图注意力网络与句法融合的医疗实体识别

Incorporating Graph Attention Network and Syntax for Medical Entity Identification

电子病历数据中包含大量的医疗实体词,对这些实体词的自动识别有益于提升计算机对电子病历数据的理解。待识别的医疗实体词通常由医疗专业术语和非规范医疗词汇构成,大量生僻词汇、长难词汇和病历行文中的省略现象给医疗实体识别任务带来了挑战。针对以上问题,该文提出一种图注意力网络与句法融合的医疗实体识别方法,该方法结合字词共现关系和句法依存关系,基于电子病历数据构建了交互式字词关系图和依存关系图,并利用图注意力网络完成多种图信息的融合。实验结果表明,在电子病历的命名实体识别中,该文方法得到88.91%的F 1值,较基线模型提高1.04%,验证了该方法的有效性。

The electronic medical record data contains a large number of medical entities,and the automatic recognition of these entities is beneficial for improving the understanding of electronic medical records.The electronic medical record data contains professional medical terms and a large number of non-standard medical vocabulary.Rare words,long difficult words and omission in medical records bring challenges to medical entity recognition.To solve this problem,this paper proposes a medical entity recognition method based on graph attention network and syntax fusion.This method combines the co-occurrence relationship between words and the rules of syntactic dependency,and it realizes the fusion of various graph information by graph-attention network based on the construction of interactive character-word relationship graph and dependency relationship graph of electronic medical record data.The experiment result reveals that the proposed method achieves 88.91%F 1 value,which is 1.04%higher than the baseline model.