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

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.