Chinese Named Entity Recognition Method Based on BERT Embedding
摘要
In Chinese Named Entity Recognition(NER) based on neural network,the vectorized representation of words is an important step.Traditional representation methods for word vectors only map a word to a single vector,and cannot represent the polysemy of a word.To address the problem,this paper introduces the BERT pretrained language model to build a BERT-BiGRU-CRF model for representation of sentence characteristics.The BERT pretrained language model with bidirectional Transformer structure is used to enhance the semantic representation of words and generate semantic vectors dynamically based on their context.On this basis,the word vector sequence is input into the BIGR-CRF model to train the whole model,or train the BIGR-CRF part only with BERT fixed.Experimental results on MSRA data show that the F1 value in the two training modes of this proposed model reaches 95.43% and 94.18% respectively,which is better than that of the BIGRU-CRF,the RADICAL-BILSTM-CRF and the GRAIN-LSTM-CRF models.