Distant Supervision Relation Extraction via Reinforcement Learning with Potential Energy Function

Distant supervision has become an essential method for relation extraction. Although distant supervision is very effective, there is a large amount of noise in the dataset produced by distant supervision. To filter out the noise in the dataset, most recent models try to handle it by reinforcement learning, which require waiting for all sentences in the bags to be selected when calculating the reward, resulting in large time consumption in model training. To solve the problem, this paper proposes a method called distant supervision relation extraction via Reinforcement Learning with Potential Energy Function (PEF-RL). The model calculates the semantic similarity of sentences and relations through the relational alias table, and we pass the semantic similarity of sentences and relations as an additional reward to the potential energy function to guide the decision making process of the instance selector during reinforcement learning, which improves the learning efficiency of the reinforcement learning relation extraction model. Experiments show that our model has significant improvements in several dimensions compared to existing models.