Interaction of discourse planning, instructional planning and dialogue management in an interactive tutoring system

We demonstrate the utility of natural language generation as the underlying model for an intelligent tutoring system (ITS) in cardiovascular physiology. We have achieved this goal by dividing it into three subgoals, each of which builds on its predecessor: (a) developing a model of the tutorial dialogue of human tutors based on current research in natural language generation, with emphasis on text planning and the Conversation Analysis school, (b) analyzing a corpus of human-to-human tutoring sessions in cardiovascular physiology in terms of this model, and (c) designing an ITS which implements the model. We develop an abstract model of tutorial dialogue in order to put text generation for ITSs on solid theoretical footing. We give a detailed analysis of our corpus using this model, including a discussion of how tutors sequence their corrections, begin and end phases of the discourse, acknowledge responses, reply to student errors, teach different kinds of information, provide hints, conduct interactive explanations and choose between domain models. We present a detailed design for an ITS which uses this model to show that it can be implemented with current technology. The system is divided into two routines running in parallel, a global tutorial planner, which makes discourse decisions for units larger than a turn, and a turn planner, which assembles individual turns. The tutorial planner does not generate text directly, but generates a series of semantic forms. The turn planner collects the semantic forms for a turn, which may include information from multiple tutorial and/or conversational goals, and generates text for them as a unit. This architecture promotes coherent dialogue while permitting the tutor to use multi-turn discourse plans and change plans in response to student input. We expect this model to produce longer, more complex, and more varied dialogues than previous work.