A Practical Solution to the Problem of Automatic Word Sense Induction

Recent studies in word sense induction are based on clustering global co-occurrence vectors, i.e. vectors that reflect the overall behavior of a word in a corpus. If a word is semantically ambiguous, this means that these vectors are mixtures of all its senses. Inducing a word’s senses therefore involves the difficult problem of recovering the sense vectors from the mixtures. In this paper we argue that the demixing problem can be avoided since the contextual behavior of the senses is directly observable in the form of the local contexts of a word. From human disambiguation performance we know that the context of a word is usually sufficient to determine its sense. Based on this observation we describe an algorithm that discovers the different senses of an ambiguous word by clustering its contexts. The main difficulty with this approach, namely the problem of data sparseness, could be minimized by looking at only the three main dimensions of the context matrices.