NOISE REDUCTION MODEL OF RELATION EXTRACTION DATA BASED ON DISTANT SUPERVISION

Aimed at the problem of error labeling in distant supervision, a new relationship extraction model is proposed. The model was divided into two parts: label learner and relationship classifier. The tag learner corresponded the reinforcement learning action to the relationship tag, explored the real tag of the instance through the deep Q network, and the corrected tag and sentence form new data to reduce the impact of noise on the model. At the same time, K-choice strategy was proposed to alleviate the problem of reward sparsity and improve the performance of relationship extraction. In addition, in the training process, the accuracy of label prediction was improved by calculating the contribution value of words in relation classification and mining trigger words. Experiments show that the model can deal with noise well, and has a good effect on sentence level relationship classification.