EmoDuo: A Dual-Tower Framework for Efficient and Robust Multimodal Sentiment Analysis

Multimodal Sentiment Analysis (MSA) has emerged as a critical research area. Traditional methods often face challenges in inter-modal feature learning and high computational cost. This paper proposes EmoDuo: Emotion analysis via Dual Tower Attention Mechanism, a novel framework for MSA that effectively addresses these challenges. EmoDuo integrates robust representation learning using RoBERTa for text and self-attention encoders for acoustic and visual modalities. The framework introduces a Cross-Attention and Residual (CAR) Block to effectively model intra- and inter-modal interactions and employs a dual-tower structure to separately predict sentiment polarity and intensity. Evaluations on CMU-MOSI and CMU-MOSEI show that EmoDuo delivers strong performance with reduced computational overhead, offering promising directions for future research in multimodal sentiment understanding.