BIRL: Bidirectional-Interaction Reinforcement Learning Framework for Joint Relation and Entity Extraction

Joint relation and entity extraction is a crucial technology to construct a knowledge graph. However, most existing methods (i) can not fully capture the beneficial connections between relation extraction and entity extraction tasks, and (ii) can not combat the noisy data in the training dataset. To overcome these problems, this paper proposes a novel Bidirectional-Interaction Reinforcement Learning (BIRL) framework, to extract entities and relations from plain text. Especially, we apply a relation calibration RL policy to (i) measure relation consistency and enhance the bidirectional interaction between entity mentions and relation types; and (ii) guide a dynamic selection strategy to remove noise from training dataset. Moreover, we also introduce a data augmentation module for bridging the gap of data-efficiency and generalization. Empirical studies on two real-world datasets confirm the superiority of the proposed model.