Utilizing the One-Sense-per-Discourse Constraint for Fully Unsupervised Word Sense Induction and Disambiguation
Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC 2004)
Abstract
Recent advances in word sense induction rely on clustering related words. In this paper, instead of using a clustering algorithm, we suggest to perform a Singular Value Decomposition (SVD) which can be guaranteed to always find a global optimum. However, in order to apply this method to the problem of word sense induction, a semantic interpretation of the dimensions computed by the SVD is required. Our finding is that in our specific setting the first dimension relates to semantic similarities between words, and the second dimension distinguishes between the two main senses of an ambiguous word. Based on this result we present an algorithm for fully unsupervised word sense induction and disambiguation.