Using distributional similarity to identify individual verb choice

Abstract

Human text is characterised by the individual lexical choices of a specific author. Significant variations exist between authors. In contrast, natural language generation systems normally produce uniform texts. In this paper we apply distributional similarity measures to help verb choice in a natural language generation system which tries to generate text similar to individual author. By using a distributional similarity (DS) measure on corpora collected from a recipe domain, we get the most likely verbs for individual authors. The accuracy of matching verb pairs produced by distributional similarity is higher than using the synonym outputs of verbs from WordNet. Furthermore, the combination of the two methods provides the best accuracy. © 2006 Association for Computational Linguistics.