Research on Chinese Naming Recognition Model Based on BERT Embedding

Abstract

Named entity recognition (NER) is one of the foundations of natural language processing(NLP). In the method of Chinese named entity recognition based on neural network, the vector representation of words is an important step. Traditional word embedding method map words or chars into a single vector, which can not represent the polysemy of words. To solve this problem, a named entity recognition method based on BERT Embedding model is proposed. The method enhances the semantic representation of words by BERT(Bidirectional Encoder Representations from Transformers) pre-trained language model. BERT can generates the semantic vectors dynamically according to the context of the words, and then inputs the word vectors into BiGRU-CRF for training. The whole model can be trained during training. It is also possible to fix the BERT and train only the BiGRU-CRF part. Experiments show that the two training methods of the model reach 95.43% F1 and 94.18% F1 in MSRA corpus, respectively, which are better than the current optimal Lattice-LSTM model.