Reproduce “Uncovering Latent Biases in Text: Method and Application to Peer Review” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Emaad Manzoor, Nihar B. Shah
Researchers at Carnegie Mellon University developed a causal framework for quantifying biases in unstructured text, defining bias as the disparity in language used for different demographic groups specifically caused by the visibility of identity indicators. Applying this methodology to ICLR peer review data, they detected a causal affiliation bias, where reviews for papers from USA/Canada institutions used more positive language when author identities were visible, a finding consistent with observed biases in review ratings.
Key insights
- Bias can be formalized and quantified as a causal estimand by comparing language disparities under different identity visibility conditions.
- Natural experiments, such as changes in peer review policies, provide a valuable means to validate text-based bias detection frameworks against 'ground truth' measures.
- A nonparametric Difference-in-Differences approach, adapted for unstructured text, can effectively identify causal disparities in language without requiring labeled biased examples.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.