Organizing Dialogue From an Incoherent Stream of Goals

ORGANIZING DIALOGUE FROM AN INCOHERENT STREAM OF GOALS* ELISE H. TURNER Department of Computer Science University of New Hampshire Durham, NH, 03824 USA Abstract--Human discourse appears coherent when it re- fleets coherent human thought. However, computers do not necessarily store or process information in the same way that people do a.ud~ therefore, cannot rely on the structure of their reuouing for the structure of thehr dialogue,s. In- stead, computer-generated conversation must rely on some other mechanism for its orgsaiisation. In this paper, we discuss one such mechanism. We describe a template that provides a guide for conversation. The template is built from schemata representing discourse convention. As goals arrive from the problem solver they are added to the tem- plate. Because accepted discourse structures are used to connect a new goal to the existing template, goals are or- ganiaed into sub-groups that follow conventional, coherent patterns of discourse. We present JUDIS, an interface to a distributed problem solver that uses this approach to orga- nise dialogues f~om an incoherent stream of goals. I INTRODUCTION Conversation seems coherent and is easy to follow be- cause it reflects the way people think. When the speaker thinks coherently, his or her communication goals will be properly organised to follow linguistic con- vention. So s models of human language generation can allow domain goals to directly motivate conversation and add clue words only when the occasional utterance violates convention [5; 8]. However~ computer-generated conversation cannot rely on problem solving for its organisation. Some prob- lem solvers make no attempt to be "cognitively plausi- ble" and do not produce goals in sequences that would appear coherent to human users. The combined goals from a distributed problem solver where several inde- pendent reasoners use a single interface to communicate with the user are also likely to be incoherent. Even if individual problem solvers produce coherent streams of goals, the stream of goals from the aggregate is likely to switch back and forth between sub-problems that are being addressed by the individual systems. We call * Some of the work described in this paper was done at the Com- puter Science Depaxtment of Georgia Institute of Technology, AtlJnts, Georgia, 30332, supported by the NSF under llrmat lST-S608362. The &uther wishes to thank the anonymous re- viewers for their comments. the sequence of goals produced by such systems an in- coherent stream of goab because the goals are ordered in a way that would not seem reasonable to a human listener. Interfaces to such systems, while being re- sponsive to the goals of problem solving, must rely on something else to give dialogue its organisation. In this paper, we describe a template that can provide computer-generated conversation with a coherent orga- nisation while meeting the needs of the problem solvers. This template is built from schemata representing ex- pected discourse structure. These schemata include general linguistic conventions as well as expected or- ganisations for specific situations. Before conversation begins the template is very abstract, giving only broad descriptions …

Content truncated to 500 words