Abstract: To address the insufficient engineering knowledge and limited experience of some frontline dam safety management personnel, this study proposes an intelligent question-answering method for dam safety knowledge. First, based on practical dam safety management work, a dam safety knowledge framework was developed and a corresponding knowledge graph was constructed. Second, the LoRA parameter-efficient fine-tuning technique was applied to further train a pretrained large language model, thereby a domain-specific model for dam-related semantic parsing was developed. Third, an entity-linking method based on an N-gram model was introduced to automatically associate the key information in user queries with corresponding entities in the knowledge graph. Finally, a standardized Cypher query template was designed for the knowledge graph to enable intelligent retrieval of dam safety knowledge. Case studies demonstrate that the constructed knowledge graph can effectively organize and structure the multi-source heterogeneous data related to dam safety management. As well, the fine-tuned domainspecific model can accurately convert natural-language queries into structured retrieval intents and keywords. The combined use of the knowledge graph and the large AI model can overcome limitations in natural-language parsing for knowledge graphs and reduce hallucinations in large models, enabling intelligent dam safety question-answering through natural-language inputs.

Key words: dam, question-answering, knowledge graph, large AI model, semantic parsing

摘要： 针对部分一线大坝安全管理人员工程知识量不足、经验欠缺等问题，提出了大坝安全知识智能问答方法。首先，从大坝安全管理的实际业务出发，提出大坝安全知识框架，构建大坝安全知识图谱；其次，利用AI大模型LoRA高效参数微调技术对预训练大模型进行二次训练，形成大坝安全语义解析专业大模型；然后，提出基于N-gram模型的实体链接方法，形成提问语句中关键信息与知识图谱中实体的自动关联机制；最后，设计固定格式的知识图谱Cypher查询语句，实现大坝安全专业知识的智能检索。工程实例应用表明，本研究构建的知识图谱能够将多源异构的大坝安全管理数据进行高效组织和结构化表达，训练形成的专业大模型能够将自然语言解析为结构化的检索意图和关键词。知识图谱和AI大模型的结合使用可同时解决知识图谱自然语言解析和AI大模型知识幻觉的短板，实现自然语言输入形式的大坝安全知识智能问答。

关键词: 大坝, 知识问答, 知识图谱, AI大模型, 语义解析