基于CNN-CRF的中文电子病历命名实体识别研究

智慧医疗技术的发展让我们不满足仅使用传统方法做医学研究。针对中文电子病历实体识别问题，设计了一种基于卷积神经网络结合条件随机场(convolutional neural network-conditional random field,CNN-CRF)的实体识别算法框架。为得到高质量的词向量，将标注实体加入词典进行分词，并将已标注和未标注文本作为语料，用word2vec工具对已分词文本进行无监督学习；为避免扩张卷积层数增加导致过拟合，采用迭代扩张卷积处理输入向量，并使用dropout随机丢弃一些连接；运用条件随机场对网络的分类结果进行修正。把该方法在中文电子病历上进行对比试验，从病历中提取出身体部位，疾病，症状，检查及治疗5类实体。实验结果表明，该方法能有效地辨别病历中的实体，其识别的准确率、召回率和f1值分别为90.01%，90.62%，90.31%，准确率和速率比传统方法都有一定提高。

Research on named entity recognition of chinese electronic medical record based on CNN-CRF

The development of intelligent medical technology makes us not satisfied with only using traditional methods for medical research. For entity recognition of Chinese electronic medical record, an algorithm framework based on convolutional neural network and conditional random field (CNN-CRF) is designed in this paper. In order to obtain high-quality word vectors, labeled entities are added to the dictionary to segment words, and labeled and unlabeled texts are used as corpus. Unsupervised learning of segmented texts is carried out with word2vec tool. The increase of the dilated layers will cause over-fitting. The iterative dilated convolution is applied to train input vector, and some connections are randomly discarded by the dropout. Finally, the conditional random field revises the classification result. Based on the Chinese electronic medical records, the proposed method can extract body parts, diseases, symptoms, examination and treatment from the records. The experimental results show that our approach can effectively identify the entities, and the accuracy rate, the recall rate and the F1 value are 90.01%, 90.62%, and 90.31%. Compared with the traditional methods, the accuracy and speed of this approach are improved.