Boosting Arabic Emotion Detection: Transformer Ensembles with Soft Voting

Abstract

Arabic Emotion Detection is essential for contexts that involve Arabic text sentiment and emotional intent, such as social media monitoring, psychological evaluation, or customer feedback. However, challenges in the Arabic language, such as deep morphology, varied dialects, and limited annotated datasets, make this task extremely challenging. Most existing studies have used classical classifiers and individual transformers such as AraBERT, MARBERT, and CamelBERT to mitigate these problems. This paper proposes an ensemble learning method using soft voting to improve Arabic emotion detection performance. Implemented with benchmarked datasets, the method proposed in this paper is ultimately more accurate at completing the task, achieving the highest accuracy of 76.32% compared to baseline models such as the Complement Naïve Bayes of accuracy 68.12%.