Word Sense Disambiguation: A Structured Learning Perspective

Abstract

This paper explores the application of structured learning methods (SLMs) to word sense disambiguation (WSD). On one hand, the semantic dependencies between polysemous words in the sentence can be encoded in SLMs. On the other hand, SLMs obtained significant achievements in natural language processing, and so it is a natural idea to apply them to WSD. However, there are many theoretical and practical problems when SLMs are applied to WSD, due to characteristics of WSD. Beginning with the method based on hidden Markov model, this paper proposes for the first time a comprehensive and unified solution for WSD based on maximum entropy Markov model, conditional random field and tree-structured conditional random field, and reduces the time complexity and running time of the proposed methods to a reasonable level by beam search, approximate training, and parallel training. The update of models brings performance improvement, the introduction of one step dependency improves performance by 1--5 percent, the adoption of non-independent features improves performance by 2--3 percent, and the extension of underlying structure to dependency parsing tree improves performance by about 1 percent. On the English all-words WSD dataset of Senseval-2004, the method based on tree-structured conditional random field outperforms the best attendee system significantly. Nevertheless, almost all machine learning methods suffer from data sparseness due to the scarcity of sense tagged data, and so do SLMs. Besides improving structured learning methods according to the characteristics of WSD, another approach to improve disambiguation performance is to mine disambiguation knowledge from all kinds of sources, such as Wikipedia, parallel corpus, and to alleviate knowledge acquisition bottleneck of WSD.