Abstract:

The purpose of knowledge graph completion is to find the missing links in triples and solve the problem of sparse data in knowledge graph. This paper proposes a knowledge graph embedding method based on capsule network, which can model relational triples（head entities, relations, tail entities）. Firstly, the triple is represented as a 3-column matrix, which is convolved with multiple filters to produce different feature maps. Secondly, these feature maps are reconstructed into corresponding capsules, each capsule is a group of neurons, and a smaller size capsule is generated by the weighted product, and then a continuous vector is generated. Finally, the vector and the weight vector are subjected to a dot product operation to obtain a corresponding score, and the results of all the score summations are used to determine the correctness of the given triple. The experimental results show that compared with other models, the proposed method effectively improves the prediction accuracy of the triples, and the knowledge graph completion is better.

Key words: knowledge graph, knowledge graph completion, link prediction, capsule network

摘要：

知识图谱补全旨在发现三元组中缺失链接，解决知识图谱数据稀疏问题。提出一种基于胶囊网络的知识图谱嵌入方法，该方法能够对关系三元组（头实体，关系，尾实体）进行建模。将三元组表示为3列矩阵，它与多个滤波器卷积以产生不同的特征映射；将这些特征图重建成相应的胶囊，每个胶囊是一组神经元，通过和权重点积生成较小尺寸的胶囊，然后生成一个连续矢量；该矢量和权重向量进行点积运算获得对应得分，所有分数求和的结果用来判断给定三元组的正确性。实验结果表明，和其他模型相比，该方法有效提高了三元组的预测精度，知识图谱补全的效果更好。

关键词: 知识图谱, 知识图谱补全, 链接预测, 胶囊网络