Modeling Dynamic Entities in Temporal Knowledge Graphs

Temporal Knowledge Graphs (TKGs) have held large appeal recently and been used in many fields gradually. TKG reasoning is aimed at forecasting new facts from existing events with timestamps and it is still faced with difficulties and challenges. In terms of different tasks in TKG reasoning, the researches can be broadly classified into interpolation and extrapolation. Extrapolated TKG reasoning attempts to predict facts in the future and can be more challenging by comparison with interpolation. Most existing works focus on modeling the time information, but only a few of them are designed definitely to model dynamic entities. Therefore, we propose a method, which deals with dynamic entities explicitly with self-attention mechanism, and adopts temporal-path-based reinforcement learning to predict future events. Through experiments on commonly used datasets for link prediction tasks, we demonstrate that our method shows good performance on most of datasets and modeling dynamic entities is of effectiveness.