【摘要】 在大规模文本语料库上预先训练的BERT（bidirectional encoder representation from transformers, BERT）等神经语言表示模型可以较好地从纯文本中捕获丰富的语义信息。但在进行中文命名实体识别任务时,由于命名实体存在结构复杂、形式多样、一词多义等问题,识别效果不佳。基于知识图谱可以提供丰富的结构化知识,从而更好地进行语言理解,提出了一种融合知识图谱信息的中文命名实体识别方法,通过知识图谱中的信息实体增强语言的外部知识表示能力。实验结果表明,与BERT、OpenAI GPT、ALBERT-BiLSTM-CRF等方法相比,所提出的方法有效提升了中文命名实体的识别效果,在MSRA（Microsoft Research Asia, MSRA）与搜狐新闻网标注数据集上,F₁值分别达到了95.4%与93.4%。更多还原

【Abstract】 Neural language representation models such as bidirectional encoder representation from transformers（BERT）,which are pre-trained on large-scale text corpora, can well capture rich semantic information from plain text.However, due to the facts of the Chinese named entity like complexity of the structure, various forms and polysemous meaning, the Chinese named entity recognition results in poor performance.In view of the fact that knowledge graphs can provide rich structured knowledge facts for better language understanding, a Chinese named entity recognition method integrating knowledge graph information was proposed in the paper to enhance the language’s external knowledge representation ability through the information entities in knowledge graphs.The experimental results showed that the proposed method effectively improved the recognition effect of Chinese named entities compared with BERT,OpenAI GPT, ALBERT-BiLSTM-CRF and other methods, and the F₁ value reached 95.4% and 93.4% on MSRA and Sohu news-marking data sets respectively.更多还原