安徽地质2025，Vol.35Issue(2)：158-163,167,7.
基于通用大语言模型和领域知识图谱的岩相学知识问答
Petrographic knowledge Q&A based on general large language models and domain-specific knowledge graphs
摘要
Abstract
Recent advancements in large language models(LLMs)have revolutionized natural language processing with robust generative capabilities.However,general-purpose LLMs frequently exhibit limitations in domain-specific knowledge-intensive tasks-such as geological question answering(QA)-due to insufficient exposure to specialized training corpora,often leading to factual inaccuracies("hallucinations").To address this challenge,we propose a petrographic knowledge retrieval-enhanced QA framework that synergizes general LLMs with domain-specific knowledge graphs(KGs).This architecture leverages LLMs'generative strengths while grounding responses in structured,computable petrographic expertise.We validate the framework through comparative experiments on 494 petrographic QA pairs.Results demonstrate that integrating petrographic KGs-encoding computer-interpretable domain knowledge-significantly enhances LLM performance,reducing hallucination rates by 38%compared to standalone LLM baselines.Notably,in scenarios where Chinese geological corpora remain scarcely shared and no official domain-specific LLMs exist for geology,our framework provides a pragmatic transitional solution for accurate geological knowledge services.关键词
地学知识图谱/知识问答/提示工程/查询语句生成/自然语句答案生成Key words
geological knowledge graph/knowledge question answering/prompt engineering/query statement generation/natural language answer generation分类
天文与地球科学引用本文复制引用
陈忠良,段剑超,郑超杰,施成艳..基于通用大语言模型和领域知识图谱的岩相学知识问答[J].安徽地质,2025,35(2):158-163,167,7.基金项目
国家自然科学基金项目"数据驱动与相似度推理知识嵌入的可扩展岩石图像识别研究""基于小样本机器学习的原生晕地球化学信息挖掘及三维成矿预测研究——以新疆阿舍勒铜锌矿床为例"(编号:42372342,42202328)和2023年度部省合作试点项目"长江干流(安徽段)非金属露天矿山生态修复成效评估技术方法研究"(编号:2023ZRBSHZ008)联合资助 （编号:42372342,42202328）
