Practical Ontologies: Requirements, Design and Applications to Semantic Integration and Knowledge Discovery

Due to their role in describing the semantics of information, knowledge representations, from formal ontologies to informal representations such as folksonomies, are becoming increasingly important in facilitating the exchange of information, as well as the semantic integration of information systems and knowledge discovery in a large number of areas, from e-commerce to bioinformatics. In this thesis we present several studies related to the development and application of knowledge representations, in the form of practical ontologies. In Chapter 2, we examine representational challenges and requirements for describing practical domain knowledge. We present representational requirements we collected by surveying current and potential ontology users, discuss current approaches to codifying knowledge and give formal representation solutions to some of the issues we identified. In Chapter 3 we introduce a practical ontology for representing data exchanges, discuss its relationship with existing standards and its role in facilitating the interoperability between information producing and consuming services. We also consider the problem of assessing similarity between ontological concepts and propose three novel measures of similarity, detailed in Chapter 4. Two of our proposals estimate semantic similarity between concepts in the same ontology, while the third measures similarity between concepts belonging to different ontologies. In Chapters 5 and 6 we discuss applications of our similarity measures. Specifically, in Chapter 5 we present a method for discovering mappings between different bio-ontologies and discuss the results of a quality assessment performed by two expert biologists on the top mappings identified by our method. In Chapter 6 we detail an integrative method for discovering gene associations: the Gene Association Predictor (GAP) and present the results of assessing the predictions made by GAP using a multi-pronged approach: (1) overlap with datasets of known, experimentally derived, gene associations, (2) comparison with other prediction services, and (3) manual curation of scientific literature.Ph