Dominant-Aware Dynamic Routing and Multi-Center Enhancement Network for Multimodal Sentiment Analysis
Purpose: Multimodal sentiment analysis (MSA) aims to integrate language, vision, and audio for precise emotion understanding. However, modality contribution varies dynamically across samples, and existing methods either predefine a fixed dominant modality or adopt static fusion rules, failing to reconstruct the fusion architecture according to sample-specific characteristics. Methods: We propose the DAMME network (Dominant-Aware Multi-center Modality Enhancement Network), comprising four components: (1) an emotion-ratio-based dominant modality identification mechanism; (2) a feature disentanglement module decomposing multimodal information into shared and specific subspaces; (3) multi-center enhancers TFM, VFM, and AFM for text-, vision-, and audio-dominated scenarios; (4) a dynamic routing module that activates the corresponding enhancer according to the identified dominant modality during training. Results: Experiments on CMU-MOSI, CMU-MOSEI, CH-SIMS, and CH-SIMSv2 demonstrate that DAMME achieves state-of-the-art performance, outperforming GRCF on 7-class classification (50.05\% vs. 49.3\% on MOSI) and surpassing MSLMamba across all metrics, while maintaining a lightweight architecture without external LLM supervision. Conclusion: The proposed dynamic architecture reconstruction paradigm offers a new research direction for multimodal sentiment analysis.
Authors
- Yajie Guo (ORCID: https://orcid.org/0000-0003-1529-7110)
- Dangguo Shao (ORCID: https://orcid.org/0009-0001-0432-4219)
- Sanli Yi (ORCID: https://orcid.org/0000-0001-5570-4927)
- Lei Ma
Institutions
- Kunming University of Science and Technology (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-18
- DOI
- https://doi.org/10.21203/rs.3.rs-11028525/v1
- Primary Topic
- Sentiment Analysis and Opinion Mining
- Type
- preprint