Knowledge graph embedding using network structure

Abstract

In knowledge graph (KG), individual knowledge is represented by three values (triples) of “the value (object) of r (predicate) is o for s (subject)”. Assuming that the subject s and the object o are nodes indicating the actual situation, and the predicate r is edge indicating their relationship, the entire knowledge group is represented by a knowledge graph. On the other hand, if the nodes and edges are represented by multidimensional vectors (embedding), it is possible to predict whether a triple is true or false, and efforts have been made in recent years to make this useful. Here, when a knowledge graph learns from data, it often takes much computation time to learn large-scale data such as WordNet. Therefore, this study aims to reduce the learning time. We pondered various features of the network and found that if there are module structures, it is possible to shorten the learning time. In particular, the identification and utilization of entities that affect the learning of other entities contribute to the learning efficiency. This high-speed learning method of embedding the knowledge graph is useful especially for the situation when a new KG model was developed and need to conduct multiple learning to find the best hyperparameter set.