Hybrid Optimizer Lifts AraBERT Accuracy in Arabic Sentiment Analysis

A study published in Cluster Computing on 22 October 2026 reports that a new hybrid metaheuristic algorithm, used to tune the hyperparameters of the pretrained Arabic BERT model known as AraBERT, produced models that outperformed the original AraBERT in both accuracy and F1-score on all three Arabic sentiment datasets tested. The work, by Abderrahmène Belhaouari, Nabil Neggaz, and Diego Oliva, also achieved the best accuracy among the compared methods on one of the three benchmarks and the second-best accuracy on the other two. The algorithm, called the hybrid Aquila Optimizer (HAO), is built by combining the Aquila Optimizer (AO), the Sine Cosine Hyperbolic Optimizer (SCHO), and opposition-based learning.

The authors describe Arabic sentiment analysis (ASA) as a demanding task given the specificity of the language and its inherent difficulty. They note that the increasing use of pretrained large language models, and their performance on other languages, makes them a viable tool for ASA, but that such models need fine-tuning in order to be applied to downstream tasks. In the experiments, the current versions of HAO, AO, and SCHO were each used to optimize eight hyperparameters of the AraBERT model.

The evaluation was carried out in two series of experiments. The second series was described by the authors as more challenging: it aimed to automatically find the optimal hyperparameters of AraBERT for the sentiment classification task. Six optimized AraBERT models were then evaluated against the base AraBERT model and against fourteen state-of-the-art approaches, using three Arabic sentiment datasets: ASTD-B, AJGT, and ArSAS.

According to the reported results, the optimized models outperformed the original AraBERT model in accuracy and F1-score on all three datasets. Compared with the broader field of fourteen state-of-the-art methods, the tuned models obtained the best accuracy on ASTD-B and the second-best accuracy on AJGT and ArSAS. The authors characterize these results as very competitive with state-of-the-art approaches.

The study is a computational evaluation of hyperparameter tuning rather than a deployment study, and its reported outcomes are limited to the three datasets, the eight optimized hyperparameters, and the comparison methods included in the experiments. Within that scope, the findings indicate that a hybrid metaheuristic combining the Aquila Optimizer, the Sine Cosine Hyperbolic Optimizer, and opposition-based learning can improve the fine-tuned performance of a pretrained Arabic language model on sentiment analysis benchmarks.

Subject of Research: Hybrid metaheuristic optimization of AraBERT hyperparameters for Arabic sentiment analysis

Article Title: Boosting aquila optimizer using sine cosine hyperbolic optimizer for tuning Bert model hyperparmeters: Arabic sentiment analysis

Article References: Belhaouari, A., Neggaz, N., & Oliva, D. (2026). Boosting aquila optimizer using sine cosine hyperbolic optimizer for tuning Bert model hyperparmeters: Arabic sentiment analysis. Cluster Computing, 29(15), Article 884. https://doi.org/10.1007/s10586-026-06650-4

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06650-4

Keywords: Arabic sentiment analysis, AraBERT, Aquila Optimizer, Sine Cosine Hyperbolic Optimizer, opposition-based learning, metaheuristics, hyperparameter tuning, fine-tuning, large language models, CEC17 benchmark, natural language processing, Cluster Computing