Deep learning models for text classification require large, annotated datasets, which are often unavailable for many languages, including Arabic. This scarcity, combined with the linguistic complexity of Arabic, presents significant challenges, especially in sentiment and emotion analysis tasks involving short and long texts. In this paper, we evaluate several data augmentation strategies aimed at enhancing Arabic text classification performance, with a particular focus on a novel text generation approach based on fine-tuned transformer models. Our method demonstrates notable improvements in classifier accuracy for both short and long texts, achieving accuracy gains of up to 92.8% in low-resource scenarios. This comparative study highlights the effectiveness of text generation techniques over traditional augmentation methods and provides practical insights into optimizing data augmentation for Arabic NLP applications.