Scholarlink

<p>The widespread occurrence of hateful messages online poses a significant societal challenge, but the emergence of Hateful Memes presents a formidable problem due to the implicit hate conveyed through complex visual-textual interplay. To overcome the limitations of traditional methods that struggle with semantic incongruities and result in high false positives, we introduce the novel, end-to-end Contextual Multi-Task Attention (CMTA) Framework. This architecture integrates a dynamic Cross-Modal Attention (CMA) mechanism for bi-directional feature fusion and a comprehensive Multi-Task Learning (MTL) paradigm, which leverages auxiliary tasks (Emotion Detection and Sentiment Analysis) to provide contextual enhancement and implicit regularization. We train and evaluate the framework using a combined dataset from Met-Meme and Memotion, and the results demonstrate that this holistic approach significantly enriches contextual understanding and achieves state-of-the-art performance. Specifically, the CMTA framework achieved F1-scores of 0.8864 on Met-Meme and 0.9043 on Memotion for the primary offensiveness task. This success, attributed to the CMA resolving semantic incongruities and MTL reducing non-offensive misclassification, provides a robust solution for combating implicit online hate.</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

CMTA-meme: dynamic cross-modal attention and multi-task learning for resolving implicit hate in memes

  • 期刊论文

  • 发表日期：

    2026年4月25日

查看全文
  • Hafiz Muhammad Arslan,
  • Tan Zhenhua,
  • Mubashar Mustafa

摘要

The widespread occurrence of hateful messages online poses a significant societal challenge, but the emergence of Hateful Memes presents a formidable problem due to the implicit hate conveyed through complex visual-textual interplay. To overcome the limitations of traditional methods that struggle with semantic incongruities and result in high false positives, we introduce the novel, end-to-end Contextual Multi-Task Attention (CMTA) Framework. This architecture integrates a dynamic Cross-Modal Attention (CMA) mechanism for bi-directional feature fusion and a comprehensive Multi-Task Learning (MTL) paradigm, which leverages auxiliary tasks (Emotion Detection and Sentiment Analysis) to provide contextual enhancement and implicit regularization. We train and evaluate the framework using a combined dataset from Met-Meme and Memotion, and the results demonstrate that this holistic approach significantly enriches contextual understanding and achieves state-of-the-art performance. Specifically, the CMTA framework achieved F1-scores of 0.8864 on Met-Meme and 0.9043 on Memotion for the primary offensiveness task. This success, attributed to the CMA resolving semantic incongruities and MTL reducing non-offensive misclassification, provides a robust solution for combating implicit online hate.