Negotiated Tutoring: An Approach to Interaction in Intelligent Tutoring Systems

This thesis describes a general approach to tutorial interaction in Intelligent Tutoring Systems, called "Negotiated Tutoring". Some aspects of the approach have been implemented as a computer program in the 'KANT' (Kritical Argument Negotiated Tutoring) system. Negotiated Tutoring synthesises some recent trends in Intelligent Tutoring Systems research, including interaction symmetry, use of explicit negotiation in dialogue, multiple interaction styles, and an emphasis on cognitive and metacognitive skill acquisition in domains characterised by justified belief. This combination of features has not been previously incorporated into models for intelligent tutoring dialogues. Our approach depends on modelling the high-level decision-making processes and memory representations used by a participant in dialogue. Dialogue generation is controlled by reasoning mechanisms which operate on a 'dialogue state', consisting of conversants' beliefs, a set of possible dialogue moves, and a restricted representation of the recent utterances generated by both conversants. The representation for conversants' beliefs is based on Anderson's (1983) model for semantic memory, and includes a model for dialogue focus based on spreading activation. Decisions in dialogue are based on preconditions with respect to the dialogue state, higher level educational preferences which choose between relevant alternative dialogue moves, and negotiation mechanisms designed to ensure cooperativity. The domain model for KANT was based on a cognitive model for perception of musical structures in tonal melodies, which extends the theory of Lerdahl and Jackendoff (1983). Our model ('GRAF' - GRouping Analysis with Frames) addresses a number of problems with Lerdahl and Jackendoff's theory, notably in describing how a number of unconscious processes in music cognition interact, including elements of top-down and bottom-up processing. GRAF includes a parser for musical chord functions, a mechanism for performing musical reductions, low-level feature detectors and a frame-system (Minsky 1977) for musical phrase structures