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

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Arabic sentiment analysis faces challenges due to linguistic diversity, dialectal variations, and limited resources.
• Developing robust sentiment classification systems requires addressing these inherent complexities.

Purpose of the Study:

• To investigate the effectiveness of ensemble machine learning methods for Arabic sentiment analysis.
• To evaluate homogeneous ensemble techniques on both balanced and imbalanced Arabic datasets.
• To assess the impact of pre-trained word embeddings and SMOTE on model performance.

Main Methods:

• Implementation and evaluation of homogeneous ensemble techniques (e.g., Naive Bayes, SVM, Decision Tree, SGD, KNN, Random Forest).
• Utilized two datasets: ArTwitter (balanced) and Syria_Tweets (imbalanced).
• Employed Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance and incorporated pre-trained word embeddings and unigram features.

Main Results:

• Ensemble models consistently outperformed individual classifiers across both datasets.
• On ArTwitter, an ensemble achieved 90.22% accuracy and 92.0% F1-score.
• On Syria_Tweets, another ensemble reached 83.82% accuracy and 83.86% F1-score.

Conclusions:

• Ensemble learning enhances the robustness and generalizability of Arabic sentiment analysis systems.
• Pre-trained embeddings further boost performance, demonstrating the value of these approaches.
• Ensemble methods effectively overcome challenges in Arabic NLP, including linguistic complexity and data imbalance.