MSAD-Moroccan dialect: a multimodal sentiment analysis dataset for Moroccan dialect (Darija)

Abstract

Sentiment analysis for Arabic dialects remains a significant challenge, mainly due to the scarcity of annotated resources and the inherently multimodal nature of real-world human communication. Among Arabic dialects, Moroccan Dar ija is particularly underrepresented in the literature, as most existing datasets focus exclusively on textual modality. To bridge this gap, we present mul timodal sentiment analysis dataset (MSAD)–Moroccan dialect, a multimodal dataset built from authentic Moroccan podcast videos, integrating synchronized textual, acoustic, and visual representations annotated across three sentiment classes: positive, negative, and neutral. To evaluate the dataset, we conducted baseline experiments using unimodal, bimodal, and trimodal configurations with support vector machines (SVM) and multi-layer perceptrons (MLP) classifiers under various fusion strategies. Our results consistently show that multimodal configurations outperform their unimodal counterparts. Notably, the trimodal setup achieved the best overall performance, reaching an accuracy of 0.6176 and a weighted F1-score of 0.6019 using an MLP classifier with concatenation based fusion. These findings highlight the effectiveness of multimodal learning for sentiment understanding in conversational contexts, and position MSAD as a new benchmark for sentiment analysis in low-resource Arabic dialects.