Human curated vs. machine generated WinoGrad schema examples.

A simple approach to generate synthetic examples using pretrained language models (GPT2, finetuned on the task corpus), and selecting the most informative and diverse set of examples for data augmentation.

Also see our project page hosted by AI2. Our paper was accepted to the Findings in EMNLP:

@inproceedings{Yang2020GDAug,
title={G-DAUG: Generative Data Augmentation for Commonsense Reasoning},
author={Yiben Yang and Chaitanya Malaviya and Jared Fernandez and Swabha Swayamdipta and Ronan Le Bras and J. Wang and Chandra Bhagavatula and Yejin Choi and Doug Downey},
booktitle={Findings of EMNLP},
year={2020},
url={https://arxiv.org/abs/2004.11546}
}