Generation of explanations and multi-turn discourse structures in tutorial dialogue, based on transcript analysis
Gregory A. Sanders · 1996
I have developed and implemented algorithms for a discourse-generator module to generate the discourse structure of tutorial dialogue. This English text will be the output of a new version of our Intelligent Tutoring System (ITS) called CircSim-Tutor. CircSim-Tutor communicates with the student in English text, displaying text on the screen and accepting it from the keyboard. I present studies of student initiatives in keyboard-to-keyboard tutoring dialogue, and the tutors' responses to them. I studied generation of extended summaries, explanations, and directed-line-of-reasoning (DLR) examples like those produced by our expert human tutors, Joel Michael and Allen Rovick. A DLR is a sequence of step-at-a-time questions and answers covering a chain of cause and effect. I suggest student understanding and recall of material will improve if CircSim-Tutor takes a consistent point of view in its explanations and summaries, as well as in question-and-answer exchanges with the students. I also suggest that CircSim-Tutor should make a point of telling students to reason in terms of physical cause-and-effect, not in teleological terms of purpose. I give a classification scheme for the student initiatives in our keyboard-to-keyboard transcripts, describe various studies of inter-rater agreement using that scheme, and discuss the results of those studies. In the opinion of all persons who used it to classify the initiatives, the classification does describe the student initiatives occurring in about 58 hours of keyboard-to-keyboard tutoring transcripts. I developed agorithms to generate summaries, explanations and multi-turn structures such as DLR exchanges. I have shown that important commonalities underlie all these tutorial discourse structures, particularly with respect to the knowledge the tutors must consult. Little or no previous work that we are aware of has been done on generating the multi-turn structures. In our keyboard-to-keyboard transcripts, DLRs serve various roles in tutoring: as summaries, as extended hints, and as a form of explaining a chain of cause and effect. I have also explained how to integrate DLRs into a proposed overall architecture of CircSim-Tutor.