ScholarWorks@Y-Scholar Hub

초록

This study investigates the potential and limitations of generative artificial intelligence (AI) in Korean language education by evaluating how well current models identify and correct syntactic errors made by learners. Multilingual models (ChatGPT, Claude) and Korean-specific models (EXAONE, QwenKO) were tested using a learner corpus. The results revealed notable performance differences. EXAONE showed high accuracy in error detection, highlighting the strengths of Korean-focused models, but its performance in error correction was limited and inconsistent across error types. Claude and GPT demonstrated more balanced results, particularly in rewriting corrected sentences, while QwenKO showed generally low accuracy. Overall, Korean-specific models excelled in identifying errors, whereas larger multilingual models performed better in sentence-level correction. All models struggled with errors requiring contextual or pragmatic understanding, suggesting that current AI systems still lack discourse-level competence. This study provides empirical insight into the capabilities of generative AI for syntactic error processing and offers pedagogical considerations for its informed application in Korean language instruction.