Automated text leveling for L2 English learners: a technology-enhanced framework with CEFR

Selecting appropriate texts for second language (L2) learners is critical for optimizing learning outcomes, yet existing methods often fail to accurately align materials with learner proficiency levels. This study addresses this challenge by adopting the Common European Framework of Reference for Languages (CEFR) as a benchmark for text classification. We employed ChatGPT-4o to categorize texts based on CEFR levels using a dataset of 120 texts annotated by both experts and L2 learners. Results demonstrate that a six-shot prompt strategy which provided examples from each of the six CEFR levels outperformed other few-shot prompt models, achieving an average classification accuracy exceeding 91%. This performance remained consistent across both the validation set (1,061 texts) and the testing set (120 texts). Notably, ChatGPT-4o surpassed eight widely-used readability formulas in text classification accuracy, underscoring the potential of advanced language models to support more effective text leveling for L2 education.