Neural Computing and Applications Software
A comparative analysis of deep learning models and fine-tuned LLMs for arabic dialect region detection
Abstract
Dialect detection in the Arabic language has gained growing attention due to the linguistic diversity and complexity of Arabic across different regions. Unlike Modern Standard Arabic (MSA), which is used in formal communication, Arabic dialects are widely used in everyday speech and on social media. These dialects vary significantly in phonology, morphology, and vocabulary, often leading to challenges in natural language processing tasks. Accurate dialect identification plays a crucial role in enhancing downstream applications such as sentiment analysis, machine translation, and speech recognition. This study presents a comparative analysis of various embedding models combined with machine learning and Deep Learning, as well as different fine-tuned large language models (LLMs), for detecting region-based Arabic dialects using the MADAR and Casablanca datasets. The models were first trained and evaluated on the MADAR dataset to assess their performance. In addition, a cross-domain evaluation was conducted by training the models on MADAR and testing them on the Casablanca dataset in order to evaluate the models’ generalization capability across different dialectal datasets. We employed various categories of embedding models, ranging from lexical methods, such as TF-IDF, to semantic methods, including Word2Vec, and contextual methods, including fine-tuned transformers. Among all experiments, we found that lexical text representation works well with machine learning methods, while deep contextualized embedding works well with deep learning models. As a complete model, we found fine-tuned models produce better results than deep learning and machine learning models. These results reflect the inherent complexity of the Arabic language. In conclusion, the findings suggest that, despite the complexity of Arabic dialects, relatively simple deep learning architectures with effective embedding models can perform competitively with large, fine-tuned language models. This suggests that model simplicity, when paired with strong linguistic representation such as MarBERT embeddings, can yield high performance in dialect detection tasks. Therefore, efficient and well-designed architectures can serve as viable alternatives to more computationally expensive models for Arabic dialect identification.