Simple features for statistical Word Sense Disambiguation

Simple Features for Statistical Word Sense Disambiguation Abolfazl K. Lamjiri, Osama El Demerdash, Leila Kosseim CLaC Laboratory Department of Computer Science Concordia University, Montreal, Canada {a keigho,osama el,kosseim}@cs.concordia.ca Abstract In this paper, we describe our experiments on statistical word sense disambiguation (WSD) using two systems based on different ap- proaches: Na¨ıve Bayes on word tokens and Max- imum Entropy on local syntactic and seman- tic features. In the first approach, we consider a context window and a sub-window within it around the word to disambiguate. Within the outside window, only content words are con- sidered, but within the sub-window, all words are taken into account. Both window sizes are tuned by the system for each word to disam- biguate and accuracies of 75% and 67% were re- spectively obtained for coarse and fine grained evaluations. In the second system, sense res- olution is done using an approximate syntac- tic structure as well as semantics of neighbor- ing nouns as features to a Maximum Entropy learner. Accuracies of 70% and 63% were ob- tained for coarse and fine grained evaluations. 1 Introduction In this paper, we present the two systems we built for our first participation in the English lexical sample task at Senseval-3. In the first system, a Na¨ıve Bayes learner based on context words as features is implemented. In the second system, an approximate syntactic structure, in addition to semantics of the nouns around the ambiguous word are selected as features to learn with a Maximum Entropy learner. In Section 2, a brief overview of related work on WSD is presented. Sections 3 and 4 provide specifications of our two systems. Section 5 discusses the results obtained and remarks on them, and finally in Section 6, conclusion and our future work direction are given. 2 Related Work In 1950, Kaplan carried out one of the earliest WSD experiments and showed that the accu- racy of sense resolution does not improve when more than four words around the target are considered (Ide and V´eronis, 1998). While re- searchers such as Masterman (1961), Gougen- heim and Michea (1961), agree with this ob- servation (Ide and V´eronis, 1998), our results demonstrate that this does not generally ap- ply to all words. A large context window pro- vides domain information which increases the accuracy for some target words such as bank.n, but not others like different.a or use.v (see Sec- tion 3). This confirms Mihalcea’s observations (Mihalcea, 2002). In our system we allow a larger context window size and for most of the words such context window is selected by the system. Another trend consists in defining and us- ing semantic preferences for the target word. For example, the verb drink prefers an ani- mate subject in its imbibe sense. Boguraev shows that this does not work for polysemous verbs because of metaphoric expressions (Ide and V´eronis, 1998). Furthermore, the grammatical structures the target word takes part in can be used as a distin- guishing tool: “the word ‘keep’, can be disam- biguated …

Content truncated to 500 words