融入频域增强自注意力机制的BTBFA混合神经网络情感分类模型  

BTBFA Hybrid Neural Network Sentiment Classification Model with Frequency Domain Enhanced Self-Attention Mechanism

作　　者：苏妍嫄 韩翠娟 张亚明[1,2,3] Su Yanyuan;Han Cuijuan;Zhang Yaming(School of Economics and Management,Yanshan University,Qinhuangdao 066004,China;Research Center of Internet Plus and Industry Development,Yanshan University,Qinhuangdao 066004,China;Research Center of Regional Economic Development,Yanshan University,Qinhuangdao 066004,China)

机构地区：[1]燕山大学经济管理学院,河北秦皇岛066004 [2]燕山大学互联网+与产业发展研究中心,河北秦皇岛066004 [3]燕山大学区域经济发展研究中心,河北秦皇岛066004

出　　处：《现代情报》2024年第12期52-63,共12页Journal of Modern Information

基　　金：国家自然科学基金项目“重大疫情跨场域耦合的网络舆情非常规突变模型与异步调控算法研究”(项目编号:72101227);河北省自然科学基金“重大疫情下复杂社会网络舆情传播演化建模与动态调控算法研究”(项目编号:G2020203003);河北省高等学校人文社会科学重点研究基地项目“数字赋能基层应急韧性管理体系构建与效能提升路径研究”(项目编号:JJ2303);河北省自然科学基金“重大突发事件多风险交互的网络舆论极化模型与跨模态预警研究”(项目编号:G2024203001);河北省自然科学基金“基于人工智能的精准国际传播机理与效果测度研究”(项目编号:G2024023024)。

摘　　要：[目的/意义]智媒时代基于神经网络模型实现用户情感精准分类,进而深入挖掘海量文本信息潜在价值具有重要意义。[方法/过程]针对现有混合模型层间依赖性强、输出特征重要性差异体现不足等导致的情感分类效果受限问题,基于Stacking集成思想,提出一种融入频域增强自注意力机制的混合神经网络情感分类模型,通过构建由Bert、TextCNN、BiLSTM组成的并行式特征提取基学习器层与融入频域增强自注意力机制的元学习器层,并与词嵌入层和全连接层相融合,系统挖掘文本深层次语义信息以及局部、全局特征,进而通过权重分配以及离散傅里叶变换提升情感分类效果。[结果/结论]酒店评论数据集上的对比实验与消融实验结果均表明,所提模型情感分类性能与其他模型相比具有显著优势,准确率、召回率、F1值分别达到91.7%、95.3%和93.9%,且随Epoch训练轮数增加,模型情感分类准确性不断提升,损失值不断降低,呈现较强的泛化能力。[Purpose/Significance]In the era of smart media,it is of great significance to achieve accurate classification of users sentiment based on neural network model and deeply explores the potential value of massive text information.[Method/Process]Aiming at the problem of the limited effect of sentiment classification caused by strong dependence between layers of existing hybrid models and insufficient reflection of the importance difference of output features and so on,the paper proposed a hybrid neural network sentiment classification model with frequency domain enhanced self-attention mechanism based on the Stacking integration algorithm.Firstly,the paper constructed the parallel feature extraction base-learner layer combining with Bert,TextCNN,BiLSTM models.Secondly,the paper constructed the meta-learner layer incorporating the frequency domain enhanced self-attention mechanism.Thirdly,the paper fused the two layers based on Stacking algorithm,and combining word embedding layer and fully connected layer,systematically mine the deep semantic information as well as the local and global features,and then through the distribution of weight and discrete Fourier transform to improve the effect of sentiment classification.[Result/Conclusion]The results of comparative experiments and ablation experiments on the hotel review dataset show that the Bert-TextCNN-BiLSTM-FAttention sentiment classification model has a significant advantage over other models,and its accuracy,recall,and F1 value reach 91.7%,95.3%,and 93.9%,respectively.Besides,with the increase in the number of Epoch training rounds,the sentiment classification accuracy of the model continues to improve,and the loss value continues to decrease,which indicates that the model has a strong generalization ability.

关 键 词：情感分类 混合神经网络 Bert-TextCNN-BiLSTM-FAttention Stacking算法 自注意力机制 离散傅里叶变换 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]