1. Word Sense Disambiguation 2002. 1. 18. Kyung-Hee Sung Foundations of Statistical NLP Chapter 7

2. 2 Contents  Methodological Preliminaries  Supervised Disambiguation –Bayesian classification / An information-theoretic approach –Disambiguation based on dictionaries, thesauri and bilingual dictionaries –One sense per discourse, one sense per collocation  Unsupervised Disambiguation

3. 3 Introduction  Word sense disambiguation –Many words have several meanings or senses. –Many words have different usages. Ex) Butter may be used as noun, or as a verb. –The task of disambiguation is done by looking at the context of the word’s use.

4. 4 Methodological Preliminaries (1/2) Supervised learningUnsupervised learning  We know the actual status (sense label) for each piece of data on which we train. (Learning from labeled data)  Classification task  We do not know the classification of the data in the training sample. (Learning from unlabeled data)  Clustering task

5. 5 Methodological Preliminaries (2/2)  Pseudowards : artificial ambiguous words –In order to test the performance of disambiguation algorithms. Ex) banana-door  Performance estimation –Upper bound : human performance –Lower bound (baseline) : the performance of the simplest possible algorithm, usually the assignment of all contexts to the most frequent sense.

6. 6 Supervised Disambiguation Bayesian classificationInformation Theory  It treats the context of occurrence as a bag of words without structure.  It integrates information from all words in the context window.  It looks at only one informative feature in the context, which may be sensitive to text structure.

7. 7 Notations SymbolMeaning wan ambiguous word s 1, …, s k, …, s K senses of the ambiguous word w (a semantic label of w) c 1, …, c i, …, c I contexts of w in a corpus v 1, …, v j, …, v J words used as contextual features for disambiguation

8. 8 Bayesian classification (1/2)  Assumption : we have a corpus where each use of ambiguous words is labeled with its correct sense.  Bayes decision rule : minimizes the probability of errors –Decide s´ if P(s´| c) > P(s k | c) ← using Bayes’s rule ← P(c) is constant for all senses

9. 9 Bayesian classification (2/2)  Naive Bayes independence assumption –All the structure and linear ordering of words within the context is ignored. → bag of words model –The presence of one word in the bag is independent of another. –  Decision rule for Naive Bayes –Decide s´ if ← computed by MLE

10. 10 An Information-theoretic approach (1/2)  The Flip-Flop algorithm applied to finding indicators for disambiguation between two senses. 1find random partition P = {P 1, P 2 } of {t 1, … t m } 2while (improving) do 3find partition Q = {Q 1, Q 2 } of {x 1, … x m } 4that maximize I(P; Q) 5find partition P = {P 1, P 2 } of {t 1, … t m } 6that maximize I(P; Q) 7end

11. 11 An Information-theoretic approach (2/2)  To translate prendre (French) based on its object –Translation, {t 1, … t m } = { take, make, rise, speak } –Values of Indicator, {x 1, … x m } = { mesure, note, exemple, décision, parole } 1. Initial partitionP 1 = { take, rise } P 2 = { make, speak } 2. Find partition Q 1, = { mesure, note, exemple} Q 2, = { décision, parole } ← This division gives us the most information for distinguishing P 1 from P 2 (maximizes I(P; Q)) 3. Find partition P 1, = { take } P 2, = { make, rise, speak } ← Always collect for take. Relations (English) take a measure take notes take an example make a decision make a speech rise to speak

12. 12 Dictionary-based disambiguation (1/2)  A word’s dictionary definitions are good indicators for the senses they define. for all senses s k of w do score(s k ) = overlap (D k, U vj in c E vj ) // number of common words end choose s´ s.t. s´= argmax sk score(s k ) SymbolMeaning D 1, …, D K dictionary definitions of the senses s 1, …, s K s j1, …, s jL senses of v j E vj dictionary definitions of a word v j / E vj = U ji D ji

13. 13 Dictionary-based disambiguation (2/2)  Simplified example : ash –The score is number of words that are shared by the sense definition and the context. SenseDefinition s1s1 treea tree of the olive family s2s2 burned stuff the solid residue left when combustible material is burned Scores Context s1s1 s2s2 01This cigar burns slowly and creates a stiff ash. 10The ash is on of the last trees to come into leaf.

