ROCK: A Causal Inference Framework for Reasoning about Commonsense Causality

Abstract

Commonsense causality reasoning (CCR) aims at identifying plausible causes and effects in natural language descriptions that deemed reasonable by an average person. Although being of great academic and practical interest, this problem is still shadowed by the lack of a well-posed theoretical framework; existing work usually relies on various notions of correlation and is susceptible to confounding co-occurrences. This paper articulates the central question of CCR and develops a novel framework, ROCK, to Reason O(A)bout Commonsense K(C)ausality based on classical causal inference principles. ROCK leverages temporal signals as incidental supervision, and makes use of temporal propensities that are analogous to propensity scores for balancing confounding effects. We implement a modular zero-shot pipeline which is effective and demonstrates good potential for CCR on various datasets.

Tasks

Reproductions