Haipeng Chen

Discriminative semantic learning for incomplete multimodal sentiment analysis

The framework first introduces learnable sentiment prototypes as semantic anchors to provide explicit sentiment-discriminative guidance for feature completion, and a gradient decoupling strategy is designed to separate the optimization paths of unimodal and multimodal objectives, preventing fusion gradients from interfering with unimodal encoders, thereby synergistically enhancing both discriminative representation learning and multimodal fusion.