Temporal-Relational Matching Network for Few-Shot Temporal Knowledge Graph Completion

Xing Gong, Jianyang Qin, Heyan Chai, Ye Ding, Yan Jia, Qing Liao

Abstract: Temporal knowledge graph completion (TKGC) is an important research task due to the incompleteness of temporal knowledge graphs. However, existing TKGC models face the following two issues: 1) these models cannot be directly applied to few-shot scenario where most relations have only few quadruples and new relations will be added; 2) these models cannot fully exploit the dynamic time and relation properties to generate discriminative embeddings of entities. In this paper, we propose a temporal-relational matching network, namely TR-Match, for few-shot temporal knowledge graph completion. Specifically, we design a multi-scale time-relation attention encoder to adaptively capture local and global information based on time and relation to tackle the dynamic properties problem. Then, we build a new matching processor to tackle the few-shot problem by mapping the query to few support quadruples in a relation-agnostic manner. Finally, we construct three new datasets for few-shot TKGC task based on benchmark datasets. Extensive experimental results demonstrate the superiority of our model over the state-of-the-art baselines.