Reproduce “Identifying and Mitigating Spurious Correlations for Improving Robustness in NLP Models” 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
Tianlu Wang, Rohit Sridhar, Diyi Yang, Xuezhi Wang
Researchers at the University of Virginia, Georgia Institute of Technology, and Google Research developed an automated three-step framework for identifying spurious correlations in NLP models. This approach, which includes token extraction, cross-dataset stability analysis, and knowledge-aware perturbation, improved the precision of identified shortcuts and enhanced model robustness, generalization, and fairness by enabling targeted mitigation strategies.
Key insights
- An automated, scalable framework is crucial for proactively identifying unknown spurious correlations in NLP models, moving beyond manual or pre-defined pattern-based approaches.
- Genuine tokens exhibit stable importance across different datasets for the same task, while spurious tokens show unstable, dataset-specific importance, allowing for their differentiation.
- Targeted mitigation by masking identified spurious tokens during training or inference effectively reduces model reliance on these shortcuts, leading to improved robustness, generalization, and fairness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.