基于IDCNN+CRF和注意力机制的电子病历命名实体识别方法及模型稳定性研究

Research on named entity recognition method and model stability of electronic medical record based on IDCNN + CRF and attention mechanism

目的:基于膨胀卷积+条件随机场(IDCNN+CRF)和注意力机制进行电子病历的实体识别,观察识别效果及模型的稳定性。方法:电子病历实体识别的实体抽取环节采用IDCNN+CRF架构,实体对齐环节采用Seq2Seq+注意力机制的翻译框架。与其他算法模型比较,观察所构建模型对电子病历实体的识别效果;逐步增加电子病历数量,分析对模型性能的提升效果。结果:IDCNN+CRF+注意力模型对于电子病历实体识别的准确率为97.17%,精确率为95.75%,召回率为95.06%,F_(β)值为95.51%,优于Bi-LSTM+CRF、CRF、HMM等其他模型;在电子病历增加15篇后F_(β)-score值趋于平稳。结论:基于IDCNN+CRF和注意力机制构建的模型具有良好的电子病历实体关系识别效果,且模型稳定性优异,为后续电子病历命名实体结构化提供了一种稳定可行的方案。

Objective To conduct entity recognition of electronic medical record(EMR) based on IDCNN+CRF and attention mechanism,and observe the recognition effect and the stability of the model.Methods IDCNN+CRF(iterated dilated convolutional neural network + conditional random field) architecture was used in entity extraction of EMR entity recognition,and Seq2Seq+attention mechanism translation framework was used in entity alignment.Compared with other algorithm models,the recognition effect of the model on EMR entities was observed.By gradually increasing the number of electronic medical records,the improvement effect on key indicators was analyzed.Results The model built based on IDCNN+CRF and attention mechanism,its accuracy was 97.17%,precision rate was 95.75%,recall rate was 95.06% and F_(β)-score value was 95.51% for EMR entity recognition,which was better than Bi-LSTM+CRF,CRF,HMM and other models.The F_(β)-score values leveled off after increasing to 15 records on the basis of the model.Conclusion The model based on IDCNN+CRF and attention mechanism has a good effect on entity relationship of EMR,and the model has excellent stability,which provides a stable and feasible scheme for the subsequent named entity structure of EMR.