Advancing Arabic Dialect Detection with Hybrid Stacked Transformer Models

Hager Saleh ^1* Abdulaziz Almohimeed ² Moatamad Hassan ³ Mandour M Ibrahim ⁴ Saeed Hamood Alsamhi ⁵ Moatamad Hassan ³ Sherif Mostafa ¹

• ¹ Faculty of Computing and Artificial Intelligence, South Valley University, Hurghada, Egypt
• ² Department of Computer Science, College of Computer and Information Sciences, Imam Muhammad ibn Saud Islamic University, Riyadh, Saudi Arabia
• ³ Faculty of Science, Aswan University, Aswan, Aswan, Egypt
• ⁴ College of Computer and Information Sciences, Imam Muhammad ibn Saud Islamic University, Riyadh, Saudi Arabia
• ⁵ Department of Computer Science and Engineering, College of Informatics, Korea University, Seoul, Republic of Korea

The rapid expansion of dialectally unique Arabic material on social media and the internet highlights how important it is to categorize dialects accurately to maximize a variety of Natural Language Processing (NLP) applications. The improvement in classification performance highlights the wider variety of linguistic variables that the model can capture, providing a reliable solution for precise Arabic dialect recognition and improving the efficacy of NLP applications. Recent advances in deep learning (DL) models have shown promise in overcoming potential challenges in identifying Arabic dialects. In this paper, we propose a novel stacking model based on two transformer models, i.e., Bert-Base-Arabertv02 and Dialectal-Arabic-XLM-R-Base, to enhance the classification of dialectal Arabic. The proposed model consists of two levels, including base models and meta-learners. In the proposed model, Level 1 generates class probabilities from two transformer models for training and testing sets, which are then used in Level 2 to 1 Sample et al.