Graph embeddings vs. Contextual embeddings: An efficient approach to Arabic sentiment analysis for improving decision making in business intelligence

The analyze of people’s opinions, feelings, and attitudes has a significant impact across various application domains, such as business intelligence, where it helps improve company services, enhance marketing strategies, and support better decision making. In recent years, sentiment analysis in the Arabic language has become increasingly challenging due to its status as one of the most widely spoken languages and its complex linguistic characteristics.In this research, we propose a novel comparative study of two efficient feature extraction approaches applied to an Arabic dataset, which include the graph embedding-based approach utilizing DeepWalk and Node2Vec techniques and the contextual embedding-based approach leveraging AraBERT. The embeddings generated by these methods are trained using various machine learning and deep learning algorithms, such as Random Forest (RF), Support Vector Machines (SVM), stacking ensemble classifiers (EC), Bidirectional Long Short-Term Memory (Bi-LSTM), Recurrent Neural Networks (RNN), and Multi-Layer Perceptron (MLP).Additionally, the AraBERT model is employed as a classifier. The performance of the models is evaluated using multiple metrics, including accuracy, precision, recall, F1-score, ROC-AUC curve, Cohen's kappa, and a novel metric assessing the time required to generate embeddings. The results indicate that the contextual embeddings approach, particularly when combined with the AraBERT classifier and Bi-LSTM, demonstrates superior performance across all evaluation metrics. This highlights the effectiveness of contextual embeddings in capturing semantic relationships between words, outperforming the graph embedding approach.