Abstract

Despite the advances in information processing systems, word-sense disambiguation tasks are far to be satisfactory as testified by numerous limitations of current translation systems and text inference systems. This paper attempts to investigate new techniques in knowledge based word-sense disambiguation field. First, by exploring the WordNet lexical database and part-of-speech conversion through the established CatVar database that translates all non-noun words into their noun counterparts, and following the spirit of Lesk’s disambiguation algorithm, a new disambiguation algorithm that maximizes the overall semantic similarity in the sense of Wu and Palmer measure between each sense of the target word and synsets of words of the context, is established. Second, motivated by the existence of WordNet domains for individual synsets, an overlapping based approach that quantifies the set intersection of synset domains, if not empty, or the hierarchy structure of the domains links through a simple path-length measure is put forward. Third, instead of exploring the whole set of words involved in the context, a selective approach that uses syntactic feature as outputted by Stanford Parser and a fixed length windowing is developed. The developed algorithms are evaluated according to two commonly employed dataset where a clear improvement to the baseline algorithm has been acknowledged.