle: “rule” At grammar-check time, choose the spelling best predicted by the surrounding words. Surprising results: Log-linear improvement even to a billion words! Getting more data is better than fine-tuning algorithms!
Marti Hearst, BYU CS 2007 The Effects of LARGE Datasets From Banko & Brill ‘01
Marti Hearst, BYU CS 2007 How to Extend this Idea? This is an exciting result … BUT relies on having huge amounts of text that has been appropriately annotated!
Marti Hearst, BYU CS 2007 How to Avoid Manual Labeling? “Web as a baseline” (Lapata & Keller 04,05) Main idea: apply web-determined counts to every problem imaginable. Example: for t in { } Compute f(w-1, t, w+1) The largest count wins
Marti Hearst, BYU CS 2007 Web as a Baseline Works very well in some cases machine translation candidate selection article generation noun compound interpretation noun compound bracketing adjective ordering But lacking in others spelling correction countability detection prepositional phrase attachment How to push this idea further? Significantly better than the best supervised algorithm. Not significantly different from the best supervised.
Marti Hearst, BYU CS 2007 Using Unambiguous Cases The trick: look for unambiguous cases to start Use these to improve the results beyond what co- occurrence statistics indicate. An Early Example: Hindle and Rooth, “Structural Ambiguity and Lexical Relations”, ACL ’90, Comp Ling’93 Problem: Prepositional Phrase attachment I eat/v spaghetti/n1 with/p a fork/n2. I eat/v spaghetti/n1 with/p sauce/n2. Question: does n2 attach to v or to n1?
Marti Hearst, BYU CS 2007 Using Unambiguous Cases How to do this with unlabeled data? First try: Parse some text into phrase structure Then compute certain co-occurrences f(v, n1, p) f(n1, p) f(v, n1) Problem: results not accurate enough The trick: look for unambiguous cases: Spaghetti with sauce is delicious. (pre-verbal) I eat with a fork. (no direct object) Use these to improve the results beyond what co- occurrence statistics indicate.
Marti Hearst, BYU CS 2007 Using Unambiguous Cases Hindle & Rooth, final algorithm: Parse text into phrase structure. Create bigram counts (v, p) and (n1, p) as follows: First, use unambiguous cases to populate bigram table Then, for the ambiguous cases: Compute a Lexical Association score comparing (v1, n1, p) to (n1, p, n2). If this is greater than a threshold, update the bigram table with the assumed attachment Else split the score and assign to both attachments The bigram table is used for further computations of the Lexical Association score.
Marti Hearst, BYU CS 2007 Unambiguous + Unlimited = Unsupervised Apply the Unambiguous Case Idea to the Very, Very Large Corpora idea The potential of these approaches are not fully realized Our work (with Preslav Nakov): Structural Ambiguity Decisions PP-attachment Noun compound bracketing Coordination grouping Semantic Relation Acquisition Hypernym (ISA) relations Verbal relations between nouns SAT Analogy problems
Marti Hearst, BYU CS 2007 Structural Ambiguity Problems Apply the U + U = U idea to structural ambiguity Noun compound bracketing Prepositional Phrase attachment Noun Phrase coordination Motivation: BioText project In eukaryotes, the key to transcriptional regulation of the Heat Shock Response is the Heat Shock Transcription Factor (HSF). Open-labeled long-term study of the subcutaneous sumatriptan efficacy and tolerability in acute migraine treatment. BimL protein interact with Bcl-2 or Bcl-XL, or Bcl-w proteins (Immuno- precipitation (anti-Bcl-2 OR Bcl-XL or Bcl-w)) followed by Western blot (anti-EEtag) using extracts human 293T cells co-transfected with EE- tagged BimL and (bcl-2 or bcl-XL or bcl-w) plasmids)
Marti Hearst, BYU CS 2007 Applying U + U = U to Structural Ambiguity We introduce the use of (nearly) unambiguous features: Surface features Paraphrases Combined with ngrams From very, very large corpora Achieve state-of-the-art results without labeled examples.
Marti Hearst, BYU CS 2007 Noun Compound Bracketing (a)[ [ liver cell ] antibody ] (left bracketing) (b)[ liver [cell line] ] (right bracketing) In (a), the antibody targets the liver cell. In (b), the cell line is derived from the liver.
