基于多特征融合和注意力机制的中文命名实体识别  

Chinese Named Entity Recognition Based on Multi-feature Fusion and Attention Mechanism

作　　者：陈奕全 吴晓鸰 占文韬 HEO Hoon CHEN Yiquan;WU Xiaoling;ZHAN Wentao;HEO Hoon(School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510006,China;SamSung Electro-mechanics,Korea)

机构地区：[1]广东工业大学计算机学院,广州510006 [2]三星电机,韩国

出　　处：《小型微型计算机系统》2025年第2期339-345,共7页Journal of Chinese Computer Systems

基　　金：广东省重点领域研发计划项目(2019B010139002)资助;广东省国际科技合作领域项目(2019A050513010)资助。

摘　　要：近年来,基于字符级嵌入的BERT模型和基于词融合的模型都在中文命名实体识别领域取得了较好的效果,但目前这些模型从字符序列中提取出的特征信息还不够充分,模型性能还有一定的提升空间.因此本文提出了一种用于中文命名实体识别的多特征融合模型,首先将输入中文语句转换为字词对序列,通过RoBERTa-wwm预训练语言模型将字词对序列中的字符序列表征为字符向量,获得全局语义特征;然后把词序列转化为词向量,再将字符向量和词向量输入到基于双线性注意力机制的词汇适配器获得字词融合特征;同时将字符向量送入到双向长短时记忆网络(BiLSTM)获得包含字符方向信息的上下文特征;最后将词汇适配器的输出和BiLSTM的输出进行动态特征融合获得包含全局语义信息、词汇信息和方向信息的上下文特征,再通过CRF解码获得最优预测序列.在多个公共数据集的实验结果验证了该模型能提取到更充分的特征信息,提高了识别性能.In recent years,both character-level embeddings-based BERT models and word fusion models have shown promising results in the field of Chinese Named Entity Recognition(NER).However,currently feature information extracted from character sequences is not sufficient,leaving room for improvement in model performance.Therefore,this paper proposes a multi-feature fusion model for Chinese named entity recognition.Firstly,it transforms input Chinese sentences into character-word pairs sequences and employs the RoBERTa-wwm pre-trained language model to represent character sequences as character vectors,capturing global semantic features.Subsequently,it converts word sequences into word vectors and feeds both character and word vectors into a vocabulary adapter based on a bilinear attention mechanism,producing character-word fusion features.Simultaneously,character vectors are input into a Bidirectional Long Short-Term Memory network(BiLSTM)to obtain contextual features that include character direction information.Finally,the outputs of the vocabulary adapter and BiLSTM are dynamically fused to generate context features that encompass global semantics information,lexical information,and direction information.The model then employs Conditional Random Fields(CRF)for decoding to obtain the optimal predicted sequences.The experimental results on multiple public datasets confirm that the model is capable of extracting more comprehensive feature information,thereby enhancing recognition performance.

关 键 词：中文命名实体识别 多特征融合 词融合 预训练模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]