水利综合知识图谱构建研究  

Research on water conservancy comprehensive knowledge graph construction

作　　者：段浩[1] 韩昆[1] 赵红莉[1] 蒋云钟[1] 李豪 毛文山 DUAN Hao;HAN Kun;ZHAO Hongli;JIANG Yunzhong;LI Hao;MAO Wenshan(China Institute of Water Resources and Hydropower Research,Beijing 100038,China)

机构地区：[1]中国水利水电科学研究院,北京100038

出　　处：《水利学报》2021年第8期948-958,共11页Journal of Hydraulic Engineering

基　　金：中国工程科技知识中心水利专业知识服务系统(CKCEST-2020-2-9)。

摘　　要：基于对水利知识特点的分析,提出了水利综合知识体系的描述方法,包括水利知识的定义、组成与关联;构建了水利知识图谱的构建框架和关键技术体系,以水利行业结构化业务数据的实体关系转换为基础,采用双向长短期记忆神经网络(Bi-directional Long Shot-Term Memory Neural Network,BiLSTM)与条件随机场(Conditional Random Fields,CRF)方法识别半结构化、非结构化学科知识文本以及互联网数据中的水利实体,使用模式匹配和共现网络分析方法抽取各实体间关系,对涉水对象及其属性进行补充,基于风险最小化的最小风险映射模型(Risk Minimization based Ontology Mapping,RiMOM)进行了多源异构水利实体的融合,实现了涉水对象与水利学科知识的融合与关联,形成水利综合知识的建模和表达。在图谱构建过程中,累计抽取水利实体136万个,构建实体关系300余万条,抽取的水利实体对象的标注准确率在80%以上。基于该图谱可实现水利知识的跨域查询与检索,学科图谱与水网图谱间关系查询,挖掘不同水利实体间的隐含关系,提高水利知识检索的效率和知识挖掘发现的能力。Based on the analysis of water conservancy knowledge characteristics,this research proposed the description method of water conservancy comprehensive knowledge graph,including the definition,composition and relevancy of water conservancy knowledge.The construction framework and key technologies of water conservancy knowledge graph were built.With the help of Bi-directional Long Shot-Term Memory Neural Network(BiLSTM)and Conditional Random Fields(CRF)models,the entities transformed from structured data were used to identify the entities in semi-structured and structured text knowledge and internet data.The relationships between entities were extracted with pattern matching and co-occurrence network analysis methods to supplement water related objects and attributes.RiMoM model was used to fuse the water conservancy entities and disciplinary knowledge.Through the graph construction,136 million water conservancy entities were extracted and more than 300 million relationships were established.The annotation accuracy of water conservancy objects was higher than 80 percent.Based on the graph constructed in this research,both knowledge and relationships could be searched and the hidden relations between different entities could be found.

关 键 词：知识图谱 水利综合知识 水资源管理 学科知识 BiLSTM CRF 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]