Bert-BLSTM-CRF模型的中文命名实体识别

Chinese Named Entity Recognition Based on Bert-BLSTM-CRF Model

中文命名实体识别方法中采用较多的是神经网络模型,但该模型在训练过程中存在字向量表征过于单一的问题,无法很好地处理字的多义性特征。因此,提出一种基于Bert-BLSTM-CRF模型的中文命名实体识别研究方法,使用Bert预训练语言模型,根据字的上下文信息来丰富字的语义向量,将输出的字向量序列作为输入送入BLSTM-CRF模型进行训练。实验结果表明,此方法在中文命名实体识别任务上其准确率、召回率和F1值分别取得了94.80%、95.44%和95.12%的成绩,相较于其他传统方法,效果显著。

Most of the Chinese named entity recognition method are based on neural network model,but the process has a problem that the vector representation is too singular during the training process,and the ambiguity of the word cannot be handled well.Therefore,a method of Chinese named entity recognition based on Bert-BLSTM-CRF model is proposed.According to the context information of the word,the Bert pretraining language model is used to enrich the semantic vector of words,and the output word vector sequence is sent as input to BLSTM-CRF model for training.The experimental results show that the accuracy,recall rate and F1 value of this method in the Chinese named entity recognition task are 94.80%,95.44%and 95.12%,respectively.The effect is remarkable compared with other traditional methods.