Acquiring Knowledge from Linguistic Models in Complex, Probabilistic Domains

This paper describes an approach to acquire qualitative and quantitative knowledge from verbally stated models in complex, probabilistic domains. This work is part of the development of an intelligent environment, MEDICUS2, that supports modelling and diagnostic reasoning in the domains of environmental medicine and human genetics. These domains are two yet new subdomains of medicine receiving increasing research efforts, but still consisting of largely fragile and uncertain knowledge. In MEDICUS, uncertainty is handled by the Bayesian network approach. Thus the modelling task for the user consists of creating a Bayesian network for the problem at hand. But since we want mathematically untrained persons to work with MEDICUS, the user may alternatively 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. The diagnostic reasoning task for the learner consists of using the network for stating diagnostic goals, and for proposing diagnostic hypotheses and examinations. 1 We thank Jörg Folckers and Karsten Rommerskirchen for doing large parts of the implementation and for assisting in the mathematical work. 2 Modelling, explanation, and di agnostic support for complex, uncertain subject matters In this paper, we first give an overview of the aims and the actual implementation state of MEDICUS. Then we will focus on the modelling component and the central concern of this paper: One of the most difficult problems in the design of a domain model represented as a Bayesian network is to acquire the necessary qualitative and quantitative information from the modeller (of course this is a problem for other uncertainty formalisms too.): • With respect to qualitative information, the dependence and independence relations implied by a Bayesian network have to be validated empirically. This can be achieved by deriving assertions from these relations and comparing them to the modeller's assertions. For example, the dependence and independence relations can be compared to descriptions of diagnostic procedures: Given a diagnostic hypothesis and some case information, what diagnostic information is considered next? What diagnostic information is not considered? • 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. The possibility to work with intervals, as for example offered by Dempster-Shafer Theory, does not solve this problem. Rather, the modeller should be able to state propositions verbally. The system should be able to assign probabilities to these "fuzzy" relations and concepts. Existing approaches have been concerned with the empirical acquisition of the semantics of fuzzy terms concerning a single variable or proposition (for example: "It is possible that it will rain tomorrow."), but not with fuzzy relations. We attempt to close this gap by extending existing approaches to relations.