Word Senses and WordNet in Semantic Analysis

Corpora and Statistical Methods Lecture 6 Semantic similarity, vector space models and word- sense disambiguation

Part 2 Word sense disambiguation

What are word senses? Cognitive definition: mental representation of meaning used in psychological experiments relies on introspection (notoriously deceptive) Dictionary-based definition: adopt sense definitions in a dictionary most frequently used resource is WordNet

WordNet Taxonomic representation of words ( concepts ) Each word belongs to a synset, which contains near- synonyms Each synset has a gloss Words with multiple senses (polysemy) belong to multiple synsets Synsets organised by hyponymy (IS-A) relations

How many senses? Example: interest pay 3% interest on a loan showed interest in something purchased an interest in a company. the national interest have X s best interest at heart have an interest in word senses The economy is run by business interests

Wordnet entry for interest (noun) 1. a sense of concern with and curiosity about someone or something (Synonym: involvement) 2. the power of attracting or holding one s interest (Synonym: interestingness) 3. a reason for wanting something done ( Synonym: sake) 4. a fixed charge for borrowing money 5. a diversion that occupies one s time and thoughts (Synonym: pastime) 6. a right or legal share of something; a financial involvement with something (Synonym: stake) 7. (usually plural) a social group whose members control some field of activity and who have common aims, (Synonym: interest group)

Some issues Are all these really distinct senses? Is WordNet too fine- grained? Would native speakers distinguish all these as different? Cf. The distinction between sense ambiguity and underspecification (vagueness): one could argue that there are fewer senses, but these are underspecified out of context

Translation equivalence Many WSD applications rely on translation equivalence Given: parallel corpus (e.g. English-German) if word w in English has n translations in German, then each translation represents a sense e.g. German translations of interest: Zins: financial charge (WN sense 4) Anteil: stake in a company (WN sense 6) Interesse: all other senses

Some terminology WSD Task: given an ambiguous word, find the intended sense in context Sense tagging: task of labelling words as belonging to one sense or another. needs some a priori characterisation of senses of each relevant word Discrimination: distinguishes between occurrences of words based on senses not necessarily explicit labelling

Some more terminology Two types of WSD task: Lexical sample task: focuses on disambiguating a small set of target words, using an inventory of the senses of those words. All-words task: focuses on entire texts and a lexicon, where every word in the text has to be disambiguated Serious data sparseness problems!

Approaches to WSD All methods rely on training data. Basic idea: Given word w in context c learn how to predict sense s of w based on various features of w Supervised learning: training data is labelled with correct senses can do sense tagging Unsupervised learning: training data is unlabelled but many other knowledge sources used cannot do sense tagging, since this requires a priori senses

Supervised learning Words in training data labelled with their senses She pays 3% interest/INTEREST-MONEY on the loan. He showed a lot of interest/INTEREST-CURIOSITY in the painting. Similar to POS tagging given a corpus tagged with senses define features that indicate one sense over another learn a model that predicts the correct sense given the features

Features (e.g. plant) Neighbouring words: plant life manufacturing plant assembly plant plant closure plant species Content words in a larger window animal equipment employee automatic

Other features Syntactically related words e.g. object, subject . Topic of the text is it about SPORT? POLITICS? Part-of-speech tag, surrounding part-of-speech tags

Some principles proposed (Yarowsky 1995) One sense per discourse: typically, all occurrences of a word will have the same sense in the same stretch of discourse (e.g. same document) One sense per collocation: nearby words provide clues as to the sense, depending on the distance and syntactic relationship e.g. plant life: all (?) occurrences of plant+life will indicate the botanic sense of plant

Training data SENSEVAL Shared Task competition datasets available for WSD, among other things annotated corpora in many languages Pseudo-words create training corpus by artificially conflating words e.g. all occurrences of man and hammer with man-hammer easy way to create training data Multi-lingual parallel corpora translated texts aligned at the sentence level translation indicates sense

Data representation Example sentence: An electric guitar and bass player stand off to one side... Target word: bass Possible senses: fish, musical instrument... Relevant features are represented as vectors, e.g.: 1 2 2 , , , i i i POS w POS w , , , , w POS w POS + + + + 1 1 1 2 2 i i i i i guitar, NN, and, CC, player, NN, stand, VB

Supervised methods

Nave Bayes Classifier Identify the features (F) e.g. surrounding words other cues apart from surrounding context Combine evidence from all features Decision rule: decide on sense s iff s , : ' s ( | ' s ) ( | ) s P F P s F k k k Example: drug. F = words in context medication sense: price, prescription, pharmaceutical illegal substance sense: alcohol, illicit, paraphernalia

Using Bayes rule We usually don t know P(sk|f) but we can compute from training data: P(sk) (the prior) and P(f|sk) ( | P ) f ( ) P f s P s = ( | ) k k P s f k ( ) P(f) can be eliminated because it is constant for all senses in the corpus = arg max ( | ) best s P s f s k k ( | P ) f ( ) P f s P s = arg max k k s ( ) P k = arg max ( | ) ( ) P f s s s k k k

The independence assumption It s called na ve because: n = j = ( | ) ( | ) P f s P f s k j k 1 i.e. all features are assumed to be independent Obviously, this is often not true. e.g. finding illicit in the context of drug may not be independent of finding pusher. cf. our discussion of collocations! Also, topics often constrain word choice.

