基于细粒度图像-方面的情感增强方面级情感分析  

Aspect-oriented affective knowledge enhanced for aspect-based sentiment analysis

作　　者：余本功[1,2] 陈明玥 Yu Bengong;Chen Mingyue(School of Management,Ministry of Education,Hefei University of Technology,Hefei 230009,China;Key Laboratory of Process Optimization&Intelligent Decision-making,Ministry of Education,Hefei University of Technology,Hefei 230009,China)

机构地区：[1]合肥工业大学管理学院,合肥230009 [2]合肥工业大学过程优化与智能决策教育部重点实验室,合肥230009

出　　处：《计算机应用研究》2025年第4期1073-1079,共7页Application Research of Computers

基　　金：国家自然科学基金资助项目(72071061)。

摘　　要：为了缩小模态间的异质性差异并缓解多个方面词带来的情感混淆,提出一种基于细粒度图像-方面的情感增强多模态方面级情感分析。具体地,该模型经过文本图像编码后,首先利用形容词-名词对将与方面词相关的图像信息加入到文本方面词中,并通过细粒度图像-方面跨模态注意力机制优化图像表征,得到细粒度方面词-图像特征;接着,基于句法结构引入情感得分,得到基于方面词的文本情感特征;最后,进行模态融合得到最终情感预测结果。在Twitter-2015和Twitter-2017数据集上,与基线模型TMSC相比,提出模型值准确率分别提高了0.25百分点和0.16百分点,充分证明了细粒度的图文匹配和情感增强操作有助于提高分类效果。To reduce the heterogeneity differences between modalities and alleviate the emotional confusion caused by multiple aspect words,this paper proposed a fine-grained image-aspect emotional enhancement model for multimodal aspect-based sentiment analysis.Specifically,after encoding text and images,the model first integrated image information related to aspect words into the textual aspect words using adjective-noun pairs.It then optimized the image representation through a fine-grained image-aspect cross-modal attention mechanism to obtain fine-grained aspect-word-image features.Next,it introduced sentiment scores based on syntactic structure to derive textual sentiment features based on aspect words.Finally,modality fusion was performed to obtain the final sentiment prediction results.This method achieves an accuracy improvement of 0.25 and 0.16 percentage points on the Twitter-2015 and Twitter-2017 datasets,respectively,compared to the baseline model TMSC,de-monstrating that fine-grained image-text matching and emotional enhancement operations contribute to improving classification performance.

关 键 词：多模态方面级情感分析 形容词-名词对 跨模态注意力机制 情感分数 模态融合 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]