基于双向GCN和CVm的实体对齐模型研究

实体对齐旨在发现并链接不同知识图谱中指向现实世界的相同实体对象.针对基于图卷积网络的实体对齐通常作用于单一关系类型的无向图,容易导致对应实体学习的嵌入结果不一致问题,构建了一种基于双向图卷积网络和变异系数法的实体对齐模型.该模型通过拆分非对称邻接权重矩阵构建双向图卷积网络方法,学习实体前后向隐藏特征,实现实体的完整表示;同时通过变异系数法为属性加权,选择最有代表性的实体局部语义信息,有效提高实体对齐精确度.通过在两组大型真实异构数据集上对模型进行验证,实验结果表明,该方法与现有基于嵌入的实体对齐方法相比Hit@1值平均提高了4％,同时保持较高的平均倒数秩,在一定程度上可以提高实体对齐效果.

Entity alignment model based on two-way gcn and cvm

Entity alignment aims to discover and link the same entity objects that point to the real word in different knowledge graphs. Entity alignment based on GCN usually acts on undirected graphs of the single relation type, which is easy to result in the problem of inconsistent embedding corresponding results to entity leaning. Therefore, this paper proposed an entity alignment model based on two-way GCN and CVm. It realized by the complete representation of the entity splitting the asymmetric adjacency weight matrix to construct a two-way GCN, so that the model could be learned the forward and backward hidden features of the entity. At the same time, for selecting the most representative entity local semantic information, it used the CVm to weight attributes, which effectively improved the accuracy of entity alignment. By verifying the model on two large real heterogeneous datasets, the Hit@1 value of this method was 4% higher than the existing embedding-based entity alignment method on average and maintains a high average reciprocal rank. It is proved that this method improves the entity alignment effect to a certain extent.