一种基于卷积神经网络和双向长短记忆神经网络的在线医疗社区问答文本实体识别方法

An online medical community Q A text entity recognition method based on CNN and BiLSTM

摘要:

[目的/意义]针对在线医疗社区问答文本复杂程度高、结构化程度低的特点，结合卷积神经网络（CNN）和双向长短记忆神经网络（BiLSTM）两种深度学习模型以及条件随机场（CRF）模型，提出一套适用于在线医疗问答文本的实体识别方法并进行验证。 [过程/方法] 将问答文本进行清洗和BIO标注后，分别用CNN和BiLSTM进行字级别的特征抽取，将两种模型抽取到的特征进行融合，后放入CRF中训练出实体预测模型，再将问答文本放入训练好的模型中得到最终的实体识别结果。[结果/结论]在所选取的乳腺癌医疗社区问答文本数据集上，所提出的方法结果优于其他模型，且识别准确率达到92.3%，召回率达到89.3%，F值达到90.8%。

Abstract:

[Purpose/Significance] In response to the characteristics of high complexity and low structure in the online medical community Q&A text, a entity recognition methods combined with two deep learning models Convolutional Neural Network (CNN) ，Bidirectional Long Short-Term Memory (BiLSTM) and Conditional Random Field (CRF) was proposed and verified, to promote the development of medical entity identification research for the online medical community. [Method/Process] After the Q&A texts are cleaned and BIO labeled, feature extraction is respectively performed in word-level by CNN and BiLSTM, then fuse the features and put the results into the CRF to train the entity prediction model, finally put Q&A text into the model to get the recognized result. [Result/Conclusion]For Q&A text about breast cancer, the proposed method is better than other models, and the recognition accuracy rate reaches 92.3%, the recall rate reaches 89.3%, and the F value reaches 90.8%.