Attention-based Learning for Multiple Relation Patterns in Knowledge Graph Embedding

Tengwei Song, Jie Luo

Abstract: Relations in knowledge graphs often exhibit multiple relation patterns. Various knowledge graph embedding methods have been proposed to modelling properties in relation patterns. However, relations with a certain relation pattern actually only account for a small proportion in the knowledge graph. Relations with no explicit relation patterns also show complicated properties which is rarely studied. To this end, we argue that a property of a relation should either be global or be partial, and propose an Attention-based Learning framework for Multi-relation Patterns (ALMP) for expressing complex properties of relations. ALMP adopts a set of affine transformations to express corresponding global relation properties. Furthermore, ALMP utilizes a module of attention mechanism to integrate the representations. Experimental results show that ALMP outperforms baseline models on the link prediction task.