Planning Natural-Language Utterances To Satisfy Multiple Goals

Abstract

This dissertation presents the results of research on a planning formalism for a theory of natural-language generation that will support the generation of utterances that satisfy multiple goals. Previous research in the area of computer generation of natural-language utterances has concentrated two aspects of language production: (1) the process of producing surface syntactic forms from an underlying representation, and (2) the planning of illocutionary acts to satisfy the speaker’s goals. This work concentrates on the interaction between these two aspects of language generation and considers the overall problem to be one of refining the specification of an illocutionary act into a surface syntactic form, emphasizing the problems of achieving multiple goals in a single utterance. Planning utterances requires an ability to reason in detail about what the hearer knows and wants. A formalism, based on a possible-worlds semantics of an intentional logic of knowledge and action, was used for representing the effects of illocutionary acts and the speaker’s beliefs about the hearer’s knowledge of the world. Techniques are described that enable a planning system to use the representation effectively. The language-planning theory and knowledge representation are embodied in a computer system called KAMP (Knowledge and Modalities Planner), which plans both physical and linguistic actions, given a high-level description of the speaker’s goals. The research has application to the design of gracefully interacting computer systems, multiple-agent planning systems, and the planning of knowledge acquisition.