Enhancing Zero-shot Commonsense Reasoning by Integrating Visual Knowledge via Machine Imagination

Researchers propose 'Imagine,' a new zero-shot commonsense reasoning framework that enhances Pre-trained Language Models by integrating machine-generated visual signals into the reasoning pipeline. The approach demonstrates superior performance over existing zero-shot methods and even advanced large language models by addressing human reporting biases through machine imagination.