Scholarlink

<p>Emotion recognition from online public opinions is essential for understanding human sentiment in diverse digital interactions. Traditional methods for multi-modal emotion recognition struggle to achieve better predictive accuracy due to modality misalignment and difficulties in obtaining important cross-modal relationships. Motivated by the hierarchical human cognitive functions of the brain, we developed TriModalEmotionNet (TME-Net) that incorporates Context-Aware Modal-Interaction Graph Attention Network (CAMIGAT) and Hierarchical Adaptive Attention Network (HAAN) for accurate emotion recognition from multi-modal data. To mitigate misalignment problems, the CAMIGAT models a heterogeneous graph that emulates the dynamic multisensory interactive mechanism of the human for understanding the cross-modal interactions between verbal (linguistic) and non-verbal (visual, auditory) signals. To resolve information redundancy, the HAAN module focuses on the intricate emotional relationship between modalities by mimicking the cognitive filtering system of the brain that regulates selective attention in human emotion perception. The TME-Net model is evaluated on three datasets across diverse metrics and compared with baseline studies. The experimental results demonstrate that TME-Net achieves impressive performance and surpasses baseline methods by achieving a higher weighted average accuracy of 97.32% on MELD, 98.37% on CMU-MOSEI, and 97.42% on IEMOCAP datasets. Moreover, the TME-Net showed lower computational complexity by obtaining a parameter of 38.8&#xa0;M, GFLOPS of 9.9, and inference time of 12.6&#xa0;ms. This makes TME-Net a suitable solution for real-time emotion recognition tasks in online public opinion and sentiment analysis.</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

A novel context-aware modal-interaction graph attention and adaptive attention approach for multi-modal emotion recognition in public opinion

  • 期刊论文

  • 发表日期：

    2026年1月26日

查看全文
  • Yuting Luo

摘要

Emotion recognition from online public opinions is essential for understanding human sentiment in diverse digital interactions. Traditional methods for multi-modal emotion recognition struggle to achieve better predictive accuracy due to modality misalignment and difficulties in obtaining important cross-modal relationships. Motivated by the hierarchical human cognitive functions of the brain, we developed TriModalEmotionNet (TME-Net) that incorporates Context-Aware Modal-Interaction Graph Attention Network (CAMIGAT) and Hierarchical Adaptive Attention Network (HAAN) for accurate emotion recognition from multi-modal data. To mitigate misalignment problems, the CAMIGAT models a heterogeneous graph that emulates the dynamic multisensory interactive mechanism of the human for understanding the cross-modal interactions between verbal (linguistic) and non-verbal (visual, auditory) signals. To resolve information redundancy, the HAAN module focuses on the intricate emotional relationship between modalities by mimicking the cognitive filtering system of the brain that regulates selective attention in human emotion perception. The TME-Net model is evaluated on three datasets across diverse metrics and compared with baseline studies. The experimental results demonstrate that TME-Net achieves impressive performance and surpasses baseline methods by achieving a higher weighted average accuracy of 97.32% on MELD, 98.37% on CMU-MOSEI, and 97.42% on IEMOCAP datasets. Moreover, the TME-Net showed lower computational complexity by obtaining a parameter of 38.8 M, GFLOPS of 9.9, and inference time of 12.6 ms. This makes TME-Net a suitable solution for real-time emotion recognition tasks in online public opinion and sentiment analysis.