Interactive ontology revision

摘要

When ontological knowledge is acquired automatically, quality control is essential. Which part of the automatically acquired knowledge is appropriate for an application often depends on the context in which the knowledge base or ontology is used. In order to determine relevant and irrelevant or even wrong knowledge, we support the tightest possible quality assurance approach - an exhaustive manual inspection of the acquired data. By using automated reasoning, this process can be partially automatized: after each expert decision, axioms that are entailed by the already confirmed statements are automatically approved, whereas axioms that would lead to an inconsistency are declined. Starting from this consideration, this paper provides theoretical foundations, heuristics, optimization strategies and comprehensive experimental results for our approach to efficient reasoning-supported interactive ontology revision. We introduce and elaborate on the notions of revision states and revision closure as formal foundations of our method. Additionally, we propose a notion of axiom impact which is used to determine a beneficial order of axiom evaluation in order to further increase the effectiveness of ontology revision. The initial notion of impact is then further refined to take different validity ratios - the proportion of valid statements within a dataset - into account. Since the validity ratio is generally not known a priori - we show how one can work with an estimate that is continuously improved over the course of the inspection process. Finally, we develop the notion of decision spaces, which are structures for calculating and updating the revision closure and axiom impact. We optimize the computation performance further by employing partitioning techniques and provide an implementation supporting these optimizations as well as featuring a user front-end. Our evaluation shows that our ranking functions almost achieve the maximum possible automatization and that the computation time needed for each reasoning-based, automatic decision takes less than one second on average for our test dataset of over 25000 statements.