Scholarlink

With the rapid growth of social media, multimodal sentiment analysis (MSA) has emerged as a critical research area. However, most existing approaches rely on direct fusion strategies that fail to capture fine-grained cross-modal interactions and often disregard the dynamic, context-dependent nature of emotional expression. Additionally, the varying reliability across modalities is frequently overlooked, leading to noisy and suboptimal representations. To address these challenges, we propose CADA, a novel fusion framework inspired by the Valence-Arousal model of emotion. CADA disentangles sentiment into two complementary dimensions: Sentiment Activation (intensity) and Sentiment Orientation (polarity). It leverages a Perceiver Compressor for modality-specific compression and a Multimodal Confidence Processor (MCP) to evaluate and weight modality reliability. Instead of flat fusion, CADA adopts a hierarchical modeling strategy where audio-text interactions capture intensity and text-visual interactions capture polarity, further modulated dynamically to ensure context-aware sentiment understanding. Experiments on CMU-MOSI and CMU-MOSEI demonstrate that CADA achieves state-of-the-art performance across multiple benchmarks.

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

CADA: A Confidence-Aware Dual-Aspect Fusion Network for Multimodal Sentiment Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tianyi Li,
  • Daming Liu

摘要

With the rapid growth of social media, multimodal sentiment analysis (MSA) has emerged as a critical research area. However, most existing approaches rely on direct fusion strategies that fail to capture fine-grained cross-modal interactions and often disregard the dynamic, context-dependent nature of emotional expression. Additionally, the varying reliability across modalities is frequently overlooked, leading to noisy and suboptimal representations. To address these challenges, we propose CADA, a novel fusion framework inspired by the Valence-Arousal model of emotion. CADA disentangles sentiment into two complementary dimensions: Sentiment Activation (intensity) and Sentiment Orientation (polarity). It leverages a Perceiver Compressor for modality-specific compression and a Multimodal Confidence Processor (MCP) to evaluate and weight modality reliability. Instead of flat fusion, CADA adopts a hierarchical modeling strategy where audio-text interactions capture intensity and text-visual interactions capture polarity, further modulated dynamically to ensure context-aware sentiment understanding. Experiments on CMU-MOSI and CMU-MOSEI demonstrate that CADA achieves state-of-the-art performance across multiple benchmarks.