基于开放域知识图谱嵌入的网络舆情人名消歧模型  

Network public opinion person name disambiguation model based on open domain knowledge graph embedding

作　　者：刘金硕[1] 张思奇 LIU Jinshuo;ZHANG Siqi(Key Laboratory of Aerospace Information Security and Trusted Computing of Ministry of Education,School of Cyber Science and Engineering,Wuhan University,Wuhan 430072,China)

机构地区：[1]武汉大学国家网络安全学院空天信息安全与可信计算教育部重点实验室,湖北武汉430072

出　　处：《武汉大学学报(工学版)》2025年第2期316-324,共9页Engineering Journal of Wuhan University

基　　金：国家自然科学基金项目(编号:U1936107);国家重点研发计划项目(编号:2020YFA0607902)。

摘　　要：针对网络舆情热点人名消歧任务难以有效获取和融合文本及外部特征的问题,提出了开放域知识图谱嵌入的方法。构建了新的知识图谱嵌入模型TransOD,将原本的知识嵌入投影到基于Transformer的双向编码器(bidirectional encoder representations for transformers,BERT)语义空间,采用多维高斯分布表示开放域实体和关系的不确定性,充分利用并结构化地理解语义信息,获取特征向量。随后定义了文本间相似度的计算方法,并通过设计聚类方法实现消歧。在网络人物搜索评测会议(web people search evaluation campaign,WePS)3年的评测数据集上,F0.5值较现有最优方法分别提升了0.034、0.043和0.102。同时验证了TransOD的开放域链接预测效果相较于现有的最优知识图谱嵌入方法,平均倒数排名(mean reciprocal rank,MRR)、Hits@1和Hits@10分别提升了0.049、19.3%和4.8%,开放域三元组补全的准确率提升了11.7%。In view of the difficulty in obtaining and integrating text and external features effectively in person name disambiguation tasks of popular online public opinions,an open domain knowledge graph embedding method is proposed.A new knowledge graph embedding model TransOD is proposed,which projects the original knowledge embedding into the bidirectional encoder representations for transformers(BERT)semantic space based on Transformer.The uncertainty of open domain entities and relations is represented by using multidimensional Gaussian distribution,and the semantic information is fully used and structurally understood to obtain the feature vector.Then the inter-text similarity calculation method is defined and the clustering method is designed to realize disambiguation.In the three-year evaluation datasets of the Web People Search Evaluation Campaign(WePS),compared with the existing optimal methods,the F0.5 values have respectively improved by 0.034,0.043 and 0.102.At the same time,the prediction effect of open domain link of TransOD is verified.Compared with the existing optimal knowledge graph embedding methods,MRR(mean reciprocal rank),Hits@1 and Hits@10 are improved by 0.049,19.3%and 4.8%,respectively.The accuracy of open domain triplet completion is improved by 11.7%.

关 键 词：网络舆情人名消歧 知识图谱 知识图谱嵌入 开放域任务 聚类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]