Interactive generation of plan descriptions and justifications
Susan Haller · 1995
Planning text to discuss domain plans interactively relates two areas of research. The first area, plan recognition, focuses on analyzing natural language that is about plans to recognize the speaker's intent and provide helpful responses. The second area, discourse planning, is concerned with using plans for selecting and structuring text that achieves communicative goals. Researchers have studied interaction more often from the perspective of plan recognition, using models where the user controls the discussion and the system's role is passive. Hence, researchers in plan recognition have developed systems for tracking discourse focus to recognize a user's domain plans and his discourse plans. When interaction is considered from the perspective of discourse planning, instead of analyzing the discourse to discover the user's intentions, systems plan to direct the discussion, making their own intentions clear. This process is called interactive generation. Situations that call for interactive generation arise when the system is acting as an expert advisor or a tutor which must be able to explain and justify what it says. During interactive generation, the system must make assumptions about the user's knowledge and his willingness to accept what is said. However, the interaction may indicate that these assumptions are wrong, and repairs must be planned. In this way, interactive generation leads to the production of efficient discussions that are tailored to the individual listener's needs. The Interactive Discourse Planner formulates and executes text plans incrementally to try to achieve its discourse goal. The system represents these plans uniformly in the same knowledge base with the domain plans that are under discussion. In this way, the text plan and the domain plan are both accessible for analyzing the listener's feedback for indications of how to continue planning text. Because text plans and domain plans are represented uniformly, it is possible to use the text plans as content for the discussion. As a testbed for my model, the system that I have implemented gives route advice interactively.