Practical reasoning for defeasable description logics.

Abstract

Description Logics (DLs) are a family of logic-based languages for formalising ontologies. They have useful computational properties allowing the development of automated reasoning engines to infer implicit knowledge from ontologies. However, classical DLs do not tolerate exceptions to speci ed knowledge. This led to the prominent research area of nonmonotonic or defeasible reasoning for DLs, where most techniques were adapted from seminal works for propositional and rst-order logic. Despite the topic's attention in the literature, there remains no consensus on what \sensible" defeasible reasoning means for DLs. Furthermore, there are solid foundations for several approaches and yet no serious implementations and practical tools. In this thesis we address the aforementioned issues in a broad sense. We identify the preferential approach, by Kraus, Lehmann and Magidor (KLM) in propositional logic, as a suitable abstract framework for de ning and studying the precepts of sensible defeasible reasoning. We give a generalisation of KLM's precepts, and their arguments motivating them, to the DL case. We also provide several preferential algorithms for defeasible entailment in DLs; evaluate these algorithms, and the main alternatives in the literature, against the agreed upon precepts; extensively test the performance of these algorithms; and ultimately consolidate our implementation in a software tool called Defeasible-Inference Platform (DIP). We found some useful entailment regimes within the preferential context that satisfy all the KLM properties, and some that have scalable performance in real world ontologies even without extensive optimisation.