Knowledge extraction from large ontologies. (Extraction de connaissances à partir d'ontologies de grandes tailles)

Hui Yang

Abstract: Because widely used real-world ontologies are often complex and large, one crucial challenge has emerged: designing tools for users to focus on sub-ontologies corresponding to their specific interests. To this end, this work investigates three different approaches for extracting knowledge from large ontologies: (1) Justification, a minimal sub-ontology of the original ontology that derives a specific conclusion; (2) Deductive module, a sub-ontology that preserves all entailments wrt a given vocabulary capturing the user interest; and (3) General module, a new ontology not necessarily a sub-ontology, that ensures to perform the same set of entailments as the original one over a given vocabulary. For computing justifications and deductive modules, we propose SAT-based methods that are conducted in two steps: (i) encoding the derivation of justifications (resp. deductive modules) as Horn-clauses; (ii) computing justifications (resp. deductive modules) by resolution over these Horn-clauses. For encoding the derivation of justifications, we construct a graph representation of ontologies and propose a new set of inference rules, which are more compact than existing ones. For encoding the derivation of deductive modules, we introduced a new notion called the forest, which relies on a graph representation, capturing all the logical entailments over a given vocabulary. For computing general modules, we proposed a new resolution-based method inspired by the existing approach for computing uniform interpolants. This method is, in general, more efficient and generates modules of better quality. Finally, each proposed method has been evaluated by implementing a prototype used to test large real-world ontologies and the experimental results have been compared to those obtained with state-of-the-art methods, showing the advantages of our method in terms of efficiency and quality.