Generating triples based on dependency parsing for contradiction detection

Contradiction detection is a task to detect the contradictory relation between two texts. In the social media, the phenomenon of contradictory descriptions of the same event is common and harmful. It is urgent to detect contradictory texts. Previous methods on detecting contradiction are mostly deriving features from shallow semantic representations like predicate-argument structures. They meet a problem of the low coverage of contradiction. We propose a joint method to extract more contradiction pairs. We utilize dependency parsing tree to generate tripes (dp-triple) which represent semantic information of the text. The dp-triple extraction method extract more contradiction pairs than present shallow semantic extraction methods like open IE or SRL. Due to the coverage limitation of triples, we also derive features from the context of the matching words between texts as backup. We demonstrate the joint method is effective in detecting contradiction. In predicting stage, we use a unsupervised method to detect contradiction relation and achieve a better performance than the state of the art method.