Reproduce “Evaluating Models' Local Decision Boundaries via Contrast Sets” 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
Matt Gardner, Yoav Artzi, Victoria Basmova, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua and 18 more
A collaborative effort led by AI2 introduced "contrast sets," a method to evaluate NLP models' local decision boundaries by manually perturbing existing test examples. This new evaluation paradigm revealed state-of-the-art models perform substantially worse on these finely targeted examples compared to standard benchmarks