Leveraging pre-trained embeddings in an ensemble machine learning approach for Arabic sentiment analysis

Introduction Arabic sentiment analysis presents unique challenges due to the linguistic complexity of the language, including its wide range of dialects, orthographic ambiguity, and limited language resources. Addressing these issues is essential to develop robust sentiment classification systems. Methods This study investigates the application of ensemble machine learning methods for Arabic sentiment analysis. Several homogeneous ensemble techniques are implemented and evaluated on two datasets: the balanced ArTwitter dataset and the highly imbalanced Syria_Tweets dataset. To mitigate class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is employed. The models incorporate pre-trained word embeddings and unigram features. Results Experimental results indicate that individual classifiers using pre-trained embeddings achieve strong performance; however, ensemble models consistently yield superior outcomes. On the ArTwitter dataset, the ensemble of Naive Bayes, Support Vector Machine, and Decision Tree classifiers achieved an accuracy of 90.22% and an F1-score of 92.0%. On the Syria_Tweets dataset, an ensemble combining Stochastic Gradient Descent, k-Nearest Neighbors, and Random Forest attained 83.82% accuracy and an 83.86% F1-score. Discussion The findings highlight the effectiveness of ensemble learning in enhancing the robustness and generalizability of Arabic sentiment analysis systems. Incorporating pre-trained embeddings further strengthens performance, demonstrating that ensemble-based approaches can overcome challenges posed by linguistic complexity and dataset imbalance in Arabic natural language processing tasks.