Marti Hearst, BYU CS 2007 Dependency Model right bracketing: [w 1 [w 2 w 3 ] ] w 2 w 3 is a compound (modified by w 1 ) home health care w 1 and w 2 independently modify w 3 adult male rat left bracketing : [ [w 1 w 2 ]w 3 ] only 1 modificational choice possible law enforcement officer w 1 w 2 w 3
Marti Hearst, BYU CS 2007 Related Work Marcus(1980), Pustejosky&al.(1993), Resnik(1993) adjacency model:Pr(w 1 |w 2 ) vs. Pr(w 2 |w 3 ) Lauer (1995) dependency model:Pr(w 1 |w 2 ) vs. Pr(w 1 |w 3 ) Keller & Lapata (2004): use the Web unigrams and bigrams Girju & al. (2005) supervised model bracketing in context requires WordNet senses to be given Our approach: Web as data 2, n-grams paraphrases surface features
Marti Hearst, BYU CS 2007 Our U + U + U Algorithm Compute bigram estimates Compute estimates from surface features Compute estimates from paraphrases Combine these scores with a voting algorithm to choose left or right bracketing. We use the same general approach for two other structural ambiguity problems.
Marti Hearst, BYU CS 2007 Using n-grams to make predictions Say trying to distinguish: [home health] care home [health care] Main idea: compare these co-occurrence probabilities “home health” vs “health care”
Marti Hearst, BYU CS 2007 Computing Bigram Statistics Dependency Model, Frequencies Compare #(w 1,w 2 ) to #(w 1,w 3 ) Dependency model, Probabilities Pr(left) = Pr(w 1 w 2 |w 2 )Pr(w 2 w 3 |w 3 ) Pr(right) = Pr(w 1 w 3 |w 3 )Pr(w 2 w 3 |w 3 ) So we compare Pr(w 1 w 2 |w 2 ) to Pr(w 1 w 3 |w 3 ) w 1 w 2 w 3 left right
Marti Hearst, BYU CS 2007 Using ngrams to estimate probabilities Using page hits as a proxy for n-gram counts Pr(w 1 w 2 |w 2 ) = #(w 1, w 2 ) / #(w 2 ) #(w 2 ) word frequency; query for “w 2 ” #(w 1, w 2 ) bigram frequency; query for “w 1 w 2 ” smoothed by 0.5 Use 2 to determine if w 1 is associated with w 2 (thus indicating left bracketing), and same for w 1 with w 3
Marti Hearst, BYU CS 2007 Association Models: 2 (Chi Squared) A = #(w i, w j ) B = #(w i ) – #(w i, w j ) C = #(w j ) – #(w i, w j ) D = N – (A+B+C) N = 8 trillion (= A+B+C+D) 8 billion Web pages x 1,000 words
Marti Hearst, BYU CS 2007 Our U + U + U Algorithm Compute bigram estimates Compute estimates from surface features Compute estimates from paraphrases Combine these scores with a voting algorithm to choose left or right bracketing.
Marti Hearst, BYU CS 2007 Web-derived Surface Features Authors often disambiguate noun compounds using surface markers, e.g.: amino-acid sequence left brain stem’s cell left brain’s stem cell right The enormous size of the Web makes these frequent enough to be useful.
Marti Hearst, BYU CS 2007 Web-derived Surface Features: Dash (hyphen) Left dash cell-cycle analysis left Right dash donor T-cell right Double dash T-cell-depletion unusable…
Marti Hearst, BYU CS 2007 Web-derived Surface Features: Possessive Marker Attached to the first word brain’s stem cell right Attached to the second word brain stem’s cell left Combined features brain’s stem-cell right
Marti Hearst, BYU CS 2007 Web-derived Surface Features: Capitalization anycase – lowercase – uppercase Plasmodium vivax Malaria left plasmodium vivax Malaria left lowercase – uppercase – anycase brain Stem cell right brain Stem Cell right Disable this on: Roman digits Single-letter words: e.g. vitamin D deficiency
Marti Hearst, BYU CS 2007 Web-derived Surface Features: Embedded Slash Left embedded slash leukemia/lymphoma cell right
Marti Hearst, BYU CS 2007 Web-derived Surface Features: Parentheses Single-word growth factor (beta) left (brain) stem cell right Two-word (growth factor) beta left brain (stem cell) right
Marti Hearst, BYU CS 2007 Web-derived Surface Features: Comma, dot, semi-colon Following the first word home. health care right adult, male rat right Following the second word health care, provider left lung cancer: patients left
Marti Hearst, BYU CS 2007 Web-derived Surface Features: Dash to External Word External word to the left mouse -brain stem cell right External word to the right tumor necrosis factor- alpha left
Marti Hearst, BYU CS 2007 Web-derived Surface Features: Problems & Solutions Problem: search engines ignore punctuation in queries “brain-stem cell” does not work Solution: query for “brain stem cell” obtain 1,000 document summaries scan for the features in these summaries
Marti Hearst, BYU CS 2007 Other Web-derived Features: Possessive Marker We can also query directly for possessives Yes, “brain stem’s cell” sort of works. Search engines: drop the possessive marker but s is kept Still, we cannot query for “brain stems’ cell”
Marti Hearst, BYU CS 2007 Other Web-derived Features: Abbreviation After the second word tumor necrosis factor (NF) right After the third word tumor necrosis (TN) factor right We query for, e.g., “tumor necrosis tn factor” Problems: Roman digits: IV, VI States: CA Short words: me
Marti Hearst, BYU CS 2007 Other Web-derived Features: Concatenation Consider health care reform healthcare : 79,500,000 carereform : 269 healthreform: 812 Adjacency model healthcare vs. carereform Dependency model healthcare vs. healthreform Triples “healthcare reform” vs. “health carereform”
Marti Hearst, BYU CS 2007 Other Web-derived Features: Using Google’s * Operator Each * allows a one-word wildcard Single star “health care * reform” left “health * care reform” right More stars and/or reverse order “care reform * * health” right
Marti Hearst, BYU CS 2007 Other Web-derived Features: Reorder Reorders for “health care reform” “care reform health” right “reform health care” left
Marti Hearst, BYU CS 2007 Other Web-derived Features: Internal Inflection Variability Vary inflection of second word tyrosine kinase activation tyrosine kinases activation
Marti Hearst, BYU CS 2007 Other Web-derived Features: Switch The First Two Words Predict right, if we can reorder adult male rat as male adult rat
Marti Hearst, BYU CS 2007 Our U + U + U Algorithm Compute bigram estimates Compute estimates from surface features Compute estimates from paraphrases Combine these scores with a voting algorithm to choose left or right bracketing.
Marti Hearst, BYU CS 2007 Paraphrases The semantics of a noun compound is often made overt by a paraphrase (Warren,1978) Prepositional stem cells in the brain right cells from the brain stem left Verbal virus causing human immunodeficiency left Copula office building that is a skyscraper right
Marti Hearst, BYU CS 2007 Paraphrases Lauer(1995), Keller&Lapata(2003), Girju&al. (2005) predict NC semantics by choosing the most likely preposition: of, for, in, at, on, from, with, about, (like) This could be problematic, when more than one preposition is possible In contrast: we try to predict syntax, not semantics we do not disambiguate, just add up all counts cells in (the) bone marrow left cells from (the) bone marrow left
Marti Hearst, BYU CS 2007 Paraphrases prepositional paraphrases: We use: ~150 prepositions verbal paraphrases: We use: associated with, caused by, contained in, derived from, focusing on, found in, involved in, located at/in, made of, performed by, preventing, related to and used by/in/for. copula paraphrases: We use: is/was and that/which/who optional elements: articles: a, an, the quantifiers: some, every, etc. pronouns: this, these, etc.
Marti Hearst, BYU CS 2007 Paraphrases: pattern (1) (1)v n1 p n2 v n2 n1(noun) Can we turn “n1 p n2” into a noun compound “n2 n1”? meet/v demands/n1 from/p customers/n2 meet/v the customer/n2 demands/n1 Problem: ditransitive verbs like give gave/v an apple/n1 to/p him/n2 gave/v him/n2 an apple/n1 Solution: no determiner before n1 determiner before n2 is required the preposition cannot be to
Marti Hearst, BYU CS 2007 Paraphrases: pattern (2) (2)v n1 p n2 v p n2 n1(verb) If “p n2” is an indirect object of v, then it could be switched with the direct object n1. had/v a program/n1 in/p place/n2 had/v in/p place/n2 a program/n1 Determiner before n1 is required to prevent “n2 n1” from forming a noun compound.
Marti Hearst, BYU CS 2007 Paraphrases: pattern (3) (3)v n1 p n2 p n2 * v n1(verb) “*” indicates a wildcard position (up to three intervening words are allowed) Looks for appositions, where the PP has moved in front of the verb, e.g. I gave/v an apple/n1 to/p him/n2 to/p him/n2 I gave/v an apple/n1
Marti Hearst, BYU CS 2007 Paraphrases: pattern (4) (4)v n1 p n2 n1 p n2 v(noun) Looks for appositions, where “n1 p n2” has moved in front of v shaken/v confidence/n1 in/p markets/n2 confidence/n1 in/p markets/n2 shaken/v
Marti Hearst, BYU CS 2007 Paraphrases: pattern (5) (5)v n1 p n2 v PRONOUN p n2(verb) n1 is a pronoun verb (Hindle&Rooth, 93) Pattern (5) substitutes n1 with a dative pronoun (him or her), e.g. put/v a client/n1 at/p odds/n2 put/v him at/p odds/n2
Marti Hearst, BYU CS 2007 Paraphrases: pattern (6) (6)v n1 p n2 BE n1 p n2(noun) BE is typically used with a noun attachment Pattern (6) substitutes v with a form of to be (is or are), e.g. eat/v spaghetti/n1 with/p sauce/n2 is spaghetti/n1 with/p sauce/n2
Marti Hearst, BYU CS 2007 Our U + U + U Algorithm Compute bigram estimates Compute estimates from surface features Compute estimates from paraphrases Combine these scores with a voting algorithm to choose left or right bracketing.
Marti Hearst, BYU CS 2007 Evaluation: Datasets Lauer Set 244 noun compounds (NCs) from Grolier’s encyclopedia inter-annotator agreement: 81.5% Biomedical Set 430 NCs from MEDLINE inter-annotator agreement: 88% ( =.606)
Marti Hearst, BYU CS 2007 Evaluation: Experiments Exact phrase queries Limited to English Inflections: Lauer Set: Carroll’s morphological tools Biomedical Set: UMLS Specialist Lexicon
Marti Hearst, BYU CS 2007 Co-occurrence Statistics Lauer set Bio set
Marti Hearst, BYU CS 2007 Paraphrase and Surface Features Performance Lauer Set Biomedical Set
Marti Hearst, BYU CS 2007 Individual Surface Features Performance: Bio
Marti Hearst, BYU CS 2007 Individual Surface Features Performance: Bio
Marti Hearst, BYU CS 2007 Results Lauer
Marti Hearst, BYU CS 2007 Results: Comparing with Others
Marti Hearst, BYU CS 2007 Results Bio
Marti Hearst, BYU CS 2007 Results for Noun Compound Bracketing Introduced search engine statistics that go beyond the n-gram (applicable to other tasks) surface features paraphrases Obtained new state-of-the-art results on NC bracketing more robust than Lauer (1995) more accurate than Keller&Lapata (2004)
Marti Hearst, BYU CS 2007 Prepositional Phrase Attachment Problem: (a) Peter spent millions of dollars. (noun attach) (b) Peter spent time with his family. (verb attach) Which attachment for quadruple: (v, n1, p, n2) Results: Much simpler than other algorithms As good as or better than best unsupervised, and better than some supervised approaches
Marti Hearst, BYU CS 2007 Related Work Supervised (Brill & Resnik, 94): transformation-based learning, WordNet classes, P=82% (Ratnaparkhi & al., 94): ME, word classes (MI), P=81.6% (Collins & Brooks, 95): back-off, P=84.5% (Stetina & Makoto, 97): decision trees, WordNet, P=88.1% (Toutanova & al., 04): morphology, syntax, WordNet, P=87.5% Unsupervised (Hindle & Rooth, 93): partially parsed corpus, lexical associations over subsets of (v,n1,p), P=80%,R=80% (Ratnaparkhi, 98): POS tagged corpus, unambiguous cases for (v,n1,p), (n1,p,n2), classifier: P=81.9% (Pantel & Lin,00): collocation database, dependency parser, large corpus (125M words), P=84.3% Unsup. state-of-the-art
Marti Hearst, BYU CS 2007 PP-attachment: Our Approach Unsupervised (v,n1,p,n2) quadruples, Ratnaparkhi test set Google and MSN Search Exact phrase queries Inflections: WordNet 2.0 Adding determiners where appropriate Models: n-gram association models Web-derived surface features paraphrases
Marti Hearst, BYU CS 2007 N-gram models (i) Pr(p|n1) vs. Pr(p|v) (ii) Pr(p,n2|n1) vs. Pr(p,n2|v) I eat/v spaghetti/n1 with/p a fork/n2. I eat/v spaghetti/n1 with/p sauce/n2. Pr or # (frequency) smoothing as in (Hindle & Rooth, 93) back-off from (ii) to (i) N-grams unreliable, if n1 or n2 is a pronoun. MSN Search: no rounding of n-gram estimates
Marti Hearst, BYU CS 2007 Web-derived Surface Features Example features open the door / with a key verb (100.00%, 0.13%) open the door (with a key) verb (73.58%, 2.44%) open the door – with a key verb (68.18%, 2.03%) open the door, with a key verb (58.44%, 7.09%) eat Spaghetti with sauce noun (100.00%, 0.14%) eat ? spaghetti with sauce noun (83.33%, 0.55%) eat, spaghetti with sauce noun (65.77%, 5.11%) eat : spaghetti with sauce noun (64.71%, 1.57%) Summing achieves high precision, low recall. PRPR sum compare
Marti Hearst, BYU CS 2007 Paraphrases v n1 p n2 v n2 n1(noun) v p n2 n1(verb) p n2 * v n1(verb) n1 p n2 v(noun) v PRONOUN p n2(verb) BE n1 p n2(noun)
Marti Hearst, BYU CS 2007 Evaluation Ratnaparkhi dataset 3097 test examples, e.g. prepare dinner for family V shipped crabs from province V n1 or n2 is a bare determiner: 149 examples problem for unsupervised methods left chairmanship of the N is the of kind N acquire securities for an N special symbols: %, /, & etc.: 230 examples problem for Web queries buy % for 10 V beat S&P-down from % V is 43%-owned by firm N
Marti Hearst, BYU CS 2007 Results Simpler but not significantly different from 84.3% (Pantel&Lin,00). For prepositions other then OF. (of noun attachment) Models in bold are combined in a majority vote.
Marti Hearst, BYU CS 2007 Noun Phrase Coordination (Modified) real sentence: The Department of Chronic Diseases and Health Promotion leads and strengthens global efforts to prevent and control chronic diseases or disabilities and to promote health and quality of life.
Marti Hearst, BYU CS 2007 NC coordination: ellipsis Ellipsis car and truck production means car production and truck production No ellipsis president and chief executive All-way coordination Securities and Exchange Commission
Marti Hearst, BYU CS 2007 NC Coordination: ellipsis Quadruple (n1,c,n2,h) Penn Treebank annotations ellipsis: (NP car/NN and/CC truck/NN production/NN). no ellipsis: (NP (NP president/NN) and/CC (NP chief/NN executive/NN)) all-way: can be annotated either way This is a problem a parser must deal with. Collins’ parser always predicts ellipsis, but other parsers (e.g. Charniak’s) try to solve it.
Marti Hearst, BYU CS 2007 Results 428 examples from Penn TB
Marti Hearst, BYU CS 2007 New Application: Machine Translation Main idea: Use syntactic paraphrases of source sentences to create more training data examples for the same target translation. Still working on this; starting to get measurable improvements
Marti Hearst, BYU CS 2007 Semantic Relation Detection Goal: automatically augment a lexical database Many potential relation types: ISA (hypernymy/hyponymy) Part-Of (meronymy) Idea: find unambiguous contexts which (nearly) always indicate the relation of interest
Marti Hearst, BYU CS 2007 Lexico-Syntactic Patterns
Marti Hearst, BYU CS 2007 Lexico-Syntactic Patterns
Marti Hearst, BYU CS 2007 Adding a New Relation
Marti Hearst, BYU CS 2007 Semantic Relation Detection Lexico-syntactic Patterns: Should occur frequently in text Should (nearly) always suggest the relation of interest Should be recognizable with little pre-encoded knowledge. These patterns have been used extensively by other researchers.
Marti Hearst, BYU CS 2007 Semantic Relation Detection What relationship holds between two nouns? olive oil – oil comes from olives machine oil – oil used on machines Assigning the meaning relations between these terms has been seen as a very difficult solution Our solution: Use clever queries against the web to figure out the relations.
Marti Hearst, BYU CS 2007 Queries for Semantic Relations Convert the noun-noun compound into a query of the form: noun2 that * noun1 “oil that * olive(s)” This returns search result snippets containing interesting verbs. In this case: Come from Be obtained from Be extracted from Made from …
Marti Hearst, BYU CS 2007 Uncovering Semantic Relations More examples: Migraine drug -> treat, be used for, reduce, prevent Wrinkle drug -> treat, be used for, reduce, smooth Printer tray -> hold, come with, be folded, fit under, be inserted into Student protest -> be led by, be sponsored by, pit, be, be organized by
Marti Hearst, BYU CS 2007 Conclusions Unambiguous + Unlimited = Unsupervised The enormous size of the web opens new opportunities for text analysis There are many words, but they are more likely to appear together in a huge dataset This allows us to do word-specific analysis To counter the labeled-data roadblock, we start with unambiguous features that we can find naturally. We’ve applied this to structural and semantic language problems. These are stepping stones towards sophisticated language understanding.
Thank you! Supported in part by NSF DBI