Hybrid Progressive Fusion Network for Multimodal Sentiment Analysis

Multimodal sentiment analysis (MSA) aims to infer human sentiments from heterogeneous modalities. Existing methods often rely on single architectures, such as Transformer, to model long-range dependencies, but struggle to achieve progressive multimodal fusion. To address these issues, we propose the Hybrid Progressive Fusion Network (HPFN) for MSA. HPFN integrates the advantages of CNN, Mamba, and Transformer networks: CNN extracts modality-specific local features, Mamba efficiently models global dependencies between modality pairs, and Transformer enables complex interactions across all modalities. Compared with Transformer-based models, HPFN provides a more comprehensive representation of both local details and global dependencies. We evaluate HPFN on the MOSI and MOSEI datasets, demonstrating its effectiveness in improving efficiency and overall performance. Our code is available at https://github.com/Anthony1013/HPFN.