Segregate, Refine, Integrate: Decomposing Multimodal Fusion for Sentiment Analysis

arXiv cs.CL · 18 hours ago

Researchers proposed SeRIn, a multimodal fusion architecture that separates the refinement of individual modality representations from cross-modal interactions using isolated pathways and a dedicated integration step. The method achieved state-of-the-art results on CH-SIMS and CMU-MOSEI benchmarks, with improvements across all metrics on both datasets. This architectural approach enables more effective sentiment analysis by preventing modality-specific signals from being contaminated during the fusion process.