From predicting Predominant senses to local context for word sense Disambiguation

Abstract

Recent work on automatically predicting the predominant sense of a word has proven to be promising (McCarthy et al., 2004). It can be applied (as a first sense heuristic) toWord Sense Disambiguation (WSD) tasks, without needing expensive hand-annotated data sets. Due to the big skew in the sense distribution of many words (Yarowsky and Florian, 2002), the First Sense heuristic forWSD is often hard to beat. However, the local context of an ambiguousword can give important clues to which of its senses was intended. The sense ranking method proposed by McCarthy et al. (2004) uses a distributional similarity thesaurus. The k nearest neighbours in the thesaurus are used to establish the predominant sense of a word. In this paper we report on a first investigation on how to use the grammatical relations the target word is involved with, in order to select a subset of the neighbours from the automatically created thesaurus, to take the local context into account. This unsupervised method is quantitatively evaluated on SemCor. We found a slight improvement in precision over using the predicted first sense. Finally, we discuss strengths and weaknesses of the method and suggest ways to improve the results in the future.