Modelling Negotiation in Intelligent Teaching Dialogues1

In this chapter we aim to establish a set of requirements for the design and implementation of models for sustaining dialogue in Intelligent Tutoring Systems, which incorporate mechanisms for negotiating cooperative teaching and learning goals. Our discussion is centred around a first prototype computational model called "KANT"`(Kritical Argument Negotiated Tutoring system), which incorporates negotiation mechanisms within the framework of a more general model for dialogue. KANT bases its decisions on what to negotiate on a set of parametrised dialogue goals, in combination with a method for controlling dialogue focus based on spreading activation in working memory. Its decisions as to whether to cooperate or not in a negotiation are based on criteria of relevance with respect to a dialogue state consisting of speakers' justified beliefs in the domain and concerning the other speaker's beliefs, a model for focus of attention in dialogue, and a restricted memory of recent negotiations. After exploring the function of negotiation in tutorial dialogues, we discuss existing research and analyse negotiation extracts from a corpus of computer-mediated teaching dialogues. We attempt to identify what may be negotiated, the causes of negotiation and its success or failure, and the associated specific interaction structures. We then describe KANT, and critique it with respect to our previous analyses in order to define an agenda for future work.