PalNLP at BAREC Shared Task 2025: Predicting Arabic Readability Using Ordinal Regression and K-Fold Ensemble Learning
Year: 2025
Type: conference-paper
Abstract: PalNLP addressed Arabic readability level prediction as a fine-grained ordinal classification problem by strictly using the Balanced Arabic Readability Evaluation Corpus (BAREC).The approach treats the 19-class ordinal classification problem as a regression task with posthoc threshold optimization, leveraging a BERTbased model and an ensemble strategy.The system achieved a Quadratic Weighted Kappa (QWK) score of 81.1 in the blind test dataset, indicating an almost perfect agreement between the system's classifications and the true labels, and placing 18 th out of 24 teams.The findings show that the model effectively learned broad readability patterns, with a competitive ±1 accuracy, but faced challenges in accurately predicting readability levels of most sentences.
Author: Mutaz Ayesh
Institution: Cardiff University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037656