ABSTRACT
Introduction
Although an intelligent tutoring system (ITS) and a human tutor may have the same goals, they do not have the same skills available for achieving those goals. Circsim-Tutor v. 3 is a natural-language based ITS which tutors students on problem solving in cardiac physiology. We examine a corpus containing over 5000 turns of human-to-human tutoring sessions in order to determine the salient features leading to student success. We determine which of those features can be replicated by an ITS and develop alternatives for the others. To reduce the input processing burden, we substitute mixed-initiative processing (Carbonell, 1970) for true cooperative conversation, short-answer questions for free-text input, and the use of explicit questions for the use of turn-taking rules. To obtain the most "bang for the buck" in text generation, we emphasize precise responses to student input, simulation of the discourse patterns of expert tutors, and the provision of variety in both pedagogy and language as substitutes for complex mental processing. Our goal is to keep the tutoring process as interactive as possible while providing both broader and deeper domain coverage.
