earticle

영어

This study examines whether ChatGPT-based text augmentation can preserve learner writing levels and how decoding temperature influences augmentation quality and stability. We augmented 141 English texts written by Korean EFL university students (CEFR B1–B2) under five temperature settings (0.3, 0.5, 0.7, 0.8, 0.9), producing 695 candidate texts. Augmentation quality was assessed using a composite hybrid score integrating level preservation, lexical diversity, and structural stability. For each source text, the highest-scoring candidate was selected, yielding 139 original–augmented text pairs for subsequent analyses. Results show that the augmented texts achieved partial level preservation: they largely maintained relative proficiency ordering and macro-level text scale (e.g., word and sentence counts), while permitting limited shifts in local lexical and syntactic features. Across conditions, temperature 0.5 showed the highest selection frequency and the lowest variability, indicating the most consistent balance between expressive diversity and structural stability for level-preserving augmentation.