Reducing wrong labels in distant supervision for relation extraction

Abstract

Distant supervision has been widely used for relation extraction recently. In the distant supervision, many labels may to wrongly marked, which exerts a bad impact on relation extraction. A method to reduce wrong labels was introduced by using the semantic Jaccard to measure semantic similarity between the relation phrases and the dependency terms. The training data after reducing wrong labels was used to train the relation extractors. The experimental results show that the proposed method can effectively reduce wrong labels and improve the relation extraction performance compared with the state-of-art methods.