Arabic Sentiment Analysis based Deep learning

Abstract

Arabic sentiment analysis is an offshoot of natural language processing (NLP) involved in classifying Arabic texts according to emotional content. Various factors hamper the successful attainment of high accuracy in sentiment classification of Arabic texts, such as rich morphology, different dialects, and poor availability of reliable annotated data. In this work, we address some of these issues by assessing and comparing different models dealing with Arabic sentiment analysis over the ar-reviews-100k dataset. Our approach involves classical models, deep learning models (LSTM, BiLSTM, GRU, and BiGRU), and transformer-based models, notably AraBERT. The experimental results show that deep learning models prove to be more competent than classical machine implementations. Considering these models, AraBERT outperformed the others in terms of accuracy, achieving a score of 74.5%. This indicates that transformer architectures could be strong candidates for Arabic sentiment analysis and potentially applicable in real-life situations, such as social media monitoring, customer feedback analysis, and market research.