Indexing and Exploiting a Discourse History to Generate Context-Sensitive Explanations

INDEXING AND EXPLOITING A DISCOURSE HISTORY TO GENERATE CONTEXT-SENSITIVE EXPLANATIONS Johanna D. Moore* Department of Computer Science, and LRDC University of Pittsburgh Pittsburgh, PA 15260 ABSTRACT A slriking difference between the interactions that students have with human tutors and those they have with computer-based instruction systems is that human tutors frequently refer to their own previous explanations. Based on a study of human-human instructional inter- actions, we are categorizing the uses of previous discourse and are developing a computational model of this behavior. In this paper, I describe the slrategies we have implemented for identifying relevant prior explanations, and the mechanisms that enable our text planner to exploit the information stored in its discourse history in order to omit information that has previously been communicated, to point out similarities and differences between entities and situations, and to mark re-explanations in circumstances where they are deemed appropriate. 1. Introduction To reap the benefits of natural language interaction, user inter- faces must be endowed with the properties that make human natural language interaction so effective. One such prop- exty is that human speakers freely exploit all aspects of the mutually known context, including the previous discourse. Computer-generated utterances that do not draw on the previ- ous discourse seem awkward, unnatural, or even incoherent. The effect of the discourse history is especially important in instructional applications because explanation is essentially incremental and interactive. To provide missing informa- tion in a way that facilitates understanding and learning, the system must have the capability to relate new information effectively to recently conveyed material, and to avoid re- peating old material that would distract the student from what is new. Strategies for using the discourse history in generating utterances are therefore crucial for building computer systems intended to engage in instructional dialogues with their users. The goal of our work is to produce a computational model of the effects of discourse context on explanations in instruc- tional dialogues, and to implement this model in an intelli- gent explanation facility. Based on a study of human-human instructional dialogues, we are developing a taxonomy that classifies the types of contextual effects that occur in our data according to the explanatory functions they serve [1]. Thus *The reseat~h described in this paper was supported by the Office of Naval Research, Cognitive and Neural Sciences Division, and the National Science Foundation, Reseerch Initiation Award. far, we have focused on four categories from our taxonomy • explicit reference to a previous explanation (or portion thereof) in order to point out similarities (differences) between the material currently being explained and ma- terial presented in earfier explanation(s), • omission of previously explained material to avoid dis- tracting student from what is new, • explicit marking of repeated material to distinguish it from new material (e.g., "As I said before,... ") • elaboration of previous material in the form of general- izations, more detail, or justifications. 1 Building on previous work [2, 3] we have implemented an ex- planation facifity that maintains a discourse …

Content truncated to 500 words