The success of an organization hinges on the opinions of its customers regarding its online ordering and delivery systems. The English-language literature has heavily concentrated research on sentiment analysis (SA). Despite the increasing use of Arabic as a web writer, there is a dearth of Arabic datasets and lexicons, and research into Arabic sentiment analysis has been limited. This study applies deep emotion mining to Arabic reviews. Our use of RNN, mBERT, RoBERTa, and AfriBERT demonstrates that the suggested method is useful in evaluating FDS sentiment. The study used oversampling approaches such as Generative Pre-trained Transformer to categorize emotions from an unbalanced dataset. When it comes to emotion categorization, we measure things like accuracy, F1-score, recall, and precision. The analysis compares methods with and without dataset preparation. Extensive experiments show that the framework works as intended and that transformer models outperform the norm in deep learning. The AfriBERTa model achieves better results than baseline norms, with an accuracy of 0.919 and an F1 score of 0.92.