The absence of face-to-face interactions between learners and instructors in online learning environments poses challenges in assessing and regulating learners’ emotions, which are crucial for optimizing learning outcomes. To address this issue, integrating emotion recognition technology into online learning platforms offers valuable insights. Textual data generated by learners, such as reviews and discussion forum posts, serve as a rich resource for emotion detection. However, these datasets often exhibit class imbalance, where certain emotions are significantly underrepresented. This study explores the effectiveness of five resampling techniques on a dataset of 401 Arabic reviews labeled with five emotions: satisfaction, enjoyment, excitement, confusion, and frustration. We applied four classifiers: Decision Tree, Logistic Regression, Random Forest, and SVM. The results indicate that oversampling techniques, particularly Random Oversampling (ROS), significantly improved the Macro F1 Score across most models, with Logistic Regression showing the highest increase from 0.48 to 0.77. A detailed analysis of individual emotion classes revealed that oversampling techniques, especially ROS, substantially enhance the recognition of minority and medium-sized classes, such as excitement and frustration. However, accurately recognizing complex emotions like confusion remains challenging, even with resampling methods.