A WordNet-based Algorithm for Word Sense Disambiguation Xiaobin Li" Stan Szpakowicz and Stan Matwin Department of Computer Science Department of Computer Science Concordia University University of Ottawa Montreal, Quebec Ottawa, Ontario Canada H3G IMS Canada MN 6N5 xiaobmlQjcs concordia ca {szpak,stan}@csi uottawa ca Abstract interpretation By design we only need the user to ap• prove the system s findings or prompt it for alternatives We present an algorithm for automatic word Also hy design we limit ourselves to information source* sense disambiguation ba.sed on lexical knowl in the public domain inexpensive dictionaries and other edge contained in WordNet and on the results lexical sources such as WordNet of surface-syntactic analvsis The algorithm lb part of a system that analyzes texts in or WordNel [Mill* r, 1990 Beckwith et al , 1991] is a very der to acquire knowledge in the presence of as rich source of lexical knowledge Since most entries have little pre-coded semantic knowledge as possi multiple senses, we fare a severe problem of ambiguity blc On the other hand, we want lo make Hit The motivation for the work described here is the desire besl us* of public-domain information sources to design a word sense disambiguation (WSD) algorithm such as WordNet Rather than depend on large that satisfies the needs of our project (learning from text) amounts of hand-crafted knowledge or statis• without large amounts of hand-crafted knowledge or sta• tical data from large corpora, we use syntac• tistical data from large corpora We concentrate on using tic information and information in WordNet information in WordNet to the fullest extent and mini• and minimize the need for other knowledg< mizing the need for other knowledge sources in the WSD sources in the word sense disambiguation pro• algorithm Semantic similarity between words (defined cess We propose to guide disambiguation by in the next section) plays an important role in the algo• semantic similarity between words and heuris• rithm We propose several heuristic rules to guide WSD tic rules based on this similarity I lu algorithm Icsted on an unrestricted, real text (the Canadian In• has been applied to the ( anadian Income fax come Tax Guide) this automatic WSD method gives Guide Test results indicate that even on a rela- encouraging results tivelv small text the proposed method produces correct noun meaning more than 11% of the Word sense disambiguation is essential in natural lan• tim< guage processing Early symbolic methods [Hirst, 1967, Small and Rieger 1962, Wilks, 1975] heavily rely on large amounts of hand-crafted knowledge As a reault, 1 Introduction they can only work in a specific domain To overcome this weakness a variety of statistical WSD methods This work is part of the project that amih at a svner [Brown et a/, 1991, Gale et a/, 1992, Resnik, 1992, gistic integration of Machine Learning and Natural Lan• Schutze, 1992,Charniak, 1993, Lehman, 1994] have been guage Processing The long-t< rm goal of the project is put forward They scale up easily and this makes them a system that performs machine learning on the results useful for large unrestricted corpora One of the most of text analysis to acquire a useful collection of produc• important steps in statistical WSD methods, however, tion rules Because such a svstem should not require is statistically motivated extraction of word-word rela• extensive domain knowledge up front, text analysis is tionships from corpora As Resnik points out [1993], to be dont in a knowledge-scant setting and with mm the corpus may fail to provide sufficient information imal user involvement A domain-independent surface- about relevant word/word relationships and word/word syntactic parser produces an analysis of a text fragment relationships, even those supported by the data, may (usually a sentence) that undergoes interactive semantic not be the appropriate relationships to look at for some tasks" Most of the statistical methods suffer from this 'Tins work was done while the first author was with the Knowledge Acquisition and Machine Learning group at the limitation Resnik proposes constraining the set of pos• University of Ottawa The authors are gralcful to the Nat sible word classes by using WordNet Rather than im• ural Sciences and Engineering Research Council of Canada proving statistical approaches [Resnik, 1993, Sussna, far having supported this work with a Strategic Grant No 1993, Voorhees, 1993] we propose a completely differ• STR0117764 ent WordNet-based algorithm for WSD 1368 NATURAL LANGUAGE 2 WordNet and Semantic Similarity a child synset property belonging, holding, material WordNet is a lexical database with a remarkably broad possession" is not a synonym of its immediate parent coverage One of its most