The task of identifying and reasoning with circumstantial preconditions associated with everyday facts is natural to humans. It is unclear whether state-of-the-art language models (LMs) understand the implicit preconditions that enable or invalidate commonsense facts, such as A glass is used for drinking water, Despite their impressive accuracy on existing commonsense tasks. In this paper, we propose a new problem of reasoning with circumstantial preconditions, and present a dataset, called CoreQuisite, which annotates commonsense facts with preconditions expressed in natural language. Based on this resource, we create three canonical evaluation tasks and use them to examine the capability of existing LMs to understand situational pre-conditions. Our results show that there is a 10-30%gap between machine and human performance on our tasks. We make all resources and software publicly available.
Download