基于注意时序网络的中文词性标注方法  

Chinese Part-of-Speech Tagging Method Based on Attention Temporal Network

作　　者：张鹏[1,2] 周志强[1,2] ZHANG Peng;ZHOU Zhi-qiang(Software Engineering Department,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Intelligent Information Technology and Service Innovation Laboratory of Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学软件工程学院,重庆400065 [2]重庆邮电大学智能信息技术与服务创新实验室,重庆400065

出　　处：《计算机仿真》2024年第5期378-382,共5页Computer Simulation

基　　金：重庆市高等教育教学改革研究重大项目(221017)。

摘　　要：针对传统的基于统计与规则的词性标注模型存在的人工特征依赖、字向量表征单一、特征提取不全面等问题,提出一种有效的基于注意时序网络的中文词性标注模型。对原始的TCN模型结构进行三点改进,并提出将注意时序网络与BiLSTM模型融合到词性标注方法中。上述模型首先通过XLNet模型获取字级别的上下文表示,利用注意时序网络的因果卷积结构获取更高层次的文本序列特征并通过注意力机制优化特征,最后通过BiLSTM进一步学习序列上下文特征,提高词性标注的准确度。实验表明,上述模型性能相较于其它模型有明显提升。Because the traditional Chinese part-of-speech tagging(CPOS)models based on statistics and rules hase many problems such as relying heavily on manually designed features,word vectors represent singleness,feature extraction is not comprehensive,this paper proposes an effective Chinese part-of-speech tagging model based on Temporal Convolutional Network with Attention(TCA).This model improved the structure of the original TCN model in three aspects,and proposed the integration of TCA and BiLSTM into CPOS method.In this model,the XLNet model was used to obtain word-level context representation,and TCN’s unique causal convolution structure was used to obtain higher-level text sequence features and optimize the features through the attention mechanism.Finally,BiLSTM was used to further learn the sequence context features to improve the accuracy of pos tagging.The experimental results show that the performance of this model is significantly improved compared with other models.

关 键 词：词性标注 时序卷积网络 注意力机制 深度学习 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]