Inference as a data management problem

Yu Liu, Peter McBrien, Peter Pietzuch, Imperial College London
Inference over OWL ontologies with large A-Boxes has been researched as a data management problem in recent years. This work adopts the strategy of applying a tableaux-based reasoner for complete T-Box classification, and using a rule-based mechanism for scalable A-Box reasoning. Specifically, we establish for the classified T-Box an inference framework, which can be used to compute and materialise inference results. The inference we focus on is type inference in A-Box reasoning, which we
more » ... as the process of deriving for each A-Box instance its memberships of OWL classes and properties. As our approach materialises the inference results, it in general provides faster query processing than non-materialising techniques, at the expense of larger space requirement and slower update speed. When the A-Box size is suitable for an RDBMS, we compile the inference framework to triggers, which incrementally update the inference materialisation from both data inserts and data deletes, without needing to re-compute the whole inference. More importantly, triggers make inference available as atomic consequences of inserts or deletes, which preserves the ACID properties of transactions, and such inference is known as transactional reasoning. When the A-Box size is beyond the capability of an RDBMS, we then compile the inference framework to Spark programmes, which provide scalable inference materialisation in a Big Data system, and our evaluation considers up to reasoning 270 million A-Box facts. Evaluating our work, and comparing with two state-of-the-art reasoners, we empirically verify that our approach is able to perform scalable inference materialisation, and to provide faster query processing with comparable completeness of reasoning.
doi:10.25560/44960 fatcat:men6j47wgnbnfhlltqagh5ircm