Social media has gained substantial interest and emerged as the main platforms for sharing information and expressing opinions. Getting insight from such opinions became one major concern. In this context, sentiment analysis is one application that studies opinions and feelings, which are mainly expressed in natural language. This paper proposes a comparative study of sentiment analysis of Arabic texts that are expressed in both modern standard Arabic and Moroccan dialect. Our methodology encapsulates four different approaches, namely, lexicon-based, Machine learning using two ensemble learning methods: Random Forest and XGBoost, a deep learning model based on the Bi-LSTM network and a large language model using BERT. This comprehensive methodology has several objectives. First, to illustrate the role of the preprocessing tasks adopted in each approach to address linguistic complexities. Second, to highlight the steps of feature representation, which is a crucial transitional step from unstructured text to structured data. Finally, to evaluate the performance of each model in sentiment classification based on the same datasets. The experimental results demonstrate the superiority of the advanced models, BERT and BI-LSTM compared to the classic Machine learning and lexicon methods.