基于LEBERT-BCF的电子病历实体识别

Entity recognition of electronic medical records based on LEBERT-BCF

针对BERT在中文电子病历实体识别过程中缺少词信息,实体边界信息被浪费和模型鲁棒性较差等问题,提出一种基于BERT并引入外部词典进行特征增强和对抗训练的实体识别模型LEBERT-BCF。该模型通过外部词典自动为电子病历进行词汇匹配构建字符-词语对,在BERT内部将字符-词语对中对应字向量与词向量经过Lexicon Adapter模块进行特征融合并使用FGM提升模型的鲁棒性。在CCKS 2019数据集上的实验结果表明,该模型的F1值比BERTBiLSTM-CRF提高了3.45%。

Aiming at the problems of BERT in entity recognition of Chinese electronic medical records, such as lack of word information, waste of entity boundary information, and poor model robustness, an entity recognition model LEBERT-BCF based on BERT and introducing external dictionaries for feature enhancement and adversarial training is proposed. An external dictionary is used to automatically match words for electronic medical records and construct character-word pairs. The corresponding word vectors in the character-word pairs are fused with the word vectors in the internal BERT by the Lexicon Adapter module, and the robustness of the model is enhanced using FGM. The experimental results on the CCKS 2019 dataset show that the F1 value of the model is 3.45% higher than that of BERT-BiLSTM-CRF.