Generating referring expressions in a domain of objects and processes

Abstract

This thesis presents a collection of algorithms and data structures for the generation of pronouns, anaphoric definite noun phrases, and one-anaphoric phrases. After a close analysis of the particular kinds of referring expressions that appear in a particular domain -that of cookery recipes -the thesis presents an appropriate ontology and a corresponding representation language. This ontology is then integrated into a wider framework for language generation as a whole, whereupon we show how the representation language can be successfully used to produce appropriate referring expressions for a range of complex object types.Amongst the more important ideas explored in the thesis are the following:• We introduce the notion of a generalized physical object as a way of representing singular entities, mass entities, and entities which are sets.• We adopt the view that planning operators are essentially underspecified events, and use this, in conjunction with a simple model of the hearer, to allow us to determine the appropriate level of detail at which a given plan should be described.• We make use of a discourse model that distinguishes local and global focus, and is closely tied to a notion of discourse structure; and we introduce a notion of DISCRIMINATORY POWER as a means to choosing the content of a referring expression.• We present a model of the generation of referring expressions that makes use of two levels of intermediate representation, and integrate this model with the use of a linguistically- founded grammar for noun phrases.The thesis ends by making some suggestions for further extensions to the work reported here.