Automatic fine-grained semantic classification for domain adaptation

Abstract

Assigning arguments of verbs to different semantic classes ('semantic typing'), or alternatively, checking the 'selectional restrictions' of predicates, is a fundamental component of many natural language processing tasks. However, a common experience has been that general purpose semantic classes, such as those encoded in resources like WordNet, or handcrafted subject-specific ontologies, are seldom quite right when it comes to analysing texts from a particular domain. In this paper we describe a method of automatically deriving fine-grained, domain-specific semantic classes of arguments while simultaneously clustering verbs into semantically meaningful groups: the first step in verb sense induction. We show that in a small pilot study on new examples from the same domain we are able to achieve almost perfect recall and reasonably high precision in the semantic typing of verb arguments in these texts.