Scholarlink

<p>The multimodal aspect-based sentiment analysis (MABSA) task aims to determine the sentiment polarity associated with each specific aspect term mentioned in the text by integrating multimodal information, such as textual and visual data. This task is of significant importance for deeper comprehension of user sentiments and opinions. However, existing research suffers from the following limitation: (1) Although a correspondence exists between visual regions and words, current methods struggle to accurately learn the correct region-word alignment. (2) 2D images contain numerous features irrelevant to sentiment analysis, and existing methods lack an effective denoising mechanism. To tackle the mentioned issues, this paper provides a novel image-text relationship-based feature interaction network (ITRIN). Specifically, this paper introduces a cross-modal alignment correction module to learn the correspondence between region-level visual features and textual tokens and employs an adaptive gating mechanism to mitigate the adverse effects of misaligned region-word pairs. Furthermore, by incorporating global contextual information, we design a gating mechanism based on the probability scores of the image-text relationship (BRGate) in the final feature fusion layer, which enables deep filtering and effective fusion of multimodal features. The proposed model, ITRIN, achieves state-of-the-art performance on two benchmark datasets, Twitter-2015 and Twitter-2017. Extensive ablation studies further validate its superiority and effectiveness. The source code is publicly released at <a href="https://github.com/MHTransKanba/ITRIN">https://github.com/MHTransKanba/ITRIN</a>.</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

Image-text relationship-based feature interaction networks for multimodal aspect-based sentiment analysis

  • 期刊论文

  • 发表日期：

    2025年10月25日

查看全文
  • Hao Ma,
  • Hai Huan

摘要

The multimodal aspect-based sentiment analysis (MABSA) task aims to determine the sentiment polarity associated with each specific aspect term mentioned in the text by integrating multimodal information, such as textual and visual data. This task is of significant importance for deeper comprehension of user sentiments and opinions. However, existing research suffers from the following limitation: (1) Although a correspondence exists between visual regions and words, current methods struggle to accurately learn the correct region-word alignment. (2) 2D images contain numerous features irrelevant to sentiment analysis, and existing methods lack an effective denoising mechanism. To tackle the mentioned issues, this paper provides a novel image-text relationship-based feature interaction network (ITRIN). Specifically, this paper introduces a cross-modal alignment correction module to learn the correspondence between region-level visual features and textual tokens and employs an adaptive gating mechanism to mitigate the adverse effects of misaligned region-word pairs. Furthermore, by incorporating global contextual information, we design a gating mechanism based on the probability scores of the image-text relationship (BRGate) in the final feature fusion layer, which enables deep filtering and effective fusion of multimodal features. The proposed model, ITRIN, achieves state-of-the-art performance on two benchmark datasets, Twitter-2015 and Twitter-2017. Extensive ablation studies further validate its superiority and effectiveness. The source code is publicly released at https://github.com/MHTransKanba/ITRIN.