An Exploratory Study on the Use of LLMs for Classical Medical Texts - Constructing and Evaluating a Source-Concept Linked Knowledge Base Inspired by the Compilatory Structure of Dongui Bogam -

고전 의서의 LLM 활용을 위한 탐색적 연구 - 『동의보감』 편제에 착안한 원문-개념어 연계형 지식베이스 구축과 평가 -

This study constructs and evaluates a source-concept-linked knowledge base, or Hub-Spoke Knowledge Base (HS-KB), based on the editorial structure of Donguibogam, in order to explore the applicability of large language models (LLMs) to classical medical texts. The HS-KB was designed to reflect the way readers navigate Donguibogam: moving from tables of contents and thematic sections to original passages, and then to related concepts such as diseases, prescriptions, materia medica, and acupoints. It consists of three components: Hub notes that preserve original textual passages, Spoke notes that collect references by conceptual terms, and Index notes that serve as navigational entry points. To assess its utility, this study compared LLM-generated answers based on the HS-KB with those based on direct reference to the original TXT files, using twenty questions. The results showed that the HS-KB achieved a higher average score in faithfulness and was advantageous in synthesizing dispersed textual evidence and relationships among concepts. Future improvements should include assigning knowledge-type tags, such as definition, pathogenesis, therapeutic principle, and prescription, to Hub passages, and prioritizing core evidence according to question type.