Alistair Knott
Human Communication Research Centre
University of Edinburgh

Title:  A feature-based account of the relations signalled by sentence/clause connectives

Computational theories of discourse often express the notion of text coherence in terms of constraints on the relationships between adjacent text spans. However, there remains much disagreement as to how these constraints should be defined.

In this talk, I'll begin by outlining a methodology for motivating and defining a set of coherence relations on the basis of a study of the set of sentence/clause connectives in a language.  The methodology rests on a conception of coherence relations as modelling psychological constructs operative in the processes of generating and interpreting text, and furthermore on the assumption that a language will contain resources for signalling such constructs explicitly.  The methodology comprises three stages: first, the gathering of a large corpus of connective phrases; second, the organisation of these connectives into a hierarchical taxonomy based on a pre-theoretical substitution test; and thirdly, the theoretical interpretation of the taxonomy. I'll concentrate on this final stage. The taxonomy lends itself to a conception of connectives (and relations) as composite constructs, defined in terms of a number of orthogonal features: I'll outline the feature definitions I'm currently thinking of, and discuss some of their advantages and shortcomings.

After the talk, I'd be happy to give a demo of the text generation system I've been working on, to illustrate how some of the theoretical ideas in the talk have been applied.

When:	1.00 pm, Wednesday 24 June

Where:	1st floor, Archway West Building