Abstract
We are interested in providing semi-automatic support for the task of integrating large knowledge bases (KBs) through the use of structural information in description logics. For that reason we set up a formal framework for the integration of KBs which enables us to investigate the potential and limits of using structural information (in contrast to just thesaurus information involving identifier strings). Structurally similar concept descriptions (role chains) are related by integration mappings, which must satisfy a previously proposed notion of ”conflict-freeness”. It turns out that computing integration mappings is closely related to unification and matching of concept descriptions (and is intractable even for simple description logics). As a formal basis for the limitations of structural information we introduce the notion of indistinguishable concepts. We state the connection between indistinguishable concepts and integration mappings theoretically and also evaluate the problem empirically, by looking for indistinguishable concepts in a sizeable ontology (galen), and integration mappings fromgalen to itself. An interesting discovery is that roughly two thirds of the concepts are distinguishable by their structural properties, indicating that semantic information is indeed being encoded in the knowledge representation, not just in the name.