中文命名实体识别的傅立叶卷积网络  

Fourier Convolutional Network for Chinese Named Entity Recognition

作　　者：李彪 LI Biao(College of Sciences,Northeastern University,Shenyang 110819,China)

机构地区：[1]东北大学理学院,辽宁沈阳110819

出　　处：《现代信息科技》2022年第2期104-106,共3页Modern Information Technology

摘　　要：针对transformer编码器架构在中文命名实体识别任务上表现不佳的问题,提出使用无参数化的傅立叶子层替换编码器中自注意力子层,使用卷积神经网络替代前馈神经网络。实验表明,采用结合傅立叶变换和卷积神经网络的transformer encoder架构的算法,可以在较小的字符嵌入和参数量下实现性能提升,且训练过程更快。In view of the poor performance of transformer encoder architecture in the task of Chinese named entity recognition,a non-parameter Fourier sublayer is proposed to replace the self attention sublayer in the encoder,and a convolutional neural network is used to replace the feedforward neural network.Experiments show that the algorithm based on the transformer encoder architecture combining Fourier transform and convolutional neural network can improve the performance with small character embedding and parameters,and the training process is faster.

关 键 词：中文命名实体识别 编码器 傅立叶变换 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]