The 1988 AAAI Workshop on Explanation

Expert system explanation is the study of how to give an expert system the ability to provide an explanation of its actions and conclusions to a variety of users (including the domain expert, knowledge engineer, and end user). The 1988 AAAI Workshop on Explanation brought together many of the world’s experts on expert system explanation in an attempt to highlight key research areas and questions that should be the focus of subsequent work. The one-day workshop was organized into five sessions of short presentations, each followed by panel-led open discussion among the 35 workshop participants. A proceedings of the workshop was compiled and is available through AAAI. The first session, Text Planning I, focused on some of the issues involved in treating the process of explanation as a complex problemsolving task requiring knowledge beyond that used to solve the expert system’s original problem. This session, led by Cecile Paris, Robert Schulman, Mike Wick, and Dan Suthers, brought up issues related to the generation of explanations, including the nature of the coupling between the processes and the knowledge involved in generating explanations and that involved in problem solving. The second session, Explanation and Knowledge Acquisition, raised several issues regarding the connection between expert system explanation and the problem of knowledge acquisition. Led by Bill Mark, Alice Kidd, Sue Abu-Hakima, and Bruce Porter, the discussion focused on identifying how explanation can be used for knowledge acquisition as well as how knowledge acquisition can be used to acquire knowledge for explanation. The third session, User Modeling, attempted to outline some of the major issues in employing user models to tailor expert system explanations to particular users or user types. Kathy McKeown, Robin Cohen, Ivan Rankin, and Robert Kass led the open discussion. The workshop participants outlined four major components central to a user model. In addition, a possible method for the automatic acquisition of some of these components was discussed. The fourth session, Question Types, highlighted work designed to find useful categorizations of the queries that an explanation system can answer. The discussion, led by G. Nigel Gilbert, Mike Tanner, and Dave Schaffer, focused on two largely distinct approaches to finding such categorizations. First, an argument was presented that called for categorizing explanation responses rather than explanation queries. The second approach, focusing on query categorization, argued for the use of the domain model and explicit considerations of potential explanation queries during the design phase of the expert system. This session also raised the issue of canned text as an explanation paradigm and discussed its usefulness. The fifth and final session, Text Planning II, centered on the issue of how to construct the information that is presented to the user. Johanna Moore, Dan Rochowiak, and Yee-Han Cheong led a discussion that focused on three main issues related to this problem: the ability to respond to follow-up questions by the user, the identification and use of implicit context in user queries, and the use of rhetorical devices to enhance explanations.