Discourse Plans, Intented Eeects and the Generation of Implicatures

A number of systems have been developed for discourse generation that use hierarchical planning techniques to generate discourse plans [Appelt 1983,Cawsey 1993,Maybury 1992,Hovy 1991,Moore and Paris 1993]. These plan-based approaches provide various solutions to the problems of content and structure determination: what communicative acts should appear in a discourse plan, what are the relationships between those acts and in what relative order should the acts occur. Can the techniques used by these approaches be used to generate implicatures related to discourse content and structure? If so, what representational extensions do they require? Thomason [Thomason 1990] advocates the use of a plan-based approach to generate implicatures in discourse. As a step in this direction, he provides a de nition of meaning in terms of the intentions of a speaker and the recognition of the speaker's discourse plan by the hearer. However, he leaves unspeci ed a computational model responsible for representing these ideas in a discourse planning system. By and large, previous discourse planning systems have been ill-equipped for the task of representing intentions, much less intentions of the type that Thomason proposes. Obviously, a discourse planning system can only take advantage of those concepts that are explicitly represented as preconditions and e ects of its actions. The representational requirements for the generation of implicature as described by Thomason center on two concepts: intention and recognition. Some headway has been made towards the representation of the rst. I will discuss the signi cance of intention in discourse planning below. About the second I will have little to say { the de nition of a representation of plan recognition useful for plan generation remains an open problem for discourse planners both in terms of its representational requirements and its computational tractability. Perhaps due in part to the representational limitations of previous discourse planners, several recent research e orts [Green 1994,Young 1995a] have described approaches to plan-based discourse generation that produce implicatures not by using a planning mechanism but by using an external editing process that modi es either the input to or output from a discourse planner. I argue that this work, while e ective at producing discourse that exploits implicature, fails to take advantage of the power of the planning algorithm to reason about action { to add content in response to the goals of a discourse plan and to structure the plan as needed to establish and maintain those goals. Without this ability, these systems are limited in a number of ways I summarize below.