Medical Relationship Classification Method Based on Dual Channel Attention

Electronic medical record mining based on relationship classification has become a hot topic in the field of healthcare. However, in existing models of relationship classification, most of them use single-layer attention, it results in relatively simple feature representation and it is easy to lose information during training. Therefore, this paper proposes a medical relationship classification method based on dual channel attention. Firstly, channel 1 combines BERT(Bidirectional Encoder Representation from Transformers), GRU(Gate Recurrent Unit) and Global Attention, while channel 2 combines Subject_object_mask_generation and So Attention. Specifically, we use Subject_object_mask_generation module to specify the corresponding positions of the subject and object within the text. And So Attention is used to focus on the attention between subject and object. Secondly, the outputs from the two channels are concatenated. Finally, we perform relationship classification on the concatenated results. We evaluated the method on the public dataset CMeIE(Chinese Medical Information Extraction), the experimental results showed that the improved model's accuracy, recall and F1 values are increased by 2.2%, 0.03% and 1.3% respectively, compared to the baseline. It indicates that our method has certain advantages compared to other relationship classification methods.