Modelling Speech Acts in Conversational Discourse

Computational pragmatics and dialogue analysis is currently a rapid growing area of research in computational linguistics. Over the past five years or so, initiatives in modelling pragmatic aspects of dialogue have led to considerably improved spoken language dialogue systems � so much so in fact that constrained human-computer interaction no longer seems out of the question. One of the main drawbacks to such systems however is highlighted by the word "constrained". Human communication is seldom confined to answering questions or solving problems within a restricted field (such as train timetable enquiries, or route finding, for instance). How can one tell whether theories of dialogue that work well in domain specific, task-oriented dialogue, can be scaled up or expanded to deal with natural conversation? In this dissertation I have carried out a critical survey of the various approaches to speech act modelling, detailing what I think are the strengths and weaknesses in the current theories. One very promising approach is that of using speech act analysis as a means of interpreting a speaker's intentions in producing an utterance. This then forms the basis for determining a hearer's response (following certain rules of conversational co-operation). I go on to present what is intended as a preliminary model, which is designed to capture the characteristic relationship and interaction of speech acts in conversational dialogue, especially those features which preceding research has failed to represent. Speech acts are defined by means of schemata that match the state of the prevailing conversational context space. Each possible contexgt space is specified in the model for the performance of a particular speech act or acts; the representation of the context space is then updated accordingly. I illustrate the theoretical model using real conversation, collected during the course of this research, and compare its performance against the analysis of a "benchmark" conversation, highlighting where the model falls short and how it could be improved in the future. I will argue that the model provides a powerful formalism for the characterisation of a wide variety of different basic speech acts