NATURAL LANGUAGE DIALOGUE IN AN INTEGRATED COMPUTATIONAL MODEL

Natural language dialogue is a continuous, unified phenomenon. Speakers use their conversational context to simplify individual utterances through a number of linguistic devices, including ellipsis and definite references. Yet most computational systems for using natural language treat individual utterances as separate entities, and have distinctly separate processes for handling ellipsis, definite references, and other dialogue phenomena.

The computational system presented here, Psli3, uses the uniform framework of a production system architecture to carry out natural language understanding and generation in a well-integrated way. This is demonstrated primarily using intersentential ellipsis resolution, in addition to examples of definite reference resolution and interactive error correction. The system's conversational context arises naturally as the result of the persistence of the internal representations of previous utterances in working memory. Natural language input is interpreted within this framework using a modification of the syntactic technique of chart parsing, extended to include semantics, and adapted to the production system architecture. It provides a graceful way of handling ambiguity within this architecture, and allows separate knowledge sources to interact smoothly across different utterances in a highly integrated fashion.

The design of this system demonstrates how flexible and natural user interactions can be carried out using a system with a naturally flexible control structure. A processing-based taxonomy for ellipsis resolution that we developed is used to analyze our coverage of intersentential ellipsis. The semantic chart parser is further extended to allow several closely related sentences to be treated in a single chart. This allows the relationship between the sentences to be used in a simple way to select between competing alternative interpretations, and provides a natural means of resolving complex elliptical utterances.

We describe this system in detail, and include a number of extensive examples of the system's processing during user interactions.