Modeling Unseen Entities from a Semantic Evidence View in Temporal Knowledge Graphs

Abstract

Temporal Knowledge Graphs (TKGs), which model dynamic events along the timeline, have attracted much attention in recent years. For temporal knowledge graphs suffer from incompleteness, quite a lot of researches are devoted to TKG reasoning, attempting to predict missing temporal facts from known events. Depending on different tasks of TKG reasoning, existing work can be divided into two categories: interpolation and extrapolation. Compared with interpolation, extrapolation faces more challenges and difficulties, which is intended to forecast facts in the future. It’s worth noting that dealing with unseen entities effectively is one of important challenges, while only a few of prior work focuses on handling unseen data explicitly. To this end, we model unseen entities from semantic perspectives in this work, combining with temporal-path-based reinforcement learning, which guarantees the interpretability of reasoning. Through extensive experiments on standard TKG tasks, we verify that our model performs well on link prediction task at future timestamps, demonstrating better extrapolation capabilities.