Phantoms at BAREC Shared Task 2025: Enhancing Arabic Readability Prediction with Hybrid BERT and Linguistic Features

This paper describes our system for the BAREC 2025 Shared Task on Arabic Readability Assessment.Our approach is centered on a hybrid model that combines the deep contextual representations of a pre-trained transformer (AraBERTv02) with a rich set of engineered linguistic features.We extracted over 200 lexical, morphological, syntactic, and semantic features, which were refined to the 100 most informative ones through a multi-stage selection process.Our final model demonstrates significant effectiveness, achieving a Quadratic Weighted Kappa (QWK) of 82.7% and an exact accuracy of 57.6% on the official blind test set.These results highlight the powerful synergy between transformer-based embeddings and explicit linguistic signals for the nuanced task of assessing Arabic text readability.