Semantic Selectional Restrictions for Disambiguating Meronymy Relations : A non-representational DRT-like discourse semantics

Abstract

We present an unsupervised approach to automatically learn lexico-syntactic patterns encoding meronymy relations from texts. Our major contribution lies in alleviating the challenge of disambiguating polysemous patterns that encode meronymy only in some contexts. We rely on the linking theory to posit that semantic features of the Part and Whole instances participating in a meronymy relation facilitate the identification of meronymy-encoding patterns. We abstract the instances to their hypernyms, enforcing semantic selectional restrictions to constrain the contexts within which patterns participate in meronymy. We disambiguate polysemous patterns using their contexts based on a modified version of Harris’ distributional hypothesis by postulating that similar patterns share similar contexts. Our experiments revealed that enforcing selectional restrictions enables detecting high-quality patterns. Furthermore, our method does not require annotated data, and has a broader coverage compared to previous studies.