基于BERT的中文命名实体识别方法

Chinese Named Entity Recognition Method Based on BERT

摘要: 由于中文与英文本身存在较大的差异，中文命名实体识别的研究存在一系列的挑战。目前来说，BLSTM-CRF模型使用最为广泛。该模型采用深度学习模型与统计模型相结合的方式进行中文命名实体识别，能够有效提取出文本中的上下文信息并考虑标签之间的关系。但由于中文存在多义字或词，存在一个句子中相同字词含义差别很大的情况，该模型在这种情况下实体识别的性能并不理想。为了更好地实现字表示既可以包含各种多样化的句法和语义表示，又可以对多义字进行建模，引入了BERT语言模型，此模型可以根据上下文信息计算出更高的全局性字词向量表示以及在句中的权重。BERT-BLSTM-CRF命名实体识别模型通过BERT预训练模型增强词向量的表示，BLSTM获取上下文语义标签序列，再使用CRF求得最优解。本文使用人民日报数据集对提出模型的进行实验测试，从实验结果可以发现，该模型的实体识别性能与传统模型相比有较大的提升。

Abstract: Due to the large differences between Chinese and English, there are a series of challenges in the re-search of Chinese named entity recognition. Currently, the BLSTM-CRF model is the most widely used. The model uses a combination of deep learning models and statistical models for Chinese named entity recognition, which can effectively extract contextual information in the text and con-sider the relationship between tags. However, due to the presence of polysemous characters or words in Chinese, the meaning of the same words in a sentence may be very different. In this case, the performance of the entity recognition of the model is not ideal. In order to better realize that the word representation can not only contain a variety of diversified syntax and semantic representa-tions, but also can model polysemous words, the BERT language model is introduced, which can calculate higher global word vector representations and weights in sentences based on context in-formation. The BERT-BLSTM-CRF named entity recognition model enhances the representation of word vectors through the BERT pre-training model, uses BLSTM to obtain contextual semantic label sequences, and then uses CRF to find the optimal solution. Using the People’s Daily data set to test the proposed model, it can be found that the entity recognition performance of the model is greatly improved compared with the traditional model.