Generating Context Templates for Word Sense Disambiguation.

This paper presents a novel approach for generating context templates for the task of word sense disambiguation (WSD). Context information of an ambiguous word, in form of feature vectors, is first classified into coarse-grained semantic categories by topic features using the latent dirichlet allocation (LDA) algorithm. To further refine the sense tags, all feature vectors of the ambiguous word, under the same topic, are recast into a network. Various centrality measures are derived to figure out the features or context words in the context templates, which are highly influential in the disambiguation. The WSD is achieved by identifying the maximum pairwise similarities between the context encoded in the templates and the sentence. The correct sense of an ambiguous word is resolved by distinguishing the most activated template without being trapped in a subjective linguistic quagmire. The approach is assessed in a corpus of more than 1,000,000 words. Experimental result shows the best measures perform comparably to the state-of-the-art.