融合MA-CapsNet的社交媒体短文本情感分类研究——基于脱氢乙酸钠微博数据

Research on Sentiment Classification of Social Media Short Texts Integrating MA-CapsNet—Based on the Weibo Data of Sodium Dehydroacetate

摘要: 针对传统情感分类模型在短文本语义捕捉、领域特异性处理及情感强度量化上的不足，本文提出融合多粒度注意力机制与胶囊网络的MA-CapsNet模型，用于社交媒体短文本情感分类研究。模型构建“双流–三阶”分析框架：输入表示层采用混合嵌入策略，融合预训练BERT向量和位置编码；多粒度注意力层通过字符级、词汇级、句子级三重特征提取与门控融合，增强对“零添加”等领域表达的识别；胶囊网络层利用动态路由算法建模特征间的层次关系，有效处理“并非不安全”等复杂否定结构；输出层实现情感五分类(非常消极至非常积极)与强度量化(1~10分)，并采用联合损失函数来优化双任务性能。结果表明，MA-CapsNet在情感分类任务上准确率达87.6%，宏平均F1值0.857，显著优于SVM (72.1%)、TextCNN (78.9%)、BERT (84.2%)等模型；情感强度预测MAE为0.73，皮尔逊相关系数0.82，体现出对情感细微差异的精准捕捉能力。研究表明，MA-CapsNet能有效提升食品安全舆情的情感分析精度，为舆情监测与引导提供技术支撑，其领域自适应设计对其他专业领域的短文本分析具有借鉴意义。

Abstract: In view of the deficiencies of traditional sentiment classification models in semantic capture of short texts, domain-specific processing and quantification of sentiment intensity, this paper proposes the MA-CapsNet model integrating multi-granularity attention mechanism and capsule network for the research of sentiment classification of short texts in social media. Model constructs “dual-stream-third-order” analysis framework: The input representation layer adopts a hybrid embedding strategy, fusing pre-trained BERT vectors and position encodings; The multi-granularity attention layer enhances the recognition of expressions in fields such as “zero addition” through the fusion of triple feature extraction at the character level, vocabulary level, and sentence level with gating. The capsule network layer uses dynamic routing algorithms to model the hierarchical relationship between features, effectively handling complex negative structures such as “not insecure”. The output layer implements five-classification of emotions (from very negative to very positive) and intensity quantification (1~10 points), and a joint loss function is adopted to optimize the performance of the dual tasks. The results show that the accuracy rate of MA-CapsNet in the emotion classification task reaches 87.6%, and the macro average F1 value is 0.857, which is significantly better than models such as SVM (72.1%), TextCNN (78.9%), and BERT (84.2%). The MAE for emotional intensity prediction is 0.73, and the Pearson correlation coefficient is 0.82, demonstrating the precise ability to capture the subtle differences in emotions. Studies show that MA-CapsNet can effectively improve the accuracy of sentiment analysis of food safety public opinions, provide technical support for public opinion monitoring and guidance, and its domain adaptive design has reference significance for short text analysis in other professional fields.