Defeasible Rules in Content Selection and Text Structuring

This paper outlines a number of ways in which defeasible rules can contribute to the content selection and discourse structuring components of a text generation system. We suggest that, for certain types of descriptive text, the characterisation of discourse struc-turing mechanisms as operations on or involving defeasible rules provides an attractive framework for addressing important issues in content selection/structuring. We describe an architecture which incorporates defeasible rules into a systemic model of generation, and illustrate its use in the description of objects in a museum gallery. While defeasible rules are traditionally used in theorem-proving applications, to make predictions about the consequences of known facts, we are here concerned with three separate issues: (i) how such rules may need to be expressed by the NLG system in order to achieve its goals; (ii) how their interaction with facts about particular objects enables the use of valuable coherence relations; and (iii) how they are relevant in taking account of the system's user model.