Linguistic Modeling of Dialogue Management

Yukinori TAKUBO and Setsuko ARITA

Faculty of Literature, Kyushu University
6-19-1 Hakozaki, Higashiku, Fukuoka, 812

We attempt to construct a performance model of speech, modeling the nature of the operations the speaker/hearer performs in speech recognition and interpretation, with the limitation of memory and various distracting factors impeding ideal delivery and perception. For this purpose, we have developed a theory of discourse management: a version of Mental Space Theory, modified to accommodate dialogic discourse by incorporating a memory management system.
We posit a cognitive interface between linguistic expressions and knowledge base. This interface contains pointers and index files linked to addresses in the knowledge base, controlling access paths to the data in the base. It also functions as a working memory buffer for the production and interpretation of utterances. Thus the interface controls the input-output operations in speech. Utterances in a dialogue exchange can be redefined as input-output operations via this interface: registering, searching, editing, etc.
In this paper we analyze particles indicating a topic, "WA" and "TTE" in the framework of Discourse Management Theory. We assume that these particles function as instruction for operations used to access data in the knowledge base. We are mainly concerned with the particle "WA", which has hitherto been treated as having various functions e.g. topic, contrast, range of value assignment, range of predication etc. We argue that these multifarious functions can be reduced one function: indicating the range of predication. We have shown how this function can be expressed as a series of steps in the discourse management model and demonstrated that our approach can capture the nature of this particle in a unifying manner.

Keywords: topic, "WA", discourse management, performance model