Persuasive and adaptive tutorial dialogues for a medical diagnosis tutoring system
The objective of this thesis is to address a key problem in the development of an intelligent tutoring system, that is, the implementation of the verbal exchange (a dialogue) that takes place between a student and the system. Here we consider TeachMed, a medical diagnosis tutoring system that teaches the students to diagnose clinical problems. However, approaches that are presented could also fit other tutoring systems. In such a system, a dialogue must be implemented that determines when and how pedagogic aid is provided to the student, that is, what to say to her, in what circumstances, and how to say it. Finite state machines and automated planning systems are so far the two most common approaches for implementing tutoring dialogues in intelligent tutoring systems. In the former approach, finite state machines of dialogues are manually designed and hard coded in intelligent tutoring systems. This is a straightforward but very time consuming approach. Furthermore, any change or extension to the hard coded finite state machines is very difficult as it requires reprogramming the system. On the other hand, automated planning has long been presented as a promising technique for automatic dialogue generating. However, in existing approaches, the requirement for the system to persuade the student is not formally acknowledged. Moreover, current dialogue planning approaches are not able to reason on uncertainties about the student's knowledge. This thesis presents two approaches for generating more effective tutorial dialogues.The first approach describes an argumentation framework for implementing persuasive tutoring dialogues. In this approach the entire interaction between the student and the tutoring system is seen as argumentation.The tutoring system and the student can settle conflicts arising during their argumentation by accepting, challenging, or questioning each other's arguments or withdrawing their own arguments. Pedagogic strategies guide the tutoring system by selecting arguments aimed at convincing the student.The second approach presents a non-deterministic planning technique which models the dialogue generation problem as one of planning with incomplete knowledge and sensing. This approach takes into account incomplete information about a particular fact of the student's knowledge by creating conditional branches in a dialogue plan such that each branch represents an adaptation of the dialogue plan with respect to a particular state of the student's knowledge or belief concerning the desired fact. In order to find out the real state of the student's knowledge and to choose the right branch at execution time, the planner includes some queries in the dialogue plan so that the tutoring system can ask the student to gather missing information. One contribution in this thesis is improving the quality of tutoring dialogues by engaging students in argumentative interactions and/or adapting the dialogues with respect to the student's knowledge. Another one is facilitating the design and implementation of tutoring by turning to automatically generated dialogues as opposed to manually generated ones.
- Sciences – Thèses