| The Turn Planner is a set of rules for planning conversational turns in a tutorial dialogue. The dialogues are carried out by CIRCSIM-Tutor, a tutoring system designed to teach medical students to solve problems using causal reasoning. The current version of the system generates sentences one at a time. The resulting output is understandable—medical students are using the system and learning from it—but the turns are not nearly as coherent as the turns generated in human dialogues.; The design of our system is entirely based on the analysis of human tutorial dialogues produced by professors of physiology at Rush Medical College. The central core of our system is an Instructional Planner, which calls the Input Understander, the Student Modeler, the Problem-Solver, and the Domain Knowledge Base to help it decide how to respond to the student. As soon as it decides on the content of the response, it calls the Discourse Planner to plan how to phase the output. In the current version, the Discourse Planner calls the Surface Generator to produce the output a sentence at a time.; In the version of the system under construction the Discourse Planner will call the Turn Planner to organize a coherent turn and the Turn Planner will call the Surface Generator. Planning in the new system will be carried out by Freedman's Atlas Planner. Therefore, the Turn Planner is implemented as Atlas planning operators. We describe the architecture of the turn planner. Then we discuss a new method of lexical analysis that allows the researcher to visualize the alternatives conveniently. Rules for discourse markers, variable descriptions, softeners and acknowledgements have been obtained by marking up human dialogues and applying machine learning techniques. Finally, we discuss alternative evaluation strategies.; Turn planning is a new concept since the need was not perceived until computers could hold conversations effectively, but it seems to make a real difference to the quality of our system output. Since our ideas are not domain dependent, these results are applicable to other natural language dialogue systems, especially Socratic dialogue systems. |
