The Predictive Power of the (Micro)Context Revisited – Behavioral Profiling and Word Sense Disambiguation.

Abstract

One of the most pressing issues in lexical semantics is the lack of solid empirical (linguistic) criteria in accounting for sense distinction. The methodology explored in the current paper starts from the premise that, if not a complete, then at least a largely criteria-based account of word senses is possible by approaching word sense discrimination through a combination of supervised and unsupervised WSD. We aim to test this claim, which has also been raised by the relatively recently re-emerging corpus-based bag-of-words approaches to WSD. The paper concludes that using exclusively the criteria of linguistic (microcontextual) data is not sufficient, distinctive, and useful for successful WSD and reveals a solely linguistic account as not applicable in practice.

Keywords: corpus-based semantics, word sense disambiguation, Behavioral Profiling, predictive power, prediction ratio