Unsupervised discovery of homograph senses using lexical context deconvolution

While much progress has been made in the automatic discovery of groups of synonymous words - or at least groups of semantically related words - very little effort has been made in the automatic discovery of the multiple meanings of polysemous words, of which the English language is full. The objective of this research, therefore, is the unsupervised discovery of the multiple senses of homographs from raw text without using any other sources of semantic or syntactic reference information. The proposed method finds compact clusters of semantically similar words - which are viewed to represent semantic classes - by applying a graph-theoretic clustering algorithm to the co-occurrence space derived from raw, unannotated text. Once these global semantic classes have been found, associating each homograph with them is treated as a problem of deconvolution. Semantic class posterior feature vectors are found by averaging the Bayesian posterior probabilities of each feature word across the members of that class, and are interpreted as context exemplars for unique word senses. A word's posterior feature vector, which includes all contexts in which the word appeared for each of its senses, is assumed to be a linear combination of the semantic class context exemplars. Thus, word-class associations are estimated by finding the non-negative least-squares solution to the system of linear equations formed by the word's and classes' posterior probability feature vectors. Preliminary results, using an assorted number of novels as a training corpus, show the ability of the method to not only discover multiple frequently used word senses, but also senses that are more obscure.