A Neural Learning Framework for Advisory Dialogue Systems

A domain independent neural learning framework for advisory dialogue systems (ADS) is suggested. A connectionist view of user and task modeling is introduced that can be implemented in a neural knowledge network. It implicitly interprets man-computer interaction and causes adaptive task support. Adaptive inference is drawn by modifying the causal connections during interaction. The interpretation of the network gives insights into the user’s knowledge and preferences. Reasons for misconceptions can be estimated and interpreted by users, designers and rules for network modification.Neural ADS learn empirically in real-time to raise future system performance but can also be programmed by experts. Additionally, the network can be used for predicting the behaviour of the whole system or its parts.Advantages are constant retrieval time for associated information, extendability, and variability. Implementing the framework does not require special hardware or neural simulators. To demonstrate the applicability, two prototypical spreadsheet applications are introduced.