Fine-tuning and Evaluation of LLaMA models for correcting particle substitution errors (centered on eun/neun (은/는), i/ga (이/가), e (에), eso (에서)) for beginning Vietnamese learners of Korean

초록

Korean grammatical particles present a persistent challenge for Vietnamese learners due to fundamentalsyntactic differences between the two languages. Vietnamese lacks case-marking particles, often leading tosubstitution errors involving eun/ neun (은/는), i/ka (이/가), e (에), and eso (에서). Traditional teachingmethods offer limited success in addressing these issues. Motivated by the need for more adaptive andlearner-sensitive solutions, this paper explores the fine-tuning of the LLaMA 3.2.1B language model to correctKorean particle substitution errors commonly made by beginner Vietnamese learners. A custom dataset wasdeveloped by generating simulated learner errors based on authentic sentence structures. The model wasfine-tuned using Low-Rank Adaptation (LoRA) and instruction-based prompts to ensure efficiency andcontextual accuracy. Evaluation on a 5,800-sentence test set demonstrated a sentence-level accuracy of91.15%, compared to just 8.36% for the pre-trained baseline. With appropriate fine-tuning, these resultsendorse the capacity of large language models for providing sound grammatical corrections that arepersonally suited to the requirements of the learners. This technology exhibits promising potential forintelligent tutoring systems in facilitating one-to-one, real-time feedback in second language learningenvironments.

키워드