Progressive fusion and guidance-aware denoising for robust multimodal sentiment analysis

With the rapid growth of digital media, multimodal sentiment analysis has attracted increasing research attention. However, existing approaches largely emphasize fusion strategies while overlooking the noise and redundancy inherent in real-world multimodal data, which limits the representational capacity of models. To address this issue, we propose a Progressive Fusion and Guidance-Aware Denoising Framework (PFGAD). It first learns joint multimodal representations using progressive fusion. Next, it applies a variational information bottleneck to remove noise that is irrelevant to the task. Finally, it uses a Guidance-Aware Denoising Module to make up for any information lost during denoising. This design achieves a good balance between informativeness and robustness. Our method demonstrates notable improvements on CMU-MOSI and CMU-MOSEI. On CMU-MOSI, it achieves Corr (0.814) and Acc-2 (87.5), which corresponds to a 1.75% relative gain in Corr and a 0.5%-point increase in Acc-2 over the strongest prior result. On CMU-MOSEI, it achieves Corr (0.786) and F1 (86.7), yielding a 0.77% relative gain in Corr and a 0.2%-point improvement in F1.