融合知识图谱和大模型的医疗智能问答方法研究

Research on Medical Intelligent Question Answering Method Combining Knowledge Graph and Big Model

摘要

医疗领域的自动问答对答案的准确性有很高的要求。尽管大语言模型（LLM）提供了通用的问答能力，但是无法满足医疗领域对答案准确性的要求。与此相对，基于知识图谱检索的自动问答依赖于客观的知识表达对回答质量提供了可靠性保证，然而目前的方法中存在知识图谱检索效率不高、检索不充分、检索过多冗余信息，以及对较复杂的问题理解不充分进而影响检索的质量等问题。为此，本文中将知识图谱与 LLM 相结合，基于 LLM 对用户的提问进行问题分解，对每一个子问题在知识图谱中进行子图搜索，再将融合后的子图交给 LLM 以生成可靠的答案。文中的方法在医疗数据集 GenMedGPT-5k、LiveQA、 HealthCareMagic-100k 和知识图谱FB15k-237 上进行了实验。实验表明，文中的方法取得了较好的性能。

Abstract

In the medical area，automated question answering places high demands on the accuracy of answers.Although large language models（LLMs）provide general-purpose question-answering capabilities，they cannot meet the stringent accuracy requirements of the medical domain.In contrast，knowledge graph-based automated question answering relies on objective knowledge representation to en- sure answer reliability.However，current methods suffer from issues such as inefficient knowledge graph retrieval，insufficient cover- age，excessive redundant information，and inadequate understanding of complex questions，which negatively impact retrieval quality.To address these challenges，this paper integrates knowledge graphs with LLMs.Specifically，the LLM is used to decompose user questions into sub-questions，each of which is then subjected to subgraph retrieval in the knowledge graph.The merged subgraphs are then fed back to the LLM to generate reliable answers.The proposed method is evaluated on medical datasets（GenMedGPT-5k，LiveQA， HealthCareMagic-100k）and the knowledge graph FB15k-237.Experimental results demonstrate that the proposed approach achieves superior performance.

关键词

大语言模型 / 知识图谱 / 检索增强 / 知识子图

Key words

Large Language Model（LLM） / knowledge graph / search enhancement / knowledge subgraph

参考文献

[1] Achiam J, Adler S, Agarwal S, et al. Gpt-4 technical report[J]. arx- iv preprint arxiv:230308774,2023.

[2] Dubey A, Jauhri A, Pandey A, et al. The llama 3 herd of models[J]. arxiv:2407. 21783,2024.

[3] Deepseek A I, Guo D, Yang D, et al. DeepSeek-R1:incentivizing reasoning capability in LLMs via reinforcement learning[J]. arxiv. org/abs/2501.12948, 2025.

[4] Sun Y, Wang S, Li Y, et al. ERNIE:enhanced representation through knowledge integration[J]. arxiv.org/abs/1904. 09223,2019.

[5] Xie Q, Schenck E J, Yang H S, et al. Faithful AI in medicine :a sys- tematic review with large language models and beyond[J]. MedRx- iv, 2023.04.18.23288752,doi: 10.1101/2023.04.18.23288752.

[6] Liu P, Yuan W, Fu J, et al. Pre-train, prompt,and predict:a system- atic survey of prompting methods in natural language processing[J]. ACM Computing Surveys, 2023, 55(9):1-35.

[7] Lewis P, Perez E, Piktus A, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[C]// Advances in Neural Infor- mation Processing Systems, 2020:9459-9474.

[8] Rau D, Wang S, Dejean H, et al. Context embeddings for efficient answer generation in retrieval-augmented generation[C]// Proceed- ings of the Web Search and Data Mining, 2025:493-502.

[9] Li X, Zhao R, Chia Y K, et al. Chain of knowledge:a framework for grounding large language models with structured knowledge ba- ses[J]. arxiv.org/abs/2305.13269,2023.

[10] Jiang B, Wang Y, Luo Y, et al. Reasoning on efficient knowledge paths:knowledge graph guides large language model for domain question answering[C]// Proceedings of the IEEE International Conference on Knowledge Graph(ICKG), 2024:142-149.

[11] Roberts A, Raffel C, Shazeer N J A P A.How much knowledge can you pack into the parameters of a language model?[J]. arxiv pre- print arxiv,2020.08910.

[12] Robertson S, Zaragoza H J F, Retrieval T I I. The probabilistic rele- vance framework:BM25 and beyond[J]. Foundations and Trends in Information Retrieval, 2009, 3(4):333-389.

[13] Peng B, Galley M, He P, et al. Check your facts and try again:Im- proving large language models with external knowledge and auto- mated feedback[J]. arxiv Preprint arxiv,2023. 12813.

[14] Kamphais C, Vries A P D, Boytsov L, et al. Which BM25 do you mean?A large-scale reproducibility study of scoring variants[C]// European Conference on Information Retrieval(ECIR), 2020:28-34.

[15] Sharma A, Kumar S J D, Engineering K. Ontology based semantic retrieval of documents using Word2vec model[J]. Data ＆ Knowl- edge Engineering, 2023,144:102110,doi:10.1016/j.datak.2022.102110.

[16] Li Y, Li Z, Zhang K, et al. ChatDoctor:a medical chat model fine- tuned on a large language model meta-AI(LLaMA)using medical domain knowledge[J]. Cureus, 2023, 15 (6):e40895,doi: 10.7759/cureus.40895.

[17] Abacha A R, Agichtein F, Pinter Y, et al. Overview of the medical question answering task at TREC 2017 LiveQA[C]// Proceedings of the Text REtrieval Conference( TREC), 2017:1-12.

[18] Bollacker K D, Evans C, Paritosh P, et al. Freebase:a collaborative- ly created graph database for structuring human knowledge[C]// Proceedings of the ACM Sigmod International Conference on Man- agement of Data, 2008:1247-1250.

[19] Zhang Z, Han X, Liu Z, et al. ERNIE: enhanced language represen- tation with informative entities[J]. arxiv preprint arxiv: 1905. 07129,2019.

[20] Lin C Y. ROUGE:a package for automatic evaluation of summaries[C]// Proceedings of Workshop on Text Summarization Branches Out, 2004:74-81.

[21] Sellam T, Das D, Parikh A P. BLEURT:learning robust metrics for text generation[J]. arxiv preprint arxiv,2004. 04696.

[22] Sun Y, Wang S, Feng S, et al. ERNIE 3.0:large scale knowledge enhanced pre-training for language understanding and generation[J]. arxiv.org/abs/2107.02137,2021.

[23] Yang R, Liu H, Marrese Taylor E, et al. KG-Rank: enhancing large language models for medical QA with knowledge graphs and rank- ing techniques[J]. arxiv preprint arxiv, 2403.05881, 2024.

[24] Wen Y, Wang Z, Sun J. MindMap:knowledge graph prompting sparks graph of thoughts in large language models[J]. arxiv pre- print arxiv,2308.09729,2023.

基金资助

上海交通大学医工交叉项目(YG2024QNB05)