融合多阶段特征的中文命名实体识别模型  

Chinese named entity recognition model based on multistage feature fusion

作　　者：杨先凤[1] 范玥 李自强 汤依磊 YANG Xian-feng;FAN Yue;LI Zi-qiang;TANG Yi-lei(School of Computer and Software,Southwest Petroleum University,Chengdu 610500,China;School of Film,Television and Media,Sichuan Normal University,Chengdu 610066,China)

机构地区：[1]西南石油大学计算机与软件学院,四川成都610500 [2]四川师范大学影视与传媒学院,四川成都610066

出　　处：《计算机工程与设计》2025年第1期37-43,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61802321);四川省科技厅重点研发计划基金项目(2020YFN0019)。

摘　　要：针对中文命名实体识别中未充分利用完整的文本表示和语句特征的问题,提出一种融合多阶段特征的中文命名实体识别模型(LM-CNER)。采用全局注意力机制文本融合字符级嵌入与其预训练词向量,同时获取字符级特征和单词级特征。采用翻转长短时记忆网络(Re-LSTM)进行上下文特征提取,采用多头自注意力机制进行句法分析,并将二者进行拼接。使用条件随机场作为解码器,得到命名实体识别结果。在微博和简历两个数据集上的实验结果表明,该模型能够获取更加准确的文本表示和语句特征,提升模型的实体识别效果。A Chinese named entity recognition model(LM-CNER)that integrated multi-stage features was proposed,in response to the issue of underutilizing complete text representations and sentence features in Chinese named entity recognition(CNER).Character-level embeddings of the text with its pre-trained word vectors were integrated using a global attention mechanism,enabling the simultaneous acquisition of character-level and word-level features.Reversed long short-term memory networks(Re-LSTM)for extracting contextual information were employed and a multi-head self-attention mechanism for syntactic analysis was utilized,followed by concatenating these two.The named entity recognition results were obtained using a conditional random field as the decoder.Experimental results on both Weibo and resume datasets demonstrate that this model can capture more accurate text representations and sentence features,thereby enhancing the entity recognition performance of the model.

关 键 词：命名实体识别 翻转长短时记忆网络 注意力机制 编码器 预训练词向量 多阶段特征 条件随机场 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]