Trimming a consistent OWL knowledge base, relying on linguistic evidence

Abstract

Intuitively absurd but logically consistent sets of statements are common in publicly available OWL datasets. This article proposes an original and fully automated method to point at erroneous axioms in a consistent OWL knowledge base, by weakening it in order to improve its compliance with linguistic evidence gathered from natural language texts. A score for evaluating the compliance of subbases of the input knowledge base is proposed, as well as a trimming algorithm to discard potentially erroneous axioms. The whole approach is evaluated on two real datasets, with automatically retrieved web pages as a linguistic input.