Kuwaiti-Arabic Sentiment Analysis: A Comparative Study

This study introduces a novel approach to the flourishing field of Arabic natural language processing (NLP) within the Arab region. Arabic NLP is increasingly vital for text sentiment analysis, supporting feedback collection for businesses, institutions, governments, and more. This paper explores a two-step Arabic-Kuwaiti sentiment classification model. Our design utilizes an initial model to distinguish Modern Standard Arabic (MSA) from the Kuwaiti dialect. Sentences classified as Kuwaiti are then passed to a subsequent model for sentiment classification. We detail the pre-processing techniques employed, emphasizing the need for data-specific adjustments for each model. Addressing the scarcity of online Kuwaiti dialect datasets, we compiled a custom dataset from scratch. We investigated various approaches, testing traditional machine learning models and experimenting with deep learning, specifically transfer learning, through fine-tuning BERT models. All models demonstrated strong performance, achieving varying accuracies. Notably, MARBERT achieved the highest performance, exceeding 98% accuracy for MSA vs. Kuwaiti dialect classification and 85% accuracy for Kuwaiti sentiment analysis.