"ACQUISITION OF SELECTIONAL PREFERENCES IN NATURAL LANGUAGE PROCESSING"

Supervised by M.Hepple

Abstract

Selectional Restrictions could be thought as the semantic type constraints that a word sense imposes on the words, with which it combines in sentences. According to influential theory (Katz and Fodor, 1964) (see Resnik, 1993), these constraints are represented using a set of conceptual semantic patterns or vocabulary.

However, languages are continuously evolving and as the scope or domain of the word usage in the language increases, this set of patterns grows. And the task of accumulating and hand-coding them becomes very complicated. This has lead to research in the feasibility of using statistical methods to automatically acquire these patterns of selectional constraints, from on-line text corpus through the analysis of word co-occurrence.

Phillip Resnik (Resnik 1993, Resnik 1996) proposed such a statistical model that learns and formalises selectional preferences in probabilistic terms. Resnik models the selectional behaviour of a predicate, as its distributional effect on the conceptual classes to which its arguments belong. Where the measure of relative entropy from information theory is used to express this effect.

In this dissertation, I have executed a computational implementation of this model. The performance of the model is evaluated by comparing its predictions to judgements obtained from human subjects.