Abstract

Identifying incorrect content (i.e., semantic error) in text is a difficult task because of the ambiguous nature of written natural language and the many factors that can make a statement semantically erroneous. Current methods identify semantic errors in a sentence by determining whether it contradicts the domain to which the sentence belongs. However, because these methods are constructed on expected logic contradictions, they cannot handle new or unexpected semantic errors. In this paper, we propose a new method for detecting semantic errors that is based on logic reasoning. Our proposed method converts text into logic clauses, which are later analyzed against a domain ontology by an automatic reasoner to determine its consistency. This approach can provide a complete analysis of the text, since it can analyze a single sentence or sets of multiple sentences. When there are multiple sentences to analyze, in order to avoid the high complexity of reasoning over a large set of logic clauses, we propose rules that reduce the set of sentences to analyze, based on the logic relationships between sentences. In our evaluation, we have found that our proposed method can identify a significant percentage of semantic errors and, in the case of multiple sentences, it does so without significant computational cost. We have also found that both the quality of the information extraction output and modeling elements of the ontology (i.e., property domain and range) affect the capability of detecting errors.