Cold-Start Multi-hop Reasoning by Hierarchical Guidance and Self-verification

Mayi Xu, Ke Sun, Yongqi Li, Tieyun Qian

Abstract: Multi-hop reasoning has attracted wide attention for knowledge graph (KG) completion since it can provide interpretable reasoning paths. Most prior multi-hop reasoning studies assume the KGs are static with fixed entities. However, in real applications, KGs are often dynamic since new entities will emerge continuously in the form of new fact triplets. In this paper, we are particularly interested in the cold-start scenario toward dynamic KGs to facilitate more practical multi-hop reasoning, which aims to explore the reasoning paths between emerging entities and existing entities. There are two challenging issues arising from this scenario: i) lacking precise guidance since available information for emerging entities is extremely limited in the cold-start scenario, ii) lacking explicit path since the emerging entities and existing ones are isolated. To address these issues, we propose a generation-based model, namely SelfHier, to explore the reasoning paths by hierarchical guidance and self-verification strategies. The hierarchical guidance strategy guides the reasoning process using hierarchical fine-grained sub-relations and coarse-grained clusters. The self-verification strategy constructs explicit reasoning paths by supplementing some missing fact triplets. Experimental results prove that SelfHier performs well in the cold-start scenario on dynamic KGs and also significantly outperforms existing multi-hop reasoning methods in the standard scenario on static KGs.