Technical reports

A relevance-based utterance processing system

Victor Poznański

February 1992, 295 pages

This technical report is based on a dissertation submitted December 1990 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Girton College.

DOI: 10.48456/tr-246

Abstract

This thesis presents a computational interpretation of Sperber and Wilson’s relevance theory, based on the use of non-monotonic logic supported by a reason maintenance system, and shows how the theory, when given a specific form in this way, can provide a unique and interesting account of discourse processing.

Relevance theory is a radical theory of natural language pragmatics which attempts to explain the whole of human cognition using a single maxim: the Principle of Optimal Relevance. The theory is seen by its originators as a computationally more adequate alternative to Gricean pragmatics. Much as it claims to offer the advantage of a unified approach to utterance comprehension, Relevance Theory is hard to evaluate because Sperber and Wilson only provide vague, high-level descriptions of vital aspects of their theory. For example, the fundamental idea behind the whole theory is that, in trying to understand an utterance, we attempt to maximise significant new information obtained from the utterance whilst consuming as little cognitive effort as possible. However, Sperber and Wilson do not make the nature of information and effort sufficiently clear.

Relevance theory is attractive as a general theory of human language communication and as a potential framework for computational language processing systems. The thesis seeks to clarify and flesh out the problem areas in order to develop a computational implementation which is used to evaluate the theory.

The early chapters examine and criticise the important aspects of the theory, emerging with a schema for an ideal relevance-based system. Crystal, a computational implementation of an utterance processing system based on this schema is then described. Crystal performs certain types of utterance disambiguation and reference resolution, and computes implicatures according to relevance theory.

An adequate reasoning apparatus is a key component of a relevance based discourse processor, so a suitable knowledge representation and inference engine are required. Various candidate formalisms are considered, and a knowledge representation and inference engine based on autoepistemic logic is found to be the most suitable. It is then shown how this representation can be used to meet particular discourse processing requirements, and how it provides a convenient interface to a separate abduction system that supplies not demonstrative inferences according to relevence theory. Crystal’s powers are illustrated with examples, and the thesis shows how the design not only implements the less precise areas of Sperber and Wilson’s theory, but overcomes problems with the theory itself.

Crystal uses rather crude heuristics to model notions such as salience and degrees of belief. The thesis thefore presents a proposal and outline for a new kind of reason maintenance system that supports non-monotonic logic whose formulae re labelled with upper/lower probability ranges intended to represent strength of belief. This system should facilitate measurements of change in semantic information and shed some light on notions such as expected utility and salience.

The thesis concludes that the design and implementation of crystal provide evidence that relevance theory, as a generic theory of language processing, is a viable alternative theory of pragmatics. It therefore merits a greater level of investigation than has been applied to it to date.

Full text

PDF (18.5 MB)

BibTeX record

@TechReport{UCAM-CL-TR-246,
  author =	 {Pozna{\'n}ski, Victor},
  title = 	 {{A relevance-based utterance processing system}},
  year = 	 1992,
  month = 	 feb,
  url = 	 {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-246.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  doi = 	 {10.48456/tr-246},
  number = 	 {UCAM-CL-TR-246}
}