From Pretraining Data to Language Models to Downstream Tasks: Tracking the Trails of Political Biases Leading to Unfair NLP Models https://aclanthology.org/2023.acl-long.656.pdf