Inductive knowledge acquisition for a UNIX coach

Abstract Research in the field of Human-Computer Interaction has brought forth several methods to formally model procedural knowledge of computer users. Such models have been mainly used for analytic purposes. It is shown that similar models can also serve as knowledge bases for intelligent user support systems, particularly plan-recognizing help systems. This is demonstrated by a prototypical application (FINIX) which provides intelligent help for UNIX file handling operations. As for other knowledge-based systems, knowledge acquisition is crucial in order to make this approach practically useful. This paper gives a detailed description of two inductive, similarity-based methods for acquiring task knowledge from the dialogue history. The first approach is semi-automatic and relies on interactions with a human referee, whereas the second is completely automated based on certain heuristics. The methods have been implemented and successfully tested in the FINIX environment. They are general in that they may be used to acquire procedural task knowledge in different domains. Both methods are analyzed according to the underlying machine learning principles.