AraRoBERTa

A. Alqahtani, C. P. Lee, K. M. Lim, A. Alsharafi, M. Alzahrani, E. Alqaysi, W. Alsarhani, Khalid Alharthi

Abstract: This paper presents a study on Arabic Sentiment Analysis using AraRoBERTa, a Transformer-based architecture optimized for Arabic text. AraRoBERTa leverages the capabilities of RoBERTa, combined with advanced preprocessing techniques, to handle the unique linguistic challenges posed by the Arabic language, including its rich morphology and diverse dialects. The model was evaluated on two benchmark datasets: the Arabic Sentiment Analysis Dataset - SS2030 and the Arabic Sentiment Tweets Dataset (ASTD). AraRoBERTa outperformed existing approaches, achieving an accuracy of 0.91 on SS2030 and 0.70 on ASTD, surpassing both traditional machine learning methods and prior deep learning models. The results highlight the model's ability to capture deep contextual relationships and adapt to diverse sentiment-rich contexts, setting a new benchmark for Arabic sentiment classification.

External IDs:doi:10.1109/icnlp65360.2025.11108553