摘要
为了解决食品安全网络舆情中评论文本情感复杂多样,且依赖讨论的话题和博文信息的问题,提出一种融合话题博文的评论文本多情感分析模型TBR-MSAM(Topic Blog Review-Multi-Sentiment Analysis Model)。首先,使用RoBERTa(Robustly optimized BERT(Bidirectional Encoder Representations from Transformers)pretraining approach)和深度学习模型构建话题博文评论特征提取(TBR-FE)模块分别对话题、博文和评论信息进行上下文特征提取;其次,构建话题博文评论的交互注意力特征融合(TBR-IAFF)模块对话题-评论和博文-评论进行两两交互以获得交互特征,并进行权重的合理分配,从而挖掘话题、博文和评论之间的复杂关系;接着,构建话题博文评论的交叉特征融合(TBR-CFF)模块对多个信息进行深层次特征融合,从而挖掘用户潜在的情感特征;最后,通过Softmax对食品安全网络舆情中评论文本的4种情感极性进行分类。在所构建的3个食品安全网络舆情数据集上的实验结果表明,相较于无话题和博文信息的最优基线模型,TBR-MSAM的Macro-F1和准确率分别至少提升了5.0和5.8个百分点;相较于融合话题和博文信息的最优基线模型,TBR-MSAM的Macro-F1和准确率分别至少提升0.2和1.1个百分点;相较于同时带有话题、博文和评论文本信息的最优基线模型,TBR-MSAM的Macro-F1和准确率分别至少提升了11.7和10.0个百分点,验证了TBR-MSAM在食品安全网络舆情的多情感分类任务中的有效性。
To address the issues that the sentiments of review texts in food safety network public opinions are various and depend on the topic and blog information,a Multi-Sentiment analysis model for review texts integrating topics and blogs named TBR-MSAM(Topic Blog Review-Multi-Sentiment Analysis Model)was proposed.Firstly,a Topic Blog Review-Feature Extraction(TBR-FE)module was constructed by using RoBERTa(Robustly optimized BERT(Bidirectional Encoder Representations from Transformers)pretraining approach)and deep learning models,and was used to extract contextual features of topic,blog and review information respectively.Secondly,a Topic Blog Review-Interactive Attention Feature Fusion(TBR-IAFF)module was built to conduct pairwise interactions between topic-review and blog-review to obtain interaction features and allocate weights reasonably,thereby exploring the complex relationships among topics,blogs and reviews.Thirdly,a Topic Blog Review-Cross Feature Fusion(TBR-CFF)module was constructed to conduct in-depth feature fusion on multiple pieces of information,thereby exploring users’potential sentimental features.Finally,Softmax was used to classify the four sentiment polarities of review texts in food safety network public opinions.Experimental results on three constructed food safety network public opinion datasets show that compared to the optimal baseline model without topic and blog information,TBR-MSAM achieves at least 5.0 and 5.8 percentage points improvements in Macro-F1 and accuracy,respectively;compared to the optimal baseline model with topic and blog information,TBR-MSAM achieves the Macro-F1 and accuracy enhanced by at least 0.2 and 1.1 percentage points,respectively;compared to the optimal baseline model with topic,blog,and review text information,TBR-MSAM achieves the Macro-F1 and accuracy increased by at least 11.7 and 10.0 percentage points,respectively.The above verifies the effectiveness of TBR-MSAM in multi-sentiment classification task for food safety network public opinion.
作者
吕星辰
林伟君
黄红星
LYU Xingchen;LIN Weijun;HUANG Hongxing(Institute of Agricultural Economics and Information,Guangdong Academy of Agricultural Sciences,Guangzhou Guangdong 510640,China;Key Laboratory of Urban Agriculture in South China,Ministry of Agriculture and Rural Affairs(Guangdong Academy of Agricultural Sciences),Guangzhou Guangdong 510640,China)
出处
《计算机应用》
北大核心
2025年第12期3786-3795,共10页
journal of Computer Applications
基金
2024年度广东省哲学社会科学规划项目(GD24XGL021)。
关键词
食品安全网络舆情
多情感分析
RoBERTa
交互注意力网络
注意力机制
特征融合
food safety network public opinion
multi-sentiment analysis
RoBERTa(Robustly optimized BERT(Bidirectional Encoder Representations from Transformers)pretraining approach)
Interactive Attention Network(IAN)
attention mechanism
feature fusion
