Plagiarism detection across languages: a comprehensive study of Arabic and English-to-Arabic long documents

Plagiarism detection in Arabic texts remains a significant challenge due to the complex morphological structure, rich linguistic diversity, and scarcity of high-quality labeled datasets. This study proposes a robust framework for Arabic plagiarism detection by integrating Siamese neural networks (SNN) with state-of-the-art transformer architectures, specifically AraT5 and Longformer. The system employs a hybrid workflow, combining transformer-based encoders and a classification objective to implicitly learn textual similarity. To address the inherent imbalance in Arabic plagiarism datasets, both weighted cross-entropy loss and Dice loss functions were utilized to optimize model training. Extensive experiments were conducted using the ExAraCorpusPAN2015 dataset, demonstrating the effectiveness of the proposed architecture. Results indicate that AraT5 with weighted cross-entropy loss outperformed other configurations, achieving an F1-score of 0.9058. Additionally, comparative analysis with existing methodologies highlights the superiority of our approach in handling nuanced semantic and structural variations within Arabic texts. This study underscores the importance of transformer-based architectures and class-specific loss functions in enhancing plagiarism detection accuracy in under-resourced languages like Arabic.