2 Content-based Information Retrieval by Computation of Least Common Subsumers in a Probabilistic Description Logic 12.1 Introduction

Due to the constantly growing number of information sources, intelligent information retrieval becomes a more and more important task. We model information sources by description logic (DL) terminologies. The commonalities of user-speciied examples can be computed by the least common subsumer (LCS) operator. However, in some cases this operator delivers too general results. In this article we solve this problem by presenting a probabilistic extension of the LCS operator for a probabilistic description logic. By computing gradual commonalities between description logic concepts, this operator serves as a crucial means for content-based information retrieval for all kinds of information sources. We also describe an extension of our operator to consider unwanted information. The probabilistic LCS can be applied for information retrieval in a scenario of multiple information sources. The number of structured but heterogeneous information sources that are available online is growing rapidly. In particular, many sources in the WorldWide Web ooer information about all kinds of themes. Often the user must manually combine information items from multiple sources. If information is distributed in diierent semi-structured formats (see the XML discussion), automatic integration techniques are required to provide adequate information systems. Basically, the same situation occurs in standard database contexts, and thus, many of the well-known integration techniques can be reused in the Web context (see, e.g., CL93]). As a remedy to the integration and combination problems, the "information agent" abstraction has been proposed (e.g., SIMS AKS96]). Information agents are understood as systems that provide a uniform query interface to multiple information sources. In the Web context, most users of information systems are only casual users. Hence, they are often overtaxed when asked to (formally) describe the exact kind of information they desire. In many applications they can, however, supply examples concerning the information of interest. In contrast to approaches where the user has to learn query languages (or agent communication languages), in this paper we focus on providing the theoretical background for information retrieval on the basis of user-speciied examples which express his information demands. An information system can automatically determine a description of the user's information demands by evaluating the