基于异构图注意力网络的药物不良反应实体关系联合抽取研究

[Purpose/ Significance] Joint extraction of entities and relations is a crucial component in the adverse drug reactions monitoring and knowledge organization. To address the issues of error propagation, entity redundancy and interac⁃ tion deficiency in traditional pipeline extraction methods, and to improve the extraction effect of overlapping ternary groups of adverse drug reactions, the paper proposes a joint extraction model of adverse drug reactions entities and relations based on heterogeneous graph attention network MF-HGAT. [Method/ Process] Firstly, the paper conducted knowledge trans⁃ fered from external medical corpus resources through pre-training with BERT to achieve the fusion of multiple semantic fea⁃ tures. Secondly, the paper introduced relations information as prior knowledge for heterogeneous graph nodes to avoid ex⁃ tracting semantically irrelevant entities. Then, the paper enhanced the representations of characters and relations nodes by iteratively fusing messages with a hierarchical graph attention network through message passing. Finally, the paper extracted drug adverse reactions entities and relations after updating the node representations. [Result/ Conclusion] Experiments on self-constructed adverse drug reactions datasets reveal that the joint extraction F1 value of MF-HGAT, which incorporates relations information and external medical and health domain knowledge, reaches 92. 75%, which is an improvement of 5. 29% over the mainstream model CasRel. The results demonstrate that the MF-HGAT model further enriches entity-rela⁃ tions semantic information by fusing character and relations node semantics through heterogeneous graph attention network, which is of great significance to the knowledge discovery of adverse drug reactions.