Analysis of Libyan Dialect Sentiment Using AraBERT and MARBERT: A Comparative Study

Abstract

Because of the wide range of spoken dialects, the Libyan dialect is regarded as an uncommon digital resource and presents significant challenges for Arabic natural language processing. By providing an experimental comparative analysis of the capability of two cutting-edge language models based on this gap, this study seeks to address and evaluate it. Trained in the Libyan dialect to use the transformer architectures AraBERTvs and MARBERTv2 for sentiment analysis. A new dataset of 500 comments collected from social media sites was created and categorized. Both models were fine-tuned using the same parameters and evaluated on a specially designed test set. The AraBERTv2 model achieved 71. 0% accuracy, compared to the MARBERTv2 model's 70. 0%, demonstrating a slight advantage in the results. Notably, a study of training practices revealed that MARBERTv2 had early over fitting, whereas AraBERTv2 exhibited superior stability and enhanced learning ability. When applied to resource-constrained local vernaculars, the research demonstrates that models trained on various Modern Standard Arabic, such as AraBERT, may provide a more robust and generalizable foundation.