In this work we show how large language models (LLMs) can learn statistical dependencies between otherwise unconditionally independent variables due to dataset selection bias. To demonstrate the effect, we developed a masked gender task that can be applied to BERT-family models to reveal spurious correlations between predicted gender pronouns and a variety of seemingly gender-neutral variables like date and location, on pre-trained (unmodified) BERT and RoBERTa large models. Finally, we provide an online demo, inviting readers to experiment further.
@article{arxiv.2207.08982,
title = {Selection Bias Induced Spurious Correlations in Large Language Models},
author = {Emily McMilin},
journal= {arXiv preprint arXiv:2207.08982},
year = {2022}
}
Comments
8 pages, 5 figures, Published at the ICML 2022 Workshop on Spurious Correlations, Invariance, and Stability