Sentiment analysis in Arabic faces unique challenges due to the language’s rich inflection and dialectal variations. This paper investigates how the composition and size of training datasets impact sentiment analysis performance for Arabic. We evaluate the effects of dataset size and preprocessing techniques (specifically stop-word removal and stemming) on a Naïve Bayes classifier’s accuracy, precision, and recall. An Arabic Twitter corpus of 150k tweets (balanced positive and negative) is used, with sentiment labels obtained via automated annotation. We experiment with incremental training set sizes and different preprocessing configurations. Results show that larger training datasets significantly improve classification performance up to a point, after which gains plateau. Moreover, appropriate preprocessing—especially light stemming and stop-word removal—yields notable improvements in accuracy and F1-score. The findings underscore the importance of dataset quality and preprocessing for sentiment analysis in highly inflected languages. Future work will explore advanced models and richer linguistic processing to further enhance Arabic sentiment analysis.