融合词性与双向时间卷积网络的中文命名实体识别方法  

Chinese named entity recognition method based on synthetical speech and bidirectional temporal convolutional networks

作　　者：张鹏[1,2] 周志强[1,2] ZHANG Peng;ZHOU Zhiqiang(School of Software Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;Intelligent Information Technology and Service Innovation Laboratory,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.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：《重庆邮电大学学报（自然科学版）》2023年第4期662-670,共9页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：重庆市高等教育教学改革研究重大项目(221017)。

摘　　要：针对目前中文命名时实体识别方法中存在的中文边界识别困难、模型梯度、文本特征不够充分等问题,提出了一种融合词性特征与双向时间卷积网络的中文命名时实体识别模型。该模型提出使用XLNet预训练语言模型生成对应的词嵌入表示,融合后使用双向时间卷积网络提取文本前向特征与后向特征。实验中对时间卷积网络的空洞因子、卷积层数和卷积核数进行参数实验并分析其影响原因,结果表明,该模型与其他模型相比,能够更准确且有效地提取文本中的实体。Aiming at the problems of Chinese boundary recognition,model gradient and insufficient text features in current entity recognition methods in Chinese naming,we propose a new entity recognition model for Chinese naming time,which integrates part of speech feature and bidirectional time convolution network.This model proposes to use XLNet pre-training language model to generate corresponding word embedding representation.After fusion,bidirectional temporal convolution network is used to extract forward and backward text features.In the experiment,three parameters of the time convolutional network,namely,the cavity factor,the number of convolution layers and the number of convolution kernel,are tested and the reasons for their influence are analyzed.Experiments on the 2014 People’s Daily dataset prove that the model is able to extract entities from the text accurately and efficiently compared to other models.

关 键 词：中文命名实体识别 词性特征 时序卷积网络 神经网络 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]