Selection Bias Induced Spurious Correlations in Large Language Models
This highlights a critical bias issue in LLMs that can affect fairness and reliability in NLP applications.
The paper demonstrates that large language models (LLMs) learn spurious correlations between gender pronouns and neutral variables like date and location due to dataset selection bias, using a masked gender task on pre-trained BERT and RoBERTa models.
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.