Transformer-based ensemble model for dialectal Arabic sentiment classification.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Social Media Analysis

Background:

• Social media is vital for public opinion, especially during global emergencies.
• Analyzing public sentiment in dialectal Arabic is challenging due to linguistic complexity.
• Sentiment analysis, or opinion mining, is key for understanding societal trends.

Purpose of the Study:

• To evaluate various models for sentiment analysis of dialectal Arabic tweets.
• To compare machine learning, deep learning, and transformer-based approaches.
• To identify the most effective model for Arabic sentiment classification.

Main Methods:

• Experimentation with machine learning (SVM, NB, DT, XGBoost), deep learning (CNN, BLSTM), and transformer models (CAMeLBERT, XLM-RoBERTa, MARBERT).
• Feature extraction using AraVec, FastText, AraBERT, and TF-IDF.
• Analysis of three benchmark Arabic tweet datasets: ASTD, ASAD, and TEAD.

Main Results:

• Transformer-based models, particularly an ensemble approach, demonstrated superior performance.
• The proposed transformer-based ensemble model achieved high accuracy (90.4%), recall (88%), precision (87.3%), and F1-score (87.7%).
• This indicates the effectiveness of advanced transformer architectures for dialectal Arabic sentiment analysis.

Conclusions:

• Transformer-based ensemble models are highly effective for dialectal Arabic sentiment analysis.
• The study provides a robust benchmark for future research in this domain.
• Improved sentiment analysis of Arabic social media can lead to better decision-making during crises.