Improving Transitive Embeddings in Neural Reasoning Tasks via Knowledge-Based Policy Networks

This paper proposes an approach to embed ontologies in order to deal with reasoning based on transitive relations, using the datasets provided for the SemRec Challenge at ISWC 2022. Knowledge Graph Embedding (KGE) methods provide a low-dimensional representation of the entities and relationships extracted from the knowledge graph and have been successfully used for a variety of applications such as question answering, reasoning, inference, and link prediction. However, most KGE methods cannot handle the underlying constraints and characteristics of ontologies, preventing them from performing important reasoning tasks such as subsumption and instance checking. We propose to extend translation-based embedding methods to include subsumption and instance checking reasoning by leveraging transitive relations. Experimental results show that our approach can achieve Hits@10 as high as %73 using samples generated by a policy network.