公文知识图谱构建与应用  

Official Document Knowledge Graph:Construction and Application

作　　者：李炜卓 周文博 卢冰洁 高辉[1] 边宇阳 张浩魏 那崇宁 许文杰 LI Weizhuo;ZHOU Wenbo;LU Bingjie;GAO Hui;BIAN Yuyang;ZHANG Haowei;NA Chongning;XU Wenjie(School of Modern Posts,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;Zhejiang Lab,Hangzhou 311121,China;School of Modern Posts,School of Computer Science,Nanjing 210023,China)

机构地区：[1]南京邮电大学现代邮政学院,南京210003 [2]之江实验室,杭州311121 [3]南京邮电大学计算机学院,南京210023

出　　处：《小型微型计算机系统》2024年第6期1281-1291,共11页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62006125,52077107)资助;国家重点研发计划项目(2022YFB4501503)资助;江苏省双创博士项目(JSSCBS20210532)资助。

摘　　要：海量电子公文的高效管理一直是社会服务研究的热门话题.现有的研究工作主要聚焦在公文归档与要素抽取上,但并未提供开源的公文数据集,同时对于公文之间的关联性与时效性缺乏深入分析.文中的工作致力于从多个领域收集公文,构建一个高质量、开源的公文知识图谱,以此来填补上述研究工作存在的鸿沟.文中详细地展示了公文知识图谱的构建框架,包括轻量级本体的设计、面向小样本场景的要素抽取算法以及面向实体对齐的知识融合算法,以此让不同来源的公文能够共享更多的链接.此外,文中进一步讨论了公文知识图谱的潜在应用,并面向公文关联发现场景提出了一种基于要素抽取增强的网络表示学习公文推荐算法.实验表明,所提方法在评估指标Rankmin与MAP上整体效果均好于已有方法.Effective management of massive electronic official documents is one of the hot topics for social services.Existing works focus on the tasks of element extraction and document archiving,but the data sets of these works are not open source.Besides,the relevance and timeliness among these official documents are not considered.To fill the above gaps,this work dedicates a continuous effort to collect official documents from various domains and construct an official document knowledge graph,called ODKG.We present the framework of ODKG comprehensively,including designed lightweight ontology,feature extraction algorithms for few-shot samples,and knowledge alignment methods for entity alignment,by which these documents could share more linkages with each other.Moreover,this paper discusses the potential applications of ODKG,and an algorithm based on network representation learning enhanced with element extraction is proposed for document recommendation.Experimental results show that our proposed method can obtain better performances than existing methods in terms of Rank min and MAP.

关 键 词：知识图谱 要素抽取 公文归档 公文推荐 网络表示学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]