用于实体对齐的多跳邻域联合采样方法  

Multi-hop neighborhood joint sampling method for entity alignment

作　　者：祁雨婷 邵玉斌[1] 杜庆治[1] 龙华[1,2] 马迪南 QI Yu-Ting;SHAO Yu-Bin;DU Qing-Zhi;LONG Hua;MA Di-Nan(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China;Yunnan Provincial Key Laboratory of Media Convergence,Kunming 650100,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,昆明650500 [2]云南省媒体融合重点实验室,昆明650100

出　　处：《四川大学学报（自然科学版）》2024年第6期144-152,共9页Journal of Sichuan University(Natural Science Edition)

基　　金：云南省媒体融合重点实验室项目(220235205)。

摘　　要：实体对齐的目的是寻找不同知识图谱中指向同一概念的实体,然而不同知识图谱之间的结构异构性增加了实体对齐的难度.现有方法主要使用实体邻域信息来降低结构异构性,但仍未对邻域的构建进行优化.因此,提出一种用于实体对齐的多跳邻域联合采样方法.首先使用图卷积神经网络得到两个知识图谱中实体的嵌入向量;其次使用多跳邻域联合采样机制,基于邻居实体在本实体多跳子邻域中的复现频次,计算出邻居实体的结构采样权重,基于邻居实体与本实体间的语义相似度,计算出邻居实体的语义采样权重,联合结构采样权重和语义采样权重,遍历所有实体,为每个实体采样多跳邻居,构建出各自的多跳邻域;之后将这些实体及其候选实体间的多跳邻域跨图交互信息聚合于实体嵌入向量中,最后计算嵌入向量间的对齐距离.在实体对齐公共数据集DBP15k的3个跨语言子集上进行实验,相较于基线模型,所提方法在Hit@1指标上分别提升3.3%,3.5%和1.8%.实验结果表明所提方法能有效提升实体对齐结果的准确性.The purpose of entity alignment is to identify entities across different knowledge graphs that refer to the same concept.However,the structural heterogeneity between different knowledge graphs increases the difficulty of entity alignment.Existing methods primarily utilize neighborhood information to mitigate the structural heterogeneity of knowledge graphs,yet optimizing the construction process of entity neighborhoods within knowledge graphs remains an unresolved issue.To address this,we propose a multi-hop neighborhood joint sampling method for entity alignment.Initially,we employ graph convolutional neural networks to ob⁃tain embedding vectors for entities in two knowledge graphs.Subsequently,through a multi-hop neighbor⁃hood joint sampling mechanism,we calculate the structural sampling weight of neighboring entities based on their recurrence counts in the multi-hop sub-neighborhoods of the entity and the semantic sampling weight based on the semantic similarity between the neighbors and the entity.By traversing all entities and combining the computed structural and semantic sampling weights,we sample multi-hop neighbors for each entity to construct their respective multi-hop neighborhoods.We then aggregate the multi-hop neighborhood cross-graph interaction information between these entities and their candidate entities into the entity embedding vec⁃tors and calculate the alignment distance between the embedding vectors.Experiments conducted on three cross-lingual subsets of the public entity alignment dataset DBP15k demonstrate that our method achieves im⁃provements of 3.3%,3.5%,and 1.8%in the Hit@1 metric compared to baseline models.This indicates that the proposed method can effectively enhance the accuracy of entity alignment results.

关 键 词：知识图谱 实体对齐 多跳邻域 邻域结构 语义相关性 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]