Exploiting Lexical Sensitivity in Performing Word Sense Disambiguation

Abstract

State-of-the-art research on word sense disambiguation (WSD) has demonstrated the superiority of supervised learning systems and necessity of multiple knowledge sources. However, despite the complex interaction observed between these two external factors, the intrinsic reason underlying such phenomena is not sufficiently understood. This calls for more qualitative analysis of disambiguation results from an interdisciplinary perspective. In this chapter, we explore the long realised lexical sensitivity issue in WSD in terms of concreteness, with reference to the context availability model in psycholinguistics and the Sketch Engine popularly used in lexicography. It will be shown that the “difficulty” of disambiguating a particular target word is a function of its information susceptibility, which depends on how the senses of the word were distinguished in the first place, thus leading to varied effectiveness of individual knowledge sources as observed. WSD could thus be treated as the reverse engineering of lexicography so that the use of knowledge sources, and the feature selection, could then be more informed with respect to individual words and their senses, and the combinations of algorithms and knowledge sources could be applied in a real lexically sensitive way.