基于改进CNN-BiGRU-att模型的文本分类研究  

Text Classification Based on Improved CNN-BiGRU-att Model

作　　者：陈农田[1] 李俊辉 满永政 CHEN Nongtian;LI Junhui;MAN Yongzheng(Civil Aviation Flight University of China,Guanghan,Sichuan 618307,China)

机构地区：[1]中国民用航空飞行学院,四川广汉618307

出　　处：《昆明理工大学学报（自然科学版）》2022年第1期30-37,共8页Journal of Kunming University of Science and Technology(Natural Science)

基　　金：国家重点研发计划项目(2018YFC0809500);四川省科技厅重点研发计划项目(22ZDYF2942);中国民航局安全能力基金项目(2022J026);中央高校教学科研专项项目(E2020063—3);民航飞行技术与飞行安全科研基地开放基金项目(F2019KF08);研究生科技创新基金项目(X2021-21)。

摘　　要：为解决单一的卷积神经网络(CNN)缺乏利用上下文本信息与单一循环神经网络(RNN)对局部信息把握不全面问题,提出一种基于注意力机制的多通道TextCNN-BiGRU分类模型.首先,通过word2vec对初始文本向量化,经实验选取窗口值组成三通道.然后利用CNN的强学习能力提取局部特征,利用双向门控循环单元(BiGRU)提取上下文全局信息,运用注意力层与池化层获取并优化重要的特征.最后采用softmax函数使误差loss极小化.仿真实验结果表明,提出的模型分类性能,准确度达94%,损失函数值稳定在0.22%左右,具有良好的泛化能力,能够有效解决单一模型挖掘信息不全问题,有效提高分类效果.To solve the problem that a single convolutional neural network(CNN) lacks the use of contextual information and a single recurrent neural network(RNN) to grasp the local information incompletely, a multi-channel TextCNN-BiGRU classification model based on the attention mechanism is proposed. Firstly, the initial text is vectorized by word2 vec, and the window values are selected through experiments to form three channels. Then, the strong learning ability of CNN is used to extract local features, the bidirectional gated recurrent unit(BiGRU) is used to extract contextual global information, and the attention layer and pooling layer are used to obtain and optimize the important features. Finally, the softmax function is used to minimize the error loss. The simulation experimental verification results show that the classification performance of the proposed model has a classification accuracy of 94%, and the loss function value is stable at about 0.22%, which can effectively solve the problem of incomplete information mining with a single model and effectively improve the classification effect.

关 键 词：深度学习 神经网络模型 文本多分类 注意力机制 文本特征提取 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]