基于Self-Attention和TextCNN-BiLSTM的中文评论文本情感分析模型  

Sentiment analysis model of Chinese commentary text based on Self-Attention and TextCNN-BiLSTM

作　　者：龙宇 李秋生[1,2] LONG Yu;LI Qiusheng(School of Physical and Electronics and Information,Gannan Normal University,Ganzhou,Jiangxi 341000,China;Research Center of Intelligent Control Engineering Technology,Gannan Normal University,Ganzhou,Jiangxi 341000,China)

机构地区：[1]赣南师范大学物理与电子信息学院,江西赣州341000 [2]赣南师范大学智能控制工程技术研究中心,江西赣州341000

出　　处：《石河子大学学报(自然科学版)》2025年第1期111-121,共11页Journal of Shihezi University(Natural Science)

基　　金：江西省教育厅科学技术研究项目(GJJ201408)。

摘　　要：目前关于中文评论文本的情感分类方法大都无法充分捕捉到句子的全局语义信息,同时也在长距离的语义连接或者情感转折理解上具有局限性,因而导致情感分析的准确度不高。针对这个问题,本文提出一种融合SelfAttention和TextCNN-BiLSTM的文本情感分析方法。该方法首先采用文本卷积神经网络(TextCNN)来提取局部特征,并利用双向长短期记忆网络(BiLSTM)来捕捉序列信息,从而综合考虑了全局和局部信息,在特征融合阶段,再采用自注意力机制来动态地融合不同层次的特征表示,对不同尺度特征进行加权,从而提高重要特征的响应。实验结果表明,所提出的模型在家电商品中文评论语料和谭松波酒店评论语料数据集上的准确率分别达到93.79%和90.05%,相较于基准模型分别提高0.69%~3.59%和4.44%~11.70%,优于传统的基于卷积神经网络(Convolutional Neural Networks, CNN)、BiLSTM或CNN-BiLSTM等的情感分析模型。Most of the current sentiment classification methods on Chinese commentary texts cannot fully capture the global semantic information of sentences,and also have limitations in the understanding of long-distance semantic connections or emotional transitions,which leads to the low accuracy of sentiment analysis.To solve this problem,this paper proposes a text sentiment analysis method that integrates self-attention mechanism and TextCNN-BiLSTM.The method uses multi-scale convolutional neural network(TextCNN)to extract local features,and utilizes bidirectional long short-term memory network(BiLSTM)to capture sequence information,so that,global and local information is considered comprehensively,in the feature fusion stage,the self-attention mechanism is adopted to dynamically integrate the feature representations of different levels and weight the features of different scales,so as to improve the response of important features.The experimental outcomes demonstrate that the proposed model achieves an accuracy of 93.79%on the Chinese comment corpus,with a specific accuracy of 90.05%in certain subsets.This performance represents a significant improvement over traditional sentiment analysis models,outperforming them by 0.69%to 3.59%and an impressive 4.44%to 11.70%.The proposed method thus surpasses conventional models that are based on Convolutional neural networks(CNN),BiLSTM,or CNN-BiLSTM architectures.

关 键 词：自注意力机制 中文评论文本 深度学习 情感分析 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]