Séminaire (organisé par l’équipe de recherche DI)


ATER, LIAS, Université de Poitiers

Analyzing and Evaluating Conflicts in Knowledge Bases

Lundi 23 mai 2016 à 14 h en salle GI042

Résumé :

Handling contradictory data is one of the most complex and important problems in reasoning under uncertainty. Inconsistencies arise easily in many applications, e. g., when several experts share their knowledge in order to solve a problem. If we are to develop computational systems that are more robust in the face of real-world situations, we need them to act intelligently when there is inconsistency. Measures of inconsistency are being developed recently as an interesting topic for this in order to allow for the degree of the problem to be quantified in a meaningful way.

In this talk, we present a logical framework to assess the severity of inconsistencies in knowledge bases of classical logic in a quantitative way. In the first part, we define a graph representation of knowledge bases, based on which we explore the logical properties that every desirable inconsistency measure should satisfy. Then, we show how the structure of the proposed graph representation can be used to evaluate, in a fine-grained way, the amount of conflict of a given knowledge base.

The second part deals with our general approach, both parameterized and parameter free, which is proposed for defining a family of measures in order to analyse inconsistency for knowledge bases. The parameterized approach allows to encompass several existing inconsistency metrics as specific cases, by properly setting its parameter. As for the parameter free approach, it is defined to avoid the difficulty in choosing a suitable parameter in practice but still keeps a desired ranking for knowledge bases by their inconsistency degrees.

Finally, SAT-based encodings are provided, which enable the use of efficient SAT solvers for the computation of the proposed measures. We implement these algorithms and test them on some real-world datasets. The experimental results for a variety of inputs show that the proposed framework gives a wide range of possibilities for evaluating large data.


FR SHIC 3272

Collegium UTC/CNRS