Scholarlink

<p>Multimodal aspect-based sentiment analysis (MABSA) seeks to determine sentiment polarity toward specific aspects within text by incorporating multimodal information. However, existing approaches encounter two key challenges: (1) limited cross-modal modeling, which may result in modality misalignment and visual noise interference, and (2) irregular textual structures–such as discontinuous or nested expressions–that hinder accurate sentiment reasoning. To address these issues, we propose a novel framework that employs the pre-trained vision-language model BLIP (Bootstrapping Language-Image Pre-training) to generate descriptive image captions. This process facilitates semantically aligned text-image fusion, effectively bridging the cross-modal semantic gap and filtering irrelevant visual noise. In addition, we introduce an aspect-guided soft prompt mechanism that enables dynamic interaction between aspect terms and multimodal features, thereby mitigating the effects of structural irregularities. To assess the robustness of our method, we construct and annotate a new dataset–ICED (International Conflict Events Dataset)–characterized by irregular sentence structures, and conduct experiments on Twitter15, Twitter17, and ICED. Experimental results show that our model consistently outperforms existing baselines in both accuracy and robustness.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multimodal sentiment analysis based on image captions and aspect-guided soft prompts

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Yuchen Li,
  • Changhong Yu,
  • Binxiao Yu

摘要

Multimodal aspect-based sentiment analysis (MABSA) seeks to determine sentiment polarity toward specific aspects within text by incorporating multimodal information. However, existing approaches encounter two key challenges: (1) limited cross-modal modeling, which may result in modality misalignment and visual noise interference, and (2) irregular textual structures–such as discontinuous or nested expressions–that hinder accurate sentiment reasoning. To address these issues, we propose a novel framework that employs the pre-trained vision-language model BLIP (Bootstrapping Language-Image Pre-training) to generate descriptive image captions. This process facilitates semantically aligned text-image fusion, effectively bridging the cross-modal semantic gap and filtering irrelevant visual noise. In addition, we introduce an aspect-guided soft prompt mechanism that enables dynamic interaction between aspect terms and multimodal features, thereby mitigating the effects of structural irregularities. To assess the robustness of our method, we construct and annotate a new dataset–ICED (International Conflict Events Dataset)–characterized by irregular sentence structures, and conduct experiments on Twitter15, Twitter17, and ICED. Experimental results show that our model consistently outperforms existing baselines in both accuracy and robustness.