Attention Mechanism in Multimodal Emotion Recognition From 2020 to 2025: Technological Evolution, Challenges, and Future Prospects

Abstract

Multimodal emotion recognition, as the core technology of human-computer interaction, has made breakthrough progress in the fusion of attention mechanisms in recent years. This article systematically reviews the research context in this field from 2020 to 2025, and for the first time establishes a classification framework from the dual dimensions of "attention refinement" and "modality interaction deepening". This paper will propose a four stage evolutionary model, including the process from foundation, to optimization, to multi-scale, and finally to cross fusion. It will also present a technical matrix for innovative induction of cross modal, recursive, multi-scale, and dynamic attention mechanisms. Furthermore, it will reveal the bottleneck of feature alignment caused by modal heterogeneity. Finally, a development path will be proposed from the directions of meta learning regulation, interpretability enhancement, etc., providing theoretical reference for constructing robust multimodal systems. Overall, this review not only summarizes the current progress of multimodal attention research but also highlights future opportunities for advancing human-computer interaction, encouraging continued exploration of more intelligent, inclusive, and adaptive multimodal systems.