Multi-Label Classification of Qur'anic Similes: A Computational Approach to Arabic Rhetorical Theory

Abstract

Similes (tashbih) are a cornerstone of Qur'anic eloquence and a cognitive tool for conveying complex theological concepts through accessible imagery. Despite the sophisticated taxonomies established by classical Arabic rhetoricians, computational approaches often rely on single-label classification, which fails to represent rhetorical overlap where a verse may embody multiple non-mutually exclusive categories. This study formalized Qur'anic simile classification as a multi-label learning task, bridging classical linguistic theory and modern natural language processing. Using an expert-annotated dataset of 364 verses grounded in authoritative classical exegeses, the study evaluated Arabic-specific Transformer models including AraBERT, CamelBERT, and MARBERT. MARBERT achieved the best performance, with a Micro F1 score of 0.7685 and a Macro F1 score of 0.6003, significantly outperforming traditional statistical baselines. The corpus showed high label density, empirically supporting the synergistic and multidimensional nature of Qur'anic rhetorical figures. The study contributes to computational hermeneutics by showing that classical rhetorical categories can function as structured and learnable knowledge, and it offers a scalable framework for automated analysis of sophisticated Classical Arabic texts.