프롬프트 엔지니어링과 파인튜닝을 활용한 대규모 언어 모델 기반 한국어 혐오 표현 다중 레이블 분류 연구

초록

The proliferation of hate speech in online spaces has emerged as a critical socio-cultural challenge, marked by rapid dissemination and continuous reproduction driven by economic incentives. Existing approaches to hate speech detection have predominantly relied on rule-based keyword filtering in traditional machine learning or deep learning models trained mainly on English datasets, leading to degraded performance in Korean contexts and limited capability for fine-grained, multi-label classification. To address these limitations, this study proposes a large language model (LLM)-based multi-label classification framework that captures the complex semantics and contextual nuances of Korean hate speech. Specifically, the publicly available Kor_Unsmile dataset was reformatted into a Chain-of-Thought (CoT) prompt structure and used to fine-tune a large-scale language model. During model development, we designed the LlamaGuard-3-8B model by integrating CoT prompt engineering with Low-Rank Adaptation (LoRA) fine-tuning, and applied systematic hyperparameter optimization to maximize performance. Experimental results show that the proposed model achieved a Macro F1 score of 0.8623 and a Micro F1 score of 0.8817, outperforming GPT-4.0 and KoELECTRA. A human evaluation alignment analysis further confirmed a high agreement rate of 90.1%, validating the model’s practical applicability. This study not only offers a practical solution for multi-label classification of Korean hate speech on online platforms but also contributes to the literature by presenting a systematic LLM-based classification model development process grounded in a design science methodology.

키워드