ReCo: Reliable Multi-hop Causal Reasoning via Structural Causal Recurrent Unit

Abstract

Multi-hop causal reasoning (MCR) is an essential ability for many decision-making AI systems, which requires the model to perform multi-step causal reasoning by connecting causal event pairs. However, MCR still suffers from two main transitive problems of threshold effect and scene drift. In this paper, we propose a novel reliable multi-hop causal reasoning framework (ReCo), which introduces a structural causal recurrent unit (SCRU) to model the causal reasoning chains. In SCRU, we devise two control mechanisms to solve the threshold effect and scene drift problems, and implement a logical constraint for better optimization of ReCo. Experiments show that ReCo achieves state-of-the-art (SOTA) results on both Chinese and English multi-hop causal reasoning datasets. Finally, BERT obtains the best results on four downstream causal-related tasks by injecting reliable causal knowledge distilled from ReCo.

Tasks

Reproductions