Training the naive Bayes classifier We need to compute: P(s) for all senses s of w ( , ) Count s w k= ( ) k P s ( ) Count w P(f|s) for all features f ( , ) Count f s j s k = ( | ) P f s j k ( ) Count k

Information-theoretic measures Find the single, most informative feature to predict a sense. E.g. using a parallel corpus: prendre (FR) can translate as take or make prendre une d cision: make a decision prendre une mesure:take a measure [to ] Informative feature in this case: direct object mesure indicates take d cision indicates make Problem: need to identify the correct value of the feature that indicates a specific sense.

Brown et als algorithm Given: translations T of word w 1. Given: values X of a useful feature (e.g. mesure, d cision as values of DO) 2. Step 1: random partition P of T 3. While improving, do: create partition Q of X that maximises I(P;Q) find a partition P of T that maximises I(P;Q) comment: relies on mutual info to find clusters of translations mapping to clusters of feature values 4.

Using dictionaries and thesauri Lesk (1986): one of the first to exploit dictionary definitions the definition corresponding to a sense can contain words which are good indicators for that sense Method: 1. Given: ambiguous word w with senses s1 snwith glosses g1 gn. 2. Given: the word w in context c 3. compute overlap between c & each gloss 4. select the maximally matching sense

Expanding a dictionary Problem with Lesk: often dictionary definitions don t contain sufficient information not all words in dictionary definitions are good informants Solution: use a thesaurus with subject/topic categories e.g. Roget s thesaurus http://www.gutenberg.org/cache/epub/22/pg22.html.utf8

Using topic categories Suppose every sense skof word w has subject/topic tk w can be disambiguated by identifying the words related to tk in the thesaurus Problems: general-purpose thesauri don t list domain-specific topics several potentially useful words can be left out e.g. Navratilova plays great tennis proper name here useful as indicator of topic SPORT

Expanding a thesaurus: Yarowsky 1992 Given: context c and topic t 1. For all contexts and topics, compute p(c|t) using Na ve Bayes by comparing words pertaining to t in the thesaurus with words in c if p(c|t) > , then assign topic t to context c 2. For all words in the vocabulary, update the list of contexts in which the word occurs. Assign topic t to each word in c 3. Finally, compute p(w|t) for all w in the vocabulary this gives the strength of association of w with t 4.

Yarowsky 1992: some results SENSE star space object celebrity shape ROGET TOPICS ACCURACY UNIVERSE ENTERTAINER INSIGNIA 96% 95% 82.% sentence punishment set of words LEGAL_ACTION 99% GRAMMAR 98%

Bootstrapping Yarowsky (1995) suggested the one sense per discourse/collocation constraints. Yarowsky s method: 1. select the strongest collocational feature in a specific context 2. disambiguate based only on this feature (similar to the information-theoretic method discussed earlier)

One sense per collocation For each sense s of w, initialise F, the collocations found in s s dictionary definition. 1. One sense per collocation: identify the set of contexts containing collocates of s for each sense s of w, update F to contain those collocates such that 2. ( | ) P s f for all s s ( | ' s ) P f (where alpha is a threshold)

One sense per discourse For each document: find the majority sense of w out of those found in previous step assign all occurrences of w the majority sense 3. This is implemented as a post-processing step. Reduces error rate by ca. 27%.

Unsupervised disambiguation

Preliminaries Recall: unsupervised learning can do sense discrimination not tagging akin to clustering occurrences with the same sense e.g. Brown et al 1991: cluster translations of a word this is akin to clustering senses

Brown et als method Preliminary categorisation: Set P(w|s) randomly for all words w and senses s of w. 1. Compute, for each context c of w the probability P(c|s) that the context was generated by sense s. 2. Use (1) and (2) as a preliminary estimate. Re-estimate iteratively to find best fit to the corpus.

Characteristics of unsupervised disambiguation Can adapt easily to new domains, not covered by a dictionary or pre-labelled corpus Very useful for information retrieval If there are many senses (e.g. 20 senses for word w), the algorithm will split contexts into fine-grained sets NB: can go awry with infrequent senses

Some issues with WSD

The task definition The WSD task traditionally assumes that a word has one and only one sense in a context. Is this true? Kilgarriff (1993) argues that co-activation (one word displaying more than one sense) is frequent: this would bring competition to the licensed trade competition = act of competing ; people/organisations who are competing

Systematic polysemy Not all senses are so easy to distinguish. E.g. competition in the agent competing vs act of competing sense. The polysemy here is systematic Compare bank/bank where the senses are utterly distinct (and most linguists wouldn t consider this a case of polysemy, but homonymy) Can translation equivalence help here? depends if polysemy is systematic in all languages

Logical metonymy Metonymy = usage of a word to stand for something else e.g. the pen is mightier than the sword pen = the press Logical metonymy arises due to systematic polysemy good cook vs. good book enjoy the paper vs enjoy the cake Should WSD distinguish these? How could they do this?

Which words/usages count? Many proper names are identical to common nouns (cf. Brown, Bush, ) This presents a WSD algorithm with systematic ambiguity and reduces performance. Also, names are good indicators of senses of neighbouring words. But this requires a priori categorisation of names. Brown s green stance vs. the cook s green curry