Cross-lingual entity alignment method based on attribute weight updating network

基于属性权重更新网络的跨语言实体对齐方法

Key Laboratory of Linguistic and Cultural Computing Ministry of Education, Northwest Minzu University, Lanzhou 730030, China

Received: 22 February 2023

Abstract

Due to the large number of attributes and the low repetition rate in a cross-lingual knowledge graph, it is difficult for an alignment task to embed attribute information efficiently. To solve the problem, an entity alignment model based on attribute weight updating network was proposed. Firstly, in order to embed attribute information efficiently, attribute embedding is approximately constructed with entity embedding through a constructor, thus avoiding their separate training. Secondly, based on the fact that different attributes make different contributions to entity alignment, an attribute weight updating module based on graph attention network was proposed to update the weight of each attribute through using attention scores in the process of training. Finally, attribute embedding and attribute weight information were aggregated into entity embedding with an attribute aggregation module to strengthen the representation of entity embedding and improve the entity alignment performance. The experimental results show that the proposed model achieves 0.751, 0.805 and 0.915 scores respectively from the Hits@1 score in three cross-lingual datasets. Its alignment performance is better than that of the current mainstream entity alignment method.

摘要

跨语言知识图谱中属性数量庞大且重复率低导致对齐任务中属性信息难以高效嵌入。针对上述问题, 提出了一种基于属性权重更新网络的跨语言实体对齐模型。为了高效地实现属性信息的嵌入, 通过一个构造器利用实体嵌入来近似地构造属性的嵌入, 避免了属性嵌入的单独训练; 基于不同属性对实体对齐贡献不同的事实, 采用了一种基于图注意力网络的属性权重更新模块, 可以在训练过程中利用注意力得分不断更新每个属性的权重; 通过一个属性聚合模块将属性嵌入和属性权重信息聚合到实体嵌入中, 强化了实体的嵌入表示, 从而提升了实体对齐的效果。提出的模型在3个跨语言数据集的实验结果显示Hits@1评价指标分别为0.751, 0.805和0.915, 对齐效果均优于目前主流的实体对齐方法。