From Evaluation to Feedback: A Feature-Based and LLM-Constrained Tool for Korean Writing Assessment

초록

Integrating LLMs with automated writing evaluation can transform scoring models into feedback tutors that translate diagnostic signals into actionable guidance. Ensuring reliable and consistent results in such systems requires addressing information overload, yet they still rely on manual intervention to extract key signals for meaningful feedback, limiting full automation. To address these challenges, we propose FEAK, a pipeline for Korean writing that dynamically identifies rubric-linked linguistic features with low values and uses them as evidence for LLM-based feedback generation. By grounding generative feedback in measurable diagnostic signals, FEAK generates interpretable and verifiable feedback. This design bridges quantitative evaluation with pedagogically meaningful guidance while operating autonomously. Experiments with automatic and human evaluations show that FEAK generates real-time feedback comparable to expert guidance and outperforms general LLM-based systems in quality and reliability. © 2026 Copyright held by the owner/author(s).

키워드