Improvements to monolingual english word sense disambiguation

Abstract

Word Sense Disambiguation remains one of the most complex problems facing computational linguists to date. In this paper we present modification to the graph based state of the art algorithm In-Degree. Our modifications entail augmenting the basic Lesk similarity measure with more relations based on the structure of WordNet, adding SemCor examples to the basic WordNet lexical resource and finally instead of using the LCH similarity measure for computing verb verb similarity in the In-Degree algorithm, we use JCN. We report results on three standard data sets using three different versions of WordNet. We report the highest performing monolingual unsupervised results to date on the Senseval 2 all words data set. Our system yields a performance of 62.7% using WordNet 1.7.1.