摘要
以往的文本情感分析模型存在忽略文本边缘信息、池化层破坏文本序列特征的问题,并且特征提取能力与识别关键信息的能力也存在不足。为了进一步提升情感分析的效果,提出了一种基于注意力机制的动态卷积神经网络(Dynamic Convolutional Neural Network,DCNN)与双向门控循环单元(Bi-directional Gated Recurrent Unit,BiGRU)的文本情感分析模型DCNN-BiGRU-Att。首先,利用宽卷积核提取文本边缘特征,采用动态k-max池化保留了文本的相对位置序列特征。其次,构建了DCNN与BiGRU的并行混合结构,避免了部分特征损失问题,并同时保留局部特征与全局上下文信息两种特征,提高了模型的特征提取能力。最后,在特征融合之后引入注意力机制,将注意力机制的作用全局化,提高了模型识别关键信息的能力。将该模型在MR与SST-2两个公开数据集上与多个深度学习模型进行对比,其准确率分别提高了1.27%和1.07%,充分证明了该模型的合理有效性。
The previous text sentiment analysis models ignore the edge information of the text,the pooling layer destroys the sequence features of the text,and the feature extraction ability and the ability to identify key information are also inadequate.In order to further improve the effect of sentiment analysis,a text sentiment analysis model of dynamic convolutional neural network(DCNN)and bi-directional gated recurrent unit(BiGRU)based on attention mechanism is proposed.Firstly,the wide convolution kernel is used to extract text edge features,and the dynamic k-max pooling layer can retain the relative position information of text.Secondly,the parallel hybrid structure of DCNN and BiGRU is constructed.This structure can avoid the problem of partial feature loss and retain both local feature and global context information,which improves the feature extraction capability of the model.Finally,the attention mechanism is introduced after feature fusion layer to globalize the impact of attention mechanism and improve the model s capability to recognize key information.Comparison between the proposed model and other deep learning models on MR and SST-2 public datasets shows,the accuracy of the proposed model is improved by 1.27%and 1.07%respectively,which fully proves the reasonable validity of the proposed model.
作者
孔韦韦
田乔鑫
滕金保
王照乾
常亮
KONG Weiwei;TIAN Qiaoxin;TENG Jinbao;WANG Zhaoqian;CHANG Liang(School of Computer,Xi’an University of Posts and Telecommunications,Xi’an 710121,China;Guangxi Key Laboratory of Trusted Software,Guilin University of Electronic Technology,Guilin 541004,China)
出处
《电讯技术》
北大核心
2023年第6期781-789,共9页
Telecommunication Engineering
基金
国家自然科学基金资助项目(61772396,61902296)
广西可信软件重点实验室研究课题(KX202061)。
