Integrating ontological metadata: Algorithms that predict semantic compatibility

This dissertation presents an architecture, algorithms, and a methodology for studying the fundamental problem of communicating with terminology subject to local variation. I assume differentiated ontologies: that concepts are defined formally by their relations to other concepts, and that they inherit definitional structure from concepts in shared ontologies. I then develop measures of description compatibility. These algorithms compare concept definitions to predict semantic compatibility, the probability that an instance of a recommendation will satisfy a request.; The description-compatibility measures vary with respect to the complexity of their calculation, and the required knowledge of the semantics of roles in concept definitions. Some of the measures quantify the proportion of definitional structure that is shared by the request and recommendation. The “matching-based” measures identify and analyze correspondences among elements of the definitions, and are thus a form of analogical reasoning. Other measures use knowledge of the interaction of roles to more closely emulate the probabilistic definition of semantic compatibility.; I use simulations to evaluate the description-compatibility measures in detail. In these experiments, I generate ontologies that describe finite worlds of artificial objects. The meaning of a concept is its denotation (a set of objects), and the semantic compatibility of two concepts is a function of the intersection of their denotations.; Statistical analyses of data generated by the simulations show that the description compatibility measures predict semantic compatibility with accuracy consistent with their requirements for input and computation. The measures based on shared inherited structure achieve ranked (Spearman's) correlations between 0.3 and 0.4. The matching-based measures achieve ranked correlations between 0.5 and 0.6. The probabilistic measures achieve ranked correlations over 0.7. These and other results indicate that differentiated ontologies with description compatibility can be used effectively for purposes such as integrating heterogeneous data, and guiding agent search for services across communities that subscribe to differentiated ontologies.