GCMA-Net: A Gated Cross-Modal Attention Network for Arabic Multimodal Sentiment Analysis

Arabic multimodal sentiment analysis (MSA) remains challenging due to linguistic diversity, modality imbalance, and the limitations of traditional fusion techniques in capturing cross-modal interactions. This study proposes GCMA-Net, a Gated Cross-Modal Attention Network designed to model fine-grained relationships across text, audio, and visual features. The model employs modality-specific projection, cross-modal attention, and gated fusion to generate a unified and contextually enriched latent representation. Experiments on the Ar-MuSA dataset show that GCMA-Net significantly outperforms existing multimodal fusion-based baselines as well as the UniTextFusion state-of-the-art method, achieving 80.17% accuracy and an F1-score of 0.8015.The proposed approach is compared against widely used multimodal fusion baseline models, and the improvements are confirmed to be statistically significant.These results demonstrate the effectiveness of interaction-centric fusion architectures in strengthening multimodal sentiment prediction. The study’s implications highlight the importance of richer latent spaces, more expressive cross-modal alignment mechanisms, and the need to expand multimodal Arabic resources to develop more robust, generalizable sentiment analysis models.