基于双子图和注意力机制的知识图谱补全方法  

Knowledge graph completion method based on bipartitegraphs and attention mechanism

作　　者：周粤 范永胜 桑彬彬 周岩 Zhou Yue;Fan Yongsheng;Sang Binbin;Zhou Yan(College of Computer&Information Science,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学计算机与信息科学学院,重庆401331

出　　处：《计算机应用研究》2025年第1期93-99,共7页Application Research of Computers

基　　金：国家自然科学基金青年科学基金资助项目(62306054);重庆师范大学研究生科研创新资助项目(YKC23022)。

摘　　要：针对现有的知识图谱补全方法捕获知识图谱结构信息能力不足的问题,提出了一种基于双子图和注意力机制以获取全局结构信息完成知识图谱自动补全的模型。该模型首先分别构建以实体和关系为中心的双子图,来分别捕获实体邻域信息和关系结构的潜在有用信息,并将双子图形成的信息输入到编码器中以更好地更新实体和关系结构信息;然后,利用注意力机制自适应地学习更新后实体和关系之间的重要交互特征;最后,将包含全局结构信息的特征向量输入到解码器中,通过一个评分函数,对输入的特征边进行打分预测,最终使用预测结果来完成知识图谱补全任务。与基线方法的性能相比,该方法在FB15K-237和NELL995数据集上的MRR和hits@10评测指标分别取得了5.1、8.8和3.4、2.2百分点的显著提升,同时在WN18RR数据集上,这两个指标也分别提高了0.1和1.9百分点。实验结果表明,所建立模型采用的结构能有效捕获知识图谱全局结构信息,进而显著增强模型的表达能力和预测性能。To address the issue of existing knowledge graph completion methods’limited capability in capturing structural information within knowledge graphs,this paper proposed a novel model that leveraged bipartite graphs and an attention mechanism to acquire global structural insights and facilitate automatic knowledge graph completion.This model firstly constructed two subgraphs centered on entities and relationships to capture potential useful information about entity neighborhood and relationship structures,and inputted the information formed by the two subgraphs into the encoder to better update entity and relationship structure information.Then,it used attention mechanisms to adaptively learn important interaction features between updated entities and relationships.Finally,it inputted the feature vectors containing global structural information into the decoder,and it actively employed a scoring function to assess and predict scores for the input feature edges,ultimately utilizing the predicted outcomes to accomplish the task of knowledge graph completion.Comparing the performance of the proposed method with the baseline method on the FB15K-237 and NELL995 datasets,the MRR and hits@10 evaluation indicators achieved significant improvements of 5.1,8.8,and 3.4,2.2 percentage points,respectively.At the same time,on the WN18RR dataset,these two indicators also were improved by 0.1 and 1.9 percentage points,respectively.The experimental results show that established model proactively adopts a structure that effectively captures the global structural information of the knowledge graph,thereby significantly enhancing the expression ability and predictive performance of the model.

关 键 词：知识图谱补全 双子图 注意力机制 编码器 解码器 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]