생성형 AI의 한국어 시제 오류 주석 가능성과 한계: 고급 한국어 학습자 오류 주석 말뭉치를 중심으로

초록

This study investigates the capabilities and constraints of generative artificial intelligence in detecting and annotating Korean tense errors by comparing AI-generated error annotations with those produced by humans. Specifically, 589 tenseerror tokens from advanced Korean learners were annotated using ChatGPT-5.2. The analysis examined not only the concordance between AI and human annotations but also the accuracy of AI-generated corrections, pinpointing of error locations, and identification of error types. The findings revealed that 55.9% of the AI annotations were considered accurate, highlighting the potential of generative AI for annotating Korean tense errors. However, the AI often produced overcorrections, localized errors according to the corrected forms rather than the original error sites, and demonstrated inconsistent performance in distinguishing among omission, addition, and substitution errors. These results suggest that AIbased annotation requires more precise annotation guidelines, richer discourse context, and systematic attention to grammatical and semantic constraints on Korean tense forms. The study advances understanding of both the usefulness and boundaries of generative AI in annotating Korean learner errors by providing detailed analyses of correction accuracy, error location identification, and error type categorization.

키워드