Distantly Supervised Relation Extraction Based On Collaborative Encoders with Hierarchy Relations

Since extracting structured information with automatic annotation, distantly supervised relation extraction (DSRE) reduces the cost of labor greatly and has become a remarkable approach to relation extraction. However, DSRE also produces a lot of mislabeled data in automatic annotation. To address this issue, this paper proposes a novel DSRE model, based on collaborative encoders with hierarchy relation of relations, namely CEH-RORs. In particular, CEH-RORs proposes collaborative encoders, which not only dynamically control the amount of information but also select useful information as effectively as possible. Moreover, this paper constructs the hierarchical graph based on the graph attention network (GAT) to aggregate the node information, in which each relation in the hierarchy of relations forms a node in the input graph. In addition, this paper further improves the performance by using pre-trained relational embeddings. Extensive experiments demonstrate that our approach improved AUC by 4.69% and average P@N to 1.78% compared to its sub-optimal value of existing remarkable models.