Generative Data Augmentation for Commonsense Reasoning
Generative Data AUGmentation for Commonsense Reasoning (G-DAUGc) is a method for generating additional training data for commonsense models, improving accuracy without requiring additional annotations. Learn more in our paper.
Here, you can view some training examples G-DAUGc produced for a commonsense task called the Winograd Schema Challenge, where the goal is to choose the word that best fits in the blank. We trained the model on a large dataset of Winograd-style questions called Winogrande, and we've mixed in some of those original examples here for comparison. Help us improve the model by guessing which is which!