Chinese medical relation extraction based on multi-hop self-attention mechanism

The medical literature is the most important way to demonstrate academic achievements and academic exchanges. Massive medical literature has become a huge treasure trove of knowledge. It is necessary to automatically extract implicit medical knowledge from the medical literature. Medical relation extraction aims to automatically extract medical relations from the medical text for various medical researches. However, there are a few kinds of research in Chinese medical literature. Currently, the popular methods are based on neural networks, which focus on semantic information on one aspect of the sentence. However, complex semantic information in the sentence determines the relation between entities, the semantic information cannot be represented by one sentence vector. In this paper, we propose an attention-based model to extract the multi-aspect semantic information for the Chinese medical relation extraction by multi-hop attention mechanism. The model could generate multiple weight vectors for the sentence through each attention step, therefore, we can generate the different semantic representation of a sentence, respectively. Our model is evaluated by using Chinese medical literature from China National Knowledge Infrastructure (CNKI). It achieves an F1 score of 93.19% for therapeutic relation tasks and 73.47% for causal relation tasks.