Temporal knowledge completion with context-aware embeddings

Yu Liu, Wen Hua, Jianfeng Qu, Kexuan Xin, Xiaofang Zhou

Abstract: Temporal knowledge graph embedding can be used to improve the coverage of temporal KGs via link predictions. Most existing works only concentrate on the target facts themselves, regardless of the rich and informative interactions between the target facts and their highly-related contexts. In this paper, we propose a novel approach to take advantage of useful contextual interactions from two aspects, namely temporal consistency and contextual consistency. More specifically, temporal consistency measures how well the target fact interacts with its surrounding contexts in the temporal dimension, while contextual consistency treats all facts as a whole integrity and captures the semantic interactions between multiple contexts. Additionally, considering the existence of useless and misleading context information, we design a crafted context selection strategy to pick out the most useful contexts with reference to the target facts, and then encode them using deep neural networks to capture the temporal and semantic interactions. Experimental results on real world datasets verify the effectiveness of our proposals comparing with competitive KGE methods and temporal KGE methods.