结合神经文本生成的FLAT模型的中文电子病历命名实体识别

Named Entity recognition of Chinese electronic medical records using FLAT combined with neural network text generation

随着医疗信息化的发展,电子病历命名实体识别受到了广泛关注。电子病历中包含大量的专业词汇,而专业词汇的切分错误会使命名实体识别效果不佳。FLAT模型在引入词边界信息时能有效避免分词错误信息的传播,提高命名实体识别效果,但FALT模型依赖于高质量的词典信息。针对这一问题,提出了结合神经文本生成的FLAT模型,使用神经文本生成方法生成大量新病历文本,通过提出的评分函数筛选通顺的文本训练词向量作为FLAT模型的词典信息。实验表明:结合神经文本生成的FLAT模型在CCKS2017数据集上取得了95.32%的F1分数,比BiLSTM CRF模型提高了1.16%,比BERT CRF模型提高了0.89%;在CCKS2019数据集上取得了85.87%的F1分数,比BiLSTM CRF模型提高了5.19%,比BERT CRF模型提高了1.34%。

With the development of medical informatization, named entity recognition of electronic medical records has attracted wide attention. Electronic medical records contain a large number of professional words, and the segmentation error of professional words will make named entity recognition ineffective. FLAT model introducing word boundary informationcan effectively avoid the spread of word segmentation error information, so as to improve the effect of named entity recognition. However, FALT model relies on high-quality dictionary information. In order to solve this problem, FLAT model combined with neural network text generation is proposed, which can generate a large number of new medical records. The proposed scoring function is used to select the smooth text to train the word vector, which is used as the dictionary information of the FLAT model. Experiments show that the LSTM-FLAT model obtains 95.32% F1 score on the CCKS 2017 dataset, which is 1.16% higher than the BiLSTM CRF model and 0.89% higher than the BERT CRF model. The F1 score on CCKS 2019 dataset is 85.87%, 5.19% higher than BiLSTM CRF model and 1.34% higher than BERT CRF model.