Interpretable Multi-hop Reasoning for Forecasting Future Links on Temporal Knowledge Graphs

Code

• github.com/lfxx123/tkbc
  OfficialIn paperpytorch★ 5

Abstract

Temporal knowledge graphs (KGs) have recently attracted growing attention. The temporal KG forecasting task, which plays a crucial role in applications such as event prediction, is predicting future links based on historical facts. The interpretability of the current temporal KG forecasting models is manifested in providing the reasoning paths. However, the comparison of reasoning paths is operated under the black box. Inspired by the observation that reasoning based on multi-hop paths is equivalent to answering questions step by step, this paper designs an Interpretable Multi-hop Reasoning (IMR) model for temporal KG forecasting. IMR transforms reasoning based on path searching into step-by-step question answering. Moreover, IMR designs three indicators according to the characteristics of temporal KGs and reasoning paths: question matching degree, answer completing level and path confidence. Unlike other models that can only utilize paths with a specified hop, IMR can effectively integrate paths of different hops; IMR can provide the reasoning paths like other interpretable models and further explain the basis for path comparison. While being more explainable, IMR has achieved state-of-the-art on four baseline datasets.

Tasks

Reproductions