Distant Supervision for Relation Extraction via Group Selection

Distant supervision DS aligns relations between name entities from a knowledge base KB with free text and automatically annotates the training corpus with relation mentions. One big challenge of DS is that the heuristically generated relation labels usually tend to be noisy, when a pair of entity has multiple and/or incomplete relations in a KB. This paper proposes two ranking-based methods to reduce noise and select effective training data for multi-instance multi-label learning MIML, one of the most popular learning paradigms for distantly supervised relation extraction. Through the proposed methods, training groups that are of low quality are excluded from the training data according to different ranking strategies. Experimental evaluation on the KBP dataset using state-of-the-art MIML algorithms in this community demonstrated that the proposed methods improved the performance significantly.