Acquiring Qualitative and Quantitative Knowledge from Verbal Statements and Dialogues in Probabilistic Domains

We describe an approach to acquire qualitative and quantitative knowledge from verbal statements and dialogues in complex, probabilistic domains. This work is part of the development of an intelligent environment, MEDICUS (M odelling, e xplanation, and di agnostic support for c omplex, u ncertain s ubject matters), that supports modelling and diagnostic reasoning in the domains of environmental medicine and human genetics. The system is designed for professional as well as for further education purposes in these two medical domains. Support for other domains of rapidly changing and uncertain knowledge will be possible as well. In MEDICUS, uncertainty is handled by the Bayesian network approach. Thus modelling consists of creating a Bayesian network for the problem at hand. Since MEDICUS is designed for users interested in the domain but not necessarily in mathematical issues, it is possible to state propositions verbally and let the system generate a Bayesian network proposal. This differs from existing reasoning systems based on Bayesian networks, i.e. in medical domains, which contain a built-in knowledge base that may be used but not created or modified by the user. Diagnostic reasoning and deciding consists of using the network for stating and testing diagnostic hypotheses, and asking for recommendations. In this paper we will focus on the modelling component. In order to design a domain model represented as a Bayesian network, it is necessary to specify the qualitative and quantitative information necessary. This is a problem for probability-based as well as for other uncertainty formalisms. We will describe our approaches how to acquire this knowledge from dialogues. • With respect to qualitative information, it is necessary to check whether the dependence and independence relations implied by a Bayesian network correspond to the intentions of the modeller. In MEDICUS, these relations are obtained from diagnostic assertions. • With respect to quantitative information, apriori and conditional distributions have to be obtained in order to be able to use the network for diagnostic reasoning. But even domain experts are usually hesitant to specify numerical relationships. So we argue that in an easily usable system the modeller should be able to state propositions verbally. We are currently developing an approach to assign probabilities to these "fuzzy" relations and concepts. We will show how to extend these approaches to natural dialogue situations between two or more participants. The purpose of this is to be able to acquire the knowledge from more natural situations, …