基于循环胶囊网络的临床语义关系识别研究

Recurrent Capsule Network for Clinical Relation Extraction

得益于医疗信息化的不断推进,医院已积累了大量的电子病历记录。然而,这些病历记录大多以自然语言的形式存在,无法为计算机所"理解",也就无法对其做进一步的处理与挖掘。由此,对病历文本进行结构化研究,识别出病历实体间的语义关系,便显得尤为重要。本文针对临床语义关系识别任务,提出循环胶囊网络模型,使用分段循环神经网络来捕捉两实体及其上下文信息,并使用胶囊网络来进行最终的关系分类。实验表明,本文提出的方法较现有监督学习方法取得了更好的识别效果(F1-score为96.51%),证明了本文方法的优越性。

A large number of electronic health records(EHRs)have been accumulated since the wide adoption of medical information systems in China.However,most of these records are written in natural language,which cannot be processed by computers directly.Thus,it is important to transform unstructured EHRs into structured ones.In this paper,a recurrent capsule network is proposed for clinical relation extraction in EHRs,where entity pairs and their contexts are captured by piece-wise recurrent neural network layers,and capsule layers are finally employed for relation classification.Experimental results show that this model performs better than the existing supervised methods,achieving a F1-score of 96.51%.