Multi-head Attention and Graph Convolutional Networks with Regularized Dropout for Biomedical Relation Extraction

Automatic extraction of biomedical relation from text becomes critical because manual relation extraction requires significant time and resources. The extracted medical relations can be used in clinical diagnosis, medical knowledge discovery, and so on. The benefits for pharmaceutical companies, health care providers, and public health are enormous. Previous studies have shown that both semantic information and dependent information in the corpus are helpful to relation extraction. In this paper, we propose a novel neural network, named RDMAGCN, for biomedical relation extraction. We use Multi-head Attention model to extract semantic features, syntactic dependency tree, and Graph Convolution Network to extract structural features from the text, and finally R-Drop regularization method to enhance network performance. Extensive results on a medical corpus extracted from PubMed show that our model achieves better performance than existing methods.