基于大模型驱动的水稻种植领域知识图谱构建

Large Model-driven Knowledge Graph Construction for Rice Cultivation Domain

摘要:

传统水稻种植知识管理存在知识碎片化、检索困难和更新滞后等问题,难以满足现代精准农业的信息需求。为此提出了基于大模型驱动的水稻种植领域知识图谱构建方法。构建了包含51篇期刊论文和2本专业书籍的RC-PKBD 数据集(40.98万字符),基于改进的LightRAG 框架设计专业提示工程策略,定义15类实体类型和12类关系类型,实现专业知识的高效抽取。建立了基于图结构的双层次检索机制,成功构建包含10723个节点和10955条边的知识图谱。质量评估显示,具体型问题准确性达91.73%,抽象型问题覆盖度达90.20%,相比GraphRAG方法准确性提升36.60个百分点。本研究为水稻种植领域知识数字化提供了有效解决方案,对推动智慧农业发展具有重要价值。

Abstract:

Traditional rice cultivation knowledge management faces challenges, including knowledge fragmentation, retrieval difficulties, and outdated information, which cannot meet the demands of modern precision agriculture. A large language model-driven knowledge graph construction method for rice cultivation domain was proposed. The RC-PKBD dataset containing 409800 Chinese characters from 51 journal articles and 2 professional books were constructed. Based on an improved LightRAG framework with domain-specific prompt engineering strategies, totally 15 entity types and 12 relationship types were defined to achieve efficient extraction of professional knowledge. The construction process involved four key steps: text processing and chunking, entity and relationship extraction using large language models, entity deduplication and alignment through vector similarity calculation, and graph structure construction with storage. A dual-level retrieval mechanism based on graph structure was established, successfully constructing a knowledge graph containing 10723 nodes and 10955 edges. The system demonstrated significant advantages over traditional deep learning methods in semantic understanding depth and domain adaptability. Quality evaluation showed that accuracy reached 91.73% for specific questions and coverage achieved 90.20% for abstract questions, with a 36.60 percentage points improvement in accuracy compared with GraphRAG. The research result can provide an effective solution for rice cultivation knowledge digitalization and has important value for advancing smart agriculture development, while offering a replicable technical approach for knowledge services in other agricultural domains.