摘要
研究目的:探索大语言模型在土地类行政处罚裁量场景中的领域适配路径,实现裁量规则结构化解析与裁量知识智能服务的深度融合。研究方法:以土地类行政处罚裁量基准文件为数据基础设计裁量知识表达模型,构建裁量知识图谱作为结构化知识库,采用LoRA微调技术对预训练大语言模型进行领域适配。研究结果:(1)“模式层—数据层”裁量知识表达范式中,模式层通过本体建模实现“事项—依据—情节—处罚”的语义化关联,数据层依托实例化数据完成裁量知识的结构化存储;(2)基于知识图谱查询语言检索算法,自动化映射裁量知识中关键节点与关系,实现了裁量领域知识形式化表达;(3)采用微调技术对模型进行参数优化,在裁量知识问答、诊断预警信息生成等核心任务中验证了大语言模型适配效能。研究结论:裁量知识图谱可有效破除裁量“知识孤岛”困境,知识赋能的裁量领域大语言模型可以为土地管理现代化提供有效帮助。
The purposes of this study are to explore the domain adaptation pathways of large language models in the context of land-related administrative penalty discretion,to achieve a deep integration of structured interpretation of discretion rules and intelligent services for discretion knowledge.The research method is to design a discretion knowledge expression model based on the benchmark file of land administrative penalties,to construct a discretion knowledge graph as a structured knowledge base,and to use LoRA that realizes domain adaptation of the pre-trained large language model.The research results show that:1)in the“schema-data”discretion knowledge representation paradigm,the schema layer establishes semantic associations among“matters-legal basis-circumstances-penalties”through ontology modeling,while the data layer relies on instantiated data to achieve structured storage of discretion.2)Based on graph query language retrieval algorithms,the key nodes and relationships in discretion knowledge were automatically mapped,achieving formal expression of domain-specific discretion knowledge.3)The fine-tune technique was employed to optimize the parameters of the large language model,validating the efficacy of domain adaptation in core tasks such as discretion knowledge question answering and early-warning information generation.In conclusion,the knowledge graph about discretion can effectively break the dilemma of“knowledge silos”in discretion.The discretion domain model empowered by knowledge can provide effective assistance for the modernization of territorial space governance.
作者
赵绍轩
周晓光
侯东阳
柳婷
贺鸿愿
符雨星
ZHAO Shaoxuan;ZHOU Xiaoguang;HOU Dongyang;LIU Ting;HE Hongyuan;FU Yuxing(School of Geosciences and Info-Physics,Central South University,Changsha 410083,China;Ningbo Institute of Surveying and Mapping and Remote Sensing Technology,Ningbo 315042,China;The First Surveying and Mapping Institute of Hunan Province,Changsha 410114,China)
出处
《中国土地科学》
北大核心
2025年第7期33-44,共12页
China Land Science
基金
国家重点研发计划(2022YFB390420501)。
关键词
土地类行政处罚裁量
知识图谱
大语言模型
land-related administrative penalty discretion
knowledge graph
large language model
