Significance of syntactic features for word sense disambiguation

Abstract

In this paper1 we explore the use of syntax in improving the performance of Word Sense Disambiguation(WSD) systems. We argue that not all words in a sentence are useful for disambiguating the senses of a target word and eliminating noise is important. Syntax can be used to identify related words and eliminating other words as noise actually improves performance significantly. CMU's Link Parser has been used for syntactic analysis. Supervised learning techniques have been applied to perform word sense disambiguation on selected target words. The Naive Bayes classifier has been used in all the experiments. All the major grammatical categories of words have been covered. Experiments conducted and results obtained have been described. Ten fold cross validation has been performed in all cases. The results we have obtained are better than the published results for the same data.