Knowledge Graph Enhanced Large Language Models for Traditional Chinese Medicine Question-Answering

Large language models (LLMs) have gained much attention in Traditional Chinese medicine (TCM) question-answering (Q&A). However, existing LLMs methods lack logical reasoning for TCM syndrome differentiation, and rely on the data collected from various data sources, easy lead to hallucinations and misdiagnosis in TCM Q&A. Moreover, knowledge graph can construct the logical connections of TCM syndrome differentiation, but it is difficult to reason independently based on the dynamic evolution of the disease condition and cannot simulate the dynamic thinking process in clinical practice of TCM. To address the above problems, we proposed a method to leverage the knowledge graph to enhance LLMs for achieving TCM Q&A. To be specific, we first construct the knowledge graph through data collection, entity extraction, knowledge modeling, and graph storage. Next, we design a triplet semantic network to solve the problems of reasoning jumps and taboos missing in existing LLMs methods, which includes the dialectical logic layer, the prescription rule layer, and the safety taboo layer. Then, we enhance the prompt engineering of LLMs through the knowledge graph and force the LLMs to follow the standardized reasoning path. Finally, we use the enhanced LLMs method to identify the entities from the new question and associate them with the knowledge graph to obtain the diagnosis and treatment plan that conforms to the theoretical logic of TCM. Experimental results on two benchmark datasets and case studies indicate that our proposed method can achieve a superior performance against existing LLMs methods.