DISCOVERING SEMANTIC REGULARITY IN LEXICAL RESOURCES

Abstract

We first describe four varieties of thesaurus: (1) Roget-style, produced to help people find synonyms when they are writing; (2) WordNet and EuroWordNet; (3) thesauruses produced (manually) to support information retrieval systems; and (4) thesauruses produced automatically from corpora. We then contrast thesauruses and dictionaries, and present a small experiment in which we look at polysemy in relation to thesaurus structure. It has sometimes been assumed that different dictionary senses for a word that are close in meaning will be near neighbours in the thesaurus. This hypothesis is explored, using as inputs the hierarchical structure of WordNet 1.5 and a mapping between WordNet senses and the senses of another dictionary. The experiment shows that pairs of `lexicographically close' meanings are frequently found in different parts of the hierarchy. These pairs often display regular polysemic relations, i.e. they constitute systematic sense combinations that are valid for more than one word. The second experiment is an attempt to gain more insight into the mechanisms that underlie lexicalized regular polysemy. The hierarchical structure of WordNet is exploited to create a working definition of systematic polysemy and extract regular polysemic patterns at a level of semantic generalization that allows the identification of fine-grained semantic relations between the senses of the words participating in the regular polysemic pattern.