Learning to Compose Relational Embeddings in Knowledge Graphs

Abstract

Knowledge Graph Embedding methods learn low-dimensional representations for entities and relations in knowledge graphs, which can be used to infer previously unknown relations between pairs of entities in the knowledge graph. This is particularly useful for expanding otherwise sparse knowledge graphs. However, the relation types that can be predicted using knowledge graph embeddings are confined to the set of relations that already exists in the KG. Often the set of relations that exist between two entities are not independent, and it is possible to predict what other relations are likely to exist between two entities by composing the embeddings of the relations in which each entity participates. We introduce relation composition as the task of inferring embeddings for unseen relations by combining existing relations in a knowledge graph. Specifically, we propose a supervised method to compose relational embeddings for novel relations using pre-trained relation embeddings for existing relations. Our experimental results on a previously proposed benchmark dataset for relation composition ranking and triple classification show that the proposed supervised relation composition method outperforms several unsupervised relation composition methods.