STBW at BAREC Shared Task 2025: AraBERT-v2 with MSE-SoftQWK Loss for Sentence-Level Arabic Readability
Year: 2025
Type: conference-paper
Abstract: Automatic Readability Assessment estimates how hard a text is for its target readers, using features such as vocabulary, spelling, morphology, etc.Based on this premise, we evaluate our experiments on Arabic language under the BAREC 2025 shared task protocol.This paper addresses the sentence-level readability assessment task with strict track, that allows only the use of BAREC train set to predict Arabic readability on a fine-grained 19-level scale.Our solution is based on a two-phase fine-tuning of AraBERT-v2 on a custom feature set of the BAREC corpus.In the blind test set, the system achieves a QWK of 85.6%.
Author: Saoussan Trigui
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037658