LIS at BAREC Shared Task 2025: Multi-Scale Curriculum Learning for Arabic Sentence-Level Readability Assessment Using Pre-trained Language Models
Year: 2025
Type: conference-paper
Abstract: Sentence-level readability assessment, which measures how easily individual sentences can be understood, has seen significant advances in English.However, Arabic readability assessment remains underexplored, primarily due to the language's morphological complexity and the scarcity of fine-grained annotated datasets.To address this gap, we leveraged the BAREC corpus, which provides 69K sentences annotated across 19 readability levels, enabling us to develop and compare five different modeling strategies ranging from lightweight classifiers to fine-tuned Arabic language models.Our experiments revealed that task-specific pretraining with CamelBERT yielded substantial performance gains, while curriculum learning offered benefits in specific scenarios.Ultimately, direct fine-tuning achieved state-of-the-art performance (QWK = 82.4).Through detailed error analysis, we identified that models struggled most with distinguishing between the lower readability level 2 and higher readability levels (15-19), highlighting the inherent challenges in fine-grained Arabic readability modeling across the full spectrum of proficiency levels.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: hybrid
OpenAlex ID: W4416037811