Distant supervision for fine-grained biomedical relation extraction from Chinese EMRs

Automatically extract relations between medical entity pairs is fundamental in biomedical research. Since the annotated dataset is very expensive, distant supervision provides an efficient solution to reduce the cost of annotation by utilizing rough corpus labeled with semantic knowledge base. However, two same entities mentioned in different sentences may express different relations, it is difficult for the traditional distant supervision methods to distinguish these different relations. In this paper, we propose a new model for biomedical relation extraction in Chinese EMRs. First, the distant supervision is used for coarse-grained relation labeling. Then, the fine-grained relations are annotated initially by measuring the distance between the contextual information of the relation instance to the semantic profile of each candidate fine-grained relation category. Finally, the high confidence fine-grained relation instances are selected as initial training set for PCNN model, in addition, a bootstrap learning is introduced in the training process to enhance the performance of fine-grained relation extraction. Experiments conducted on a real-word dataset and the results show that our method outperforms all baseline systems.