Pragmatic modeling in information system interfaces (goals, dialogue, plans, ill-formedness)

Natural language interfaces currently treat each query as an isolated request for information, with little use of the dialogue context within which the utterance occurs. This thesis investigates how natural language systems can assimilate an on-going dialogue and use the resulting knowledge to increase their robustness. The thesis first presents a strategy for inferring a model of the task-related plan motivating an information-seeker's queries. Focusing heuristics are used to relate each new utterance to the existing plan context and construct an enlarged context model. It then develops a pragmatics-based approach to handling two classes of problematic utterances: utterances involving pragmatic overshoot and intersentential elliptical fragments. The framework for handling pragmatic overshoot rephrases the pragmatically illformed query based on the speaker's perceived intentions in making the utterance. The ellipsis interpretation strategy identifies the discourse goal which the speaker is pursuing with the utterance and interprets it relative to the speaker's inferred task-related plan. The results of this research indicate that natural language interfaces must place greater emphasis upon established dialogue context, both the inferred task-related plan and anticipated conversational goals, in understanding utterances.