Modeling Extemporaneous Elaboration

MODELING EXTEMPORANEOUS ELABORATION Marie A. Bienkowski Bell Communications Research 445 South Street, MRE 2P358 Morristown, NJ 07960-1961, USA ABSTRACT Intelligent problem solving systems must be able to ex- press their results in a coherent and flexible manner. One way this can be done is by eztemporaneous elaboration, the method of language production that underlies more skilled tasks such as explanation. This paper outlines a computa- tional model for extemporaneous elaboration that is imple- mented in a computer model called Extemper, shows ex- amples of its operation, and compares it with other models of language production. Extemper contains the four com- ponents minimally required for elaboration: 1) an efficient method for linearizing a knowledge structure, 2) a translation/selection mechanism for producing a conceptual textbase from the knowledge structure, 3) local coherence operators which provide local connections between textbase elements, and 4) a conceptual generator to translate the coherent textbase into English. Introduction An intelligent problem solving system may be required to participate in a purposive discourse with a user. Purposive discourse is dialogue where a welbdefined goal is pursued in a stereotyped and efficient - but not inflexible - way. Many common types of purposive discourse that could occur with an intelligent system, such as tutoring, explaining, and advis- ing, require lengthy elaborations. Automated reasoners or problem solvers, in particular, are often required to explain, describe, or justify, that is, to elaborate on, their solutions. The process of producing unrehearsed and unedited exposi- tions during purposive discourse is called ezternporaneous or spontaneous elaboration. This paper describes a computational model for extem- poraneous elaboration implemented in a computer model called Extemper. The model consists of four components. The first is a linearizer which provides the overall structure of the elaboration by either a) following a trace of the processing of a problem solver (rehashing) or, b) directing a rehearsal of the knowledge used by the system (i.e., instan- tinting the knowledge structures normally used for problem solving for use in language production). The second com- ponent is a set of selectors which determine what will be ex- pressed based on considering the type of discourse, the dis- course goals of the listener, and the relevance of domain- specific items. The linearizer and selectors produce the overall concep- tual form of the elaboration. It is given a coherent linguistic form by the combination of local coherence operators, which examine conceptual forms to track focus, add connective words, etc., and a sentence generator (Cullingford, 1986) which renders it in English. Extemper's core model of elaboration, then, consists of four components: a linearizer, selectors, local coherence operators, and a sentence gener- ator. These constitute a minimal cognitive architecture for elaboration: minimal because more may be needed for other elaboration types (e.g., a component to build and maintain a listener model for tutoring), and cognitive because com- ponents like these have been proposed by psychologists for models of language production (e.g., vanDijk and Kintsch, 1083). In the next section, I discuss how Extemper's components cooperate to …

Content truncated to 500 words