A Framework for Representing Tutorial Discourse

We set forth general techniques for managing discourse in an intelligent tutor. These techniques are being implemented in a structure that dynamically reasons about the discourse, a student's response, and the tutor's move. The structure is flexible, domain-independent, and designed to be rebuilt-decision points and machine actions are modifiable through a visual editor. We discuss this formal reasoning structure and its application in an intelligent tutor. I. Tutorial Discourse We have built a process model for tutorial discourse that provides custom-tailored feedback to students in the form of examples, analogies, and simulations. The processing model views discourse as navigation through a set of possible discourse situations. Transition from one situation to another is dynamically generated so the system is capable of tracking and responding to contingencies in discourse alternatives. Fundamental to our perspective is the view that tutoring conversation is motivated by general rules (principles) of discourse and selection of intervention techniques based on error and misconception analysis. Effective tutoring requires sophisticated and dynamic reasoning about selection of tutoring strategy, the choice of a path through the curriculum, updating the student model, and assessment of student errors and misconceptions. Conversational actions produced by the tutor and responses from students will change the state of the discourse , and the structure must decide what the tutor should say and how it should interpret and act on subsequent student responses. Hence a desiderata on the design of any discourse manager is that it respond fluidly to the user and that it coordinate its utterances in a more flexible manner than has been required for question/answer or summarization systems. Networks and production rule formalisms have been used to identify admissible instructional actions in tutoring formalisms are often domain-dependent and restricted to a narrow set of didactic responses. We have extended such formalisms by modifying the basic ATN architecture, adding greater functionality to it, and by developing a domain-independent structure that can be used in a variety of tutors. We have also incorporated results from cognitive research about effective tutoring into the framework, including principles that drive production of good discourse. We are making the discourse framework modifiable through a visual editor that allows a knowledge engineer to access the machine's response decisions and actions. In the long term, we intend to make the steps of this reasoning process available to human teachers who can then modify the tutor for use in a classroom. …