Comparative evaluation of domain-specific and general-purpose transformer models for Arabic poet classification.

Arabic poet classification presents distinctive challenges stemming from the morphological richness and stylistic diversity inherent in both classical and modern Arabic verse. This study conducts an extensive comparative evaluation of several neural language models to assess their ability to represent poetic expression and capture authorial characteristics. Two carefully curated datasets are utilised: FrequentPoets, representing prolific authors with extensive verse collections, and CrossEraPoets, encompassing poets from distinct historical periods to examine temporal stylistic variation. A comparative evaluation framework is introduced to contrast domain-specific and general-purpose language models across prolific authorship and cross-era stylistic variation. The domain-adapted AraPoemBERT consistently achieves superior performance, attaining 73.11% accuracy (73.00% F1) on FrequentPoets and 77.06% accuracy (77.04% F1) on CrossEraPoets, whereas the general-purpose GPT-4o demonstrates considerably lower performance under zero-shot and few-shot evaluation settings. The results highlight the significance of domain-adapted pretraining for morphologically complex languages like Arabic and suggest the potential advantage of transformer-based architectures in modelling stylistic and linguistic nuances unique to Arabic verse. These findings also suggest that temporal diversity may play an important role in model generalisation across different poetic styles. The study contributes to Arabic Natural Language Processing (NLP) and digital humanities by enabling computational authorship attribution, stylistic analysis, and cross-historical exploration of Arabic literary heritage. Overall, the proposed framework provides a robust foundation for future research in Arabic poetry analytics and domain-specific language modelling.