Akademik Veri Yönetim Sistemi
ACM Transactions on Asian and Low-Resource Language Information Processing, cilt.25, sa.4, 2026 (SCI-Expanded, Scopus)
This study aims to enhance Arabic Sentiment Analysis (ASA) by developing and evaluating a hybrid Deep Learning (DL) model, AraBERT_CNN_MHA_BiLSTM, which integrates Arabic Bidirectional Encoder Representations from Transformers (AraBERT) with Convolutional Neural Networks (CNN), Multi-Head Attention (MHA), and Bidirectional Long Short-Term Memory (BiLSTM) for Sentiment Analysis (SA) in the Iraqi Arabic dialect. Through an ablation study, we systematically assess the contribution of each component by testing the model on two datasets: “IQAD31K,” comprising 31,324 reviews scraped from Google Play Store, and the Iraqi Arabic Dialect (IAD) dataset, comprising 2,000 annotated comments from Iraqi Facebook pages. The complete model achieves an F1-score of 95.63% on “IQAD31K” and 94.50% on the IAD dataset, exceeding traditional Machine Learning (ML) methods such as Support Vector Machine (SVM) and Random Forest (RF), which struggle with Arabic linguistic complications. These findings highlight the importance of integrating AraBERT's contextual embeddings, CNN's local feature extraction, MHA's long-range dependency capture, and BiLSTM's sequential modeling to achieve the best performance in dialectal Arabic sentiment analysis, thereby contributing a robust architectural framework for future research in dialectal Arabic Natural Language Processing (ANLP).