outstanding qualities is a svnset 'possession" at all Using the information about word sense network A word sense node in this net• a parent synset to decide the intended meaning of a word work is a group of strict synonyms called synset' which in its immediate child svnset is different from using the is regarded as a basic object in WordNet Each sense information about a child synset to decide the intended of a word is mapped to such a word sens*, node (1 e meaning of a word in its immediate parent synset So a synset) in WordNet and all word sense nodes in here we have divided the semantic similarity between a parent synset and its immediate child synset into two WordNet are linked by a \anety of semantic relation• levels (Level 2 and Level 3) ships, such as IS-A (hypernymy/hyponymy) PART-OF (meronymy/holonymy), synonymy and antonymy The Although we have only applied the algorithm to WSD IS-A relationship is dominant - synsels are grouped by of noun objects in a text (i e nouns that are objects of it into hicrarchies Our algorithm only accesses informa• verbs in sentences analyzed by, the system), it can also be tion about nouns and verbs, for which thtre exist such applied to other noun phrases in a sentence, in particular lexical hierarchies subjects It has become common to use some messure of se• In this approach, we must consider contexts that are mantic similarity between words to support word sense relevant lo our method and tht semantic similarity in disambiguation [Resmk 1992, Schutze, 1992] In this these contexts work, we have adopted the following defintion seman• Tor all practical purposes, the possible senses of a word tic similarity between words is inverselv proportional (o can be found in WordNet, but due to its extremely broad the semantic distance between words in a Wordnet IS coverage most words have multiple word senses A con• A hierarchy By investigating the semantic relationships text must be considered in order to decide a particular word sense The notion of context and its use could between two given words in WordNet hierarchies seman- differ widely across WSD methods One may consider tic similarity can be measured and roughly divided into a whole text, a 100-word window, a sentence or some the following four levels specific words nearby, and so on In our work, we as• Level 1 The words are strict synonynis according to sume that a group of words co-occurring in a sentence WordNet one word is in the same synset as the will determine earh other s sense despite each of them other word being multiply ambiguous A simple and effective way is Lo consider as context the verbs that dominate noun Level 2 The words are extended synonyms accord• objects in sentences That is, we investigate verb-noun ing to WordNet one word is the immediate parent pairs to determine tht intended meaning of noun objects node of the other word in a IS-A hierarchy of Word- in sentences Net When used to get the synonyms of a word, WordNet not only produces the strict synonyms (a In this work, then we focus on investigating two as• synset), but also the immediate parent nodes of this pects of semantic similarity synset in a IS-A hierarchy of WordNet Here, these • The semantic similarity of the noun objects immediate parent nodes are called extended syn- onyms • The semantic similarity of their verb contexts Level 3 Tht words are hyponyms according to Word- 3 Heuristic Rules and Confidence of Net one word is a child node of the other word in a IS-A hierarchy of WordNet Results Level 4 The words have a coordinate relationship in We have set out to determine the intended meaning of a WordNet a word is a sibling node of the other word noun object in a given verb context from all candidate (i e both words have tht same parent node) m a word senses in WordNet of the noun object, select one IS-A hierarchy of WordNet sense that best fits the given verb context Suppose that the algorithm seeks the intended mean• Level 1 concerns the semantic similarity between ing of a noun object NOBJ in its verb context VC (that words inside a synset, Level 2 - Level 4 describe the se• is, the intended meaning of NOBJ in a verb-object pair mantic similarity between words that, belong to different (VC, NOBJ)) NOBJ has n candidate word senses in synsets (a synset which is composed of a group of strict WordNet s(k) means the kth word sense of NOBJ in synonyms is a node in the hierarchy of WordNet) The WordNet, l<k<n SS means the semantic similarity be• semantic similarity at Level 1 and Level 4 are symmet tween words in a IS-A hierarchy of WordNet An arrow ric, but the semantic similarity at Level 2 and Level 3 is used to describe the relationship between words or be• are not symmetric although both are about the relation tween a word and its word