A Study on the Classification of the Semantic Functions of the Korean Postposition -ey: Comparing Generative AI Prompting Conditions

This study investigates the automatic classification of the semantic functions of the Korean postposition -ey using the generative AI model GPT-4o and examines how different prompting strategies affect classification performance. Drawing on the APIK corpus, 300 sentences representing six semantic functions—Locative(LOC), Goal(GOL), Effector(EFF), Criterion (CRT), Theme(THM), and Instrument(INS)—were analyzed under three prompting conditions: Zero-shot, Few-shot, and Chain-of-Thought(CoT). The results show that Few-shot prompting achieved the highest overall performance, whereas CoT did not consistently improve classification accuracy, despite its usefulness for examining the model’s reasoning processes and error patterns. The error analysis revealed three major tendencies: overgeneralization toward LOC, boundary confusion between GOL and LOC, and dispersed misclassification of INS. These findings suggest that LLM-based prompting can serve as a flexible and low-cost tool for exploring polysemy classification, although it cannot fully replace task-specific fine-tuning approaches. The study also provides empirical evidence relevant to understanding acquisition difficulty and informing the teaching of -ey in Korean language education.