The Multimodal Sentiment Analysis (MSA) land-scape for Arabic content is strikingly underexplored, mainly due to limited datasets and a lack of robust integration methods across text, audio, and image. While transformer-based models like MarBERT and ArBERT achieve strong results on Arabic text, most research remains unimodal and does not fully exploit multimodal synergy. In this work, we propose a three-fold approach for Arabic MSA. First, we finetune robust transformers for each modality, namely ViT, MarBERT, and HuBert for image, Text, and Audio, respectively. Second, we perform an early feature fusion. Third, we use classifiers for sentiment prediction. On the recent Ar-MuSA benchmark released on 2025, our tri-modal fusion system, achieves state-of-the-art performance (F1=0.7756, Accuracy=0.7759), significantly exceeding the multimodal models benchmarked on the Ar-MuSa dataset, as well as the unimodal and bimodal methods. This demonstrates that comprehensive tri-modal fusion and thoughtful classifier selection are essential for accurate, human-centric Arabic sentiment analysis.
Cheikhi et al. (Thu,) studied this question.