Commonality-Based Information Retrieval with a Terminological Knowledge Representation System

The World-Wide Web as a vast information repository consists of a largely unorganized collection of documents, and mostinformation systems in the Web use standard database techniques for information retrieval. As argued in the knowledgediscovery, data mining, and Artificial Intelligence literature, this may not always be appropriate in view of the growing numberof private users accessing the internet. Researchers have reached a consensus that there is a need to structure informationand provide intelligent information retrieval techniques, especially adapted to the needs of the new user class.

Terminological knowledge representation systems based on description logics as descendants of the system Kl-Onehave proven to be an excellent means for structurally representing the knowledge of an application domain and reasoning aboutit. In this thesis we present a theoretical framework for a terminological knowledge representation system whose basicfeature is a facility for commonality-based information retrieval. With this technique, information retrieval is performed on thebasis of the commonalities of example information items specified by the user. The results in this thesis are thus a steptowards a unification of the query-by-example information retrieval paradigm and the logic-based knowledge representationmethodology.

Due to the formal syntax and well-defined semantics of description logics, it is possible to develop specialized and efficientreasoning services which are used to derive implicit knowledge from explicitly represented knowledge. The reasoning servicescan be used to support the design of the knowledge base as well as its application. In this thesis we present a number ofreasoning services and integrate them into a terminological knowledge representation system based on expressive descriptionlogics. We will particularly focus on the development of specialized non-standard reasoning services to formalizecommonality-based information retrieval.

Robustness issues play an important role in Artificial Intelligence research. In our application context robustness of a systemis strongly influenced by the amount and quality of information items returned by the system. It is a crucial objective to keepthe retrieval set processable, especially if strongly diversified or unexpectedly homogeneous collections of user-specifiedexamples are present. Therefore, a great portion of this work will be concerned with the problem of avoiding inappropriatelylarge or small retrieval sets. In addition, we will deal with the problem of incorporating so-called negative examples. Thespecification of negative examples enables the user to express examples of undesired information items.

The reasoning procedures developed in this thesis are not only of use for commonality-based information retrieval, but formany other description logic applications including innovative forms of example-based knowledge base design and learningmethods. Aside from theoretical usefulness, the presented retrieval framework can be applied to build information systems formany purposes including e-commerce applications as a field of growing economic relevance. However, the implementation ofthe theoretical results into a full-featured application is not part of this thesis.