A path-based relation networks model for knowledge graph completion

Abstract

We consider the problem of learning and inference in a large-scale knowledge graph containing incomplete knowledge. We show that a simple neural network module for relational reasoning through the path extracted from the knowledge base can be used to reliably infer new facts for the missing link. In our work, we used path ranking algorithm to extract the relation path from knowledge graph and use it to build train data. In order to learn the characteristics of relation, a detour path between nodes was created as training data using the extracted relation path. Using this, we trained a model that can predict whether a given triple (Head entity, relation, tail entity) is valid or not. Experiments show that our model obtains better link prediction, relation prediction and triple classification results than previous state-of-the-art models on benchmark datasets WN18RR, FB15k-237, WN11 and FB13.