融合BERT与领域本体规则的农业机械化管理知识图谱构建与智能问答应用研究

Constructing of Agricultural Mechanization Management Knowledge Graph by Integrating BERT and Domain Ontology Rules and Its Application in Intelligent Question Answering

摘要:

针对农业机械化管理领域专业知识分散、异构文档格式多样以及智能决策需求迫切等问题，本研究提出了一种基于 BERT 预训练语言模型与领域本体规则相结合的自动化知识抽取与融合方法，旨在构建高质量的农业机械化管理知识库。首先，设计涵盖农机装备、维护活动、故障诊断和政策法规等类别的领域本体;其次，借助 BERT 预训练模型在农机管理学著作、学术文献、技术手册和政策法规等多源文本上精确提取实体与关系，并通过本体规则对提取结果进行校验与去重;最后，将实体、关系与高质量三元组载人图数据库，支撑智能问答与决策分析应用。试验结果表明：关系抽取模型在各数据源中高置信三元组占比最高达 88.9%; 智能问答系统在 450 条典型业务测试中准确率达 90.9%, 幻觉率低至 3.1%, 并且具备完全可追溯性，其性能显著优于 CPT-4o 等通用大模型;系统端到端平均响应时延 150ms, 吞吐率 200req/s, 资源利用率控制在合理范围内。该方法不仅实现了农业机械化管理领域知识的自动化高效整合，还为智慧农机决策支持提供了可复制和持续演进的技术路径。

Abstract:

Aiming to address the issues of fragmented domain knowledge, diverse heterogeneous document formats, and the urgent demand for intelligent decision-making in the field of agricultural machinery management, an automated knowledge extraction and fusion method that combined a BERT pretrained language model with domain ontological rules was proposed, with the aim of constructing a high-quality knowledge base for agricultural machinery management. Firstly, a domain ontology covering categories such as agricultural machinery equipment, maintenance activities, fault diagnosis, and policies and regulations was designed. Secondly, with the help of the BERT pretrained model, entities and relations were accurately extracted from multi-source texts, including monographs on agricultural machinery management, academic literature, technical manuals, and policy and regulatory documents, and the extraction results were validated and de-duplicated by using ontological rules. Finally, the entities, relations, and high-quality triples were loaded into a graph database to support intelligent question answering and decision analysis applications. Experimental results showed that the proportion of high-confidence triples produced by the relation extraction model reached up to 88. 9% across different data sources; the intelligent question answering system achieved an accuracy of 90. 9% on 450 typical business test cases, with a hallucination rate as low as 3.1% and fully traceable answers, and its performance was significantly better than that of general-purpose large models such as CPT - 4o. The system attained an average end-to-end response latency of 150 ms, a throughput of 200 req/s, and kept resource utilization within a reasonable range. This method not only enabled automated and efficient integration of knowledge in the field of agricultural machinery management, filling a gap in related research, but also provided a replicable and continuously evolvable technical path for decision support in smart agricultural machinery.