저자원 텍스트분류의 성능 향상을 위한 KoBERT기반 카테고리 매핑과 LLM 결합 연구

초록

Text classification tasks with insufficient training data remain a challenging problem. To address this, we propose a method that combines KoBERT-based fine-tuning with Large Language Model(LLM)-based zero-shot classification. Focusing on the Science and Technology Standard Classification and Future Emerging Technologies(6T) systems, we fine-tuned KoBERT to implement a model for the Science and Technology Standard Classification and established a mapping strategy between the two systems. For subcategories where mapping was not feasible, zero-shot classification was applied. We also employed explanation-based and verification prompts to enhance the reliability of classification results. Experimental results confirmed that the proposed method can be highly effective in scenarios with limited training data. This provides an approach for addressing classification tasks with insufficient data.

키워드