Leveraging Pre-trained Embeddings in an Ensemble Machine Learning Approach for Arabic Sentiment Analysis

Areej Jaber^1*Israa Bahati¹Paloma Martinez²

• ¹Palestine technical university - Kadoorie, Tulkarm, Palestine
• ²Universidad Carlos III de Madrid, Getafe, Spain

Arabic sentiment analysis presents unique challenges due to the linguistic complexity of Arabic, including its wide range of dialects, orthographic ambiguity, and limited language resources. This study explores the application of ensemble machine learning methods to improve sentiment classification performance in Arabic. Several homogeneous ensemble techniques are implemented and evaluated using two datasets: the balanced ArTwitter dataset and the highly imbalanced Syria_Tweets dataset. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is applied. The models incorporate pre-trained word embeddings and unigram features. Experimental results demonstrate that individual classifiers using pre-trained embed-dings perform well, but ensemble models yield superior performance. On the ArTwitter dataset, the ensemble of Naive Bayes, Support Vector Machine, and Decision Tree classifiers achieves an accuracy of 90.22% and 92.0% F1-score, while on the Syria_Tweets dataset, an ensemble combining Stochastic Gradient Descent, k-Nearest Neighbors, and Random Forest reaches 83.82% accuracy and 83.86% F1-score. These findings highlight the effectiveness of ensemble learning in enhancing the robustness and generalizability of Arabic sentiment analysis systems when supported by pre-trained embedding representations.