联合实体类别及邻域信息指导的实体对齐方法  

Joint classes and neighborhood information guided embedding for entity alignment

作　　者：蒋敏[1] 王淑营[1] Jiang Min;Wang Shuying(School of Computing&Artificial Intelligence,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学计算机与人工智能学院,成都611756

出　　处：《计算机应用研究》2024年第12期3614-3621,共8页Application Research of Computers

基　　金：四川省重大科技专项资助项目(2022ZDZX0003)。

摘　　要：针对当前的一些实体对齐工作仅利用关系结构信息,而忽略了实体与实体类别之间的联系以及属性信息的重要性的问题,提出一种联合实体类别及邻域信息指导的实体对齐方法。该方法通过嵌入关系三元组、属性三元组和实体类别信息,利用BERT预训练模型进行编码,解决缺失语义信息的问题,采用卷积神经网络提取属性特征,利用实体类别和种子词构建类别冲突矩阵。最后通过实体和属性指导实体对齐,并在迭代训练过程中将高质量的实体对加入到种子词中。在四个公开数据集上hits@1分别可达到90.1%、94%、92.5%和94.5%,相较于基准模型平均提升了6.225百分点,实验结果表明提出方法可以有效地实现实体对齐。This paper proposed an entity alignment method leveraging entity categories and neighborhood information.The method overcame the shortcomings of some approaches that did not consider the importance of class and attribute information.The method embedded relation triples,attribute triples,and entity class information,and encoded them using the BERT model to address the issue of missing semantic information.It used convolutional neural networks to extract attribute features.Then,it constructed a class conflict matrix with entity classes and seed words to guide the alignment process.During the iterative training,the method added high-quality entity pairs to the seed words.The hits@1 on the four public datasets can reach 90.1%,94%,92.5%,and 94.5%,respectively,which is an average improvement of 6.225%compared with the benchmark model.The experimental results show that the proposed model performs well in the task of entity alignment.

关 键 词：知识图谱 实体对齐 BERT 实体类别 属性信息 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]