Improving Sentence-Level Relation Classification via Machine Reading Comprehension and Reinforcement Learning

Distant supervision (DS) has been proposed to automatically annotate data and achieved significant success in relation classification. However, despite its efficiency, distant supervision often suffers from the noisy labeling problem. To solve the problem, existing methods can be divided into two major approaches: (1) Some works adopt multi-instance learning (MIL) for relation classification to reduce the impact of noisy data. However, they do not perform well at the sentence level. (2) Other works focus on finding the noisy instances directly. They mainly use reinforcement learning to filter out the noisy instances. The key component is the instance selector, which is used to select the correct instances from the noisy data. However, current instance selectors usually use simple neural network models and initialize the models with random parameters, which leads to limited improvement and slower convergence. In this paper, we propose a novel instance selector to directly select the high-quality instances from DS-generated data as the refined training data to improve the performance of sentence-level relation classification. Specifically, the instance selector consists of a machine reading comprehension (MRC) estimator and an instance sampler. The MRC estimator is used to evaluate the quality of the instances, and the instance sampler is used to select the high-quality instances. Moreover, due to the lack of explicit knowledge about which instances are mislabeled, we use reinforcement learning to train the MRC estimator. Experiments show that our method achieves state-of-the-art performance on two human-annotated NYT10 datasets. The source code of this paper can be found in https://github.com/xubodhu/MRCRL.