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

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.