다중 필터링 기반 지식 그래프 트리플 주입을 통한 한국어 의료 LLM 추론 성능 향상

초록

While combining knowledge graphs (KGs) with large language models (LLMs) has proven effective for medical question answering, irrelevant triples injected from KGs can act as noise, degrading overall performance. To address this, we retrieve triples from UMLS, a biomedical knowledge base, and apply a three-stage filtering pipeline consisting of semantic similarity filtering, confidence filtering, and LLM Relevance Filtering to select only high-quality triples. In particular, we introduce a dual-query construction strategy that formulates queries differently depending on the source of each triple, enabling confidence scoring from diverse perspectives for both option entities and question entities respectively. By injecting the refined triples into LLMs in combination with few-shot and chain-of-thought prompting, our method achieves consistent performance improvements over existing baselines across four Korean medical licensing domains: physician, nurse, pharmacist, and dentist.

키워드