14. 14 Thesaurus-based disambiguation (1/2)  Semantic categories of the words in a context determine the semantic categories of the context as a whole, and that this category in turn determines which word senses are used.  Each word is assigned one or more subject codes in the dictionary. –t(s k ) : subject code of sense s k. –The score is the number of words that compatible with the subject code of sense s k. for all senses s k of w do score(s k ) = Σ vj in c δ ( t(s k ), v j ) end choose s´ s.t. s´= argmax sk score(s k )

15. 15 Thesaurus-based disambiguation (2/2) WordSenseRoget categoryAccuracy bass[beis] musical [bæs] fish MUSIC ANIMAL, INSECT 99% 100% interestcuriosity advantage financial share RESIONING INJUSTICE DEBT PROPERTY 88% 34% 90% 38%  Self-interest (advantage) is not topic-specific.  When a sense is spread out over several topics, the topic- based classification algorithm fails.

16. 16 Disambiguation based on translations in a second-language corpus  In order to disambiguate an occurrence of interest in English (first language), we identify the phrase it occurs in and search a German (second language) corpus for instances of the phrase. –The English phrase showed interest : show(E) → ‘zeigen’(G) –‘zeigen’(G) will only occur with Interesse(G) since ‘legal shares’ are usually not shown. –We can conclude that interest in the phrase to show interest belongs to the sense attention, concern.

17. 17 One sense per discourse constraint  The sense of a target word is highly consistent within any given document. –If the first occurrence of plant is a use of the sense ‘living being’, then later occurrences are likely to refer to living beings too. for all documents d m do determine the majority sense s k of w in d m assign all occurrences of w in d m to s k end

18. 18 One sense per collocation constraint  Word senses are strongly correlated with certain contextual features like other words in the same phrasal unit. –Collocational features are ranked according to the ratio: –Relying on only the strongest feature has the advantage that no integration of different sources of evidence is necessary. ← The number of occurrences of sense s k with collocation f m

19. 19 Unsupervised Disambiguation (1/3)  There are situations in which even such a small amount of information is not available.  Sense tagging requires that some characterization of the senses be provided. However, sense discrimination can be performed in a completely unsupervised fashion.  Context-group discrimination : a completely unsupervised algorithm that clusters unlabeled occurrences.

20. 20 An EM algorithm (1/2) 1.Initialize the parameters of the model μ randomly. Compute the log likelihood of the corpus C 2. While l(C|μ) is improving repeat: (a) E-step. Estimate h ik ← Naive Bayes assumption

21. 21 An EM algorithm (2/2) (b) M-step. Re-estimate the parameters P(v j |s k ) and P(s k ) by way of MLE

22. 22 Unsupervised Disambiguation (2/3)  Unsupervised disambiguation can be easily adapted to produce distinctions between usage types. –Ex) The distinction between physical bank ( in the context of bank robberies ) banks as abstract corporations ( in the context of corporate mergers )  The unsupervised algorithm splits dictionary senses into fine-grained contextual variants. –Usually, the induced clusters do not line up well with dictionary senses. Ex) ‘lawsuit’ → ‘civil suit’, ‘criminal suit’

23. 23 Unsupervised Disambiguation (3/3)  Infrequent senses and senses that have few collocations are hard to isolate in unsupervised disambiguation.  Results of the EM algorithm –The algorithm fails for words whose senses are topic-independent such as ‘to teach’ for train. WordSense Accuracy Meanσ suitlawsuit the suit you wear 95 96 0000 trainline of railroad cars to teach 79 55 19 31 ← for ten experiments with different initial conditions