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

Abstract: To address the issues that the sentiment of comment texts in food safety network public opinions is various and depends on the topics or blog, a Multiple Sentiment Analysis Model for Comment Texts Integrating Topic and Blog Post Information (TBR-MSAM) was proposed. Firstly, the Topic Blog Review Feature Extraction Module (TBR-FE) was constructed by using RoBERTa and deep learning models, with which the contextual features of topic, blog post and review information were respectively extracted. Secondly, the Interactive Attention Feature Fusion Module for Topic Blog Review (TBR-IAFF) was built to conduct pairwise interactions between topic-review and blog-review to obtain interaction features, allocate weights reasonably and explore the complex relationships among topics, blog posts and reviews. Next, a Cross Feature Fusion Module for Topic Blog Review (TBR-CFF) was constructed to conduct in-depth feature fusion on multiple pieces of information and mine users' potential emotional features. Finally, softmax was used to classify the four emotional polarities of review texts in food safety network public opinions. The experimental results on the three food safety network public opinion datasets constructed show that, compared to the optimal baseline model without topic and blog information, the TBR-MSAM achieved at least 5.0% and 5.8% improvement in Macro-F1 and accuracy, compared to the optimal baseline model with cross-fused topic and blog information, the Macro-F1 and accuracy were enhanced by at least 0.2% and 1.1%, compared to the optimal baseline model incorporating topic and blog, the Macro-F1 and accuracy were increased by at least 11.7% and 10.0%. These findings verified the effectiveness of the proposed TBR-MSAM model in the multi-sentiment classification task for food safety network public opinion.

Key words: food safety network public opinion, multi-sentiment analysis, RoBERTa, interactive attention network, attention mechanism, feature fusion