Entity embedding and relational path on small samples for knowledge graph completion

Abstract

In the modeling work of Knowledge Graph Completion (KGC), we propose a KGC model considering both entity embeddings and relational paths for small sample data. The relational path information between entities can identify the relative positions of two entities in the knowledge graph, and if the distance is too far from there is no need for link prediction with high probability, so limiting the number of hopes of relational path can reduce the cost of link prediction. There are many link prediction models that only consider relations, but such models are not suitable for small sample datasets because there are too few types of relations in small datasets, and considering only relations is not a good way to characterize entities, so we added entity embeddings to consider relational paths, aggregated entity neighborhoods and relational neighborhoods around entities to target entities, and finally combined entity embeddings with relational paths to perform link prediction tasks. We have tested on three small sample datasets and achieved remarkable results.