Multimodal Sentiment Analysis Based on Multi-View Attention and Cross-Modal Contrastive Learning

Multimodal Sentiment Analysis (MSA) leverages text, visual, and auditory data to detect sentiments. Existing methods often overlook explicit textual sentiment cues and false sentiments. We propose MAFM, integrating Multi-view Attention Fusion (MAF) to reduce cross-modal noise and Cross-modal Distance-aware Contrastive Learning (CDCL) to enhance intra-modal feature learning. Experiments on CMU-MOSI and CMU-MOSEI show MAFM outperforms state-of-the-art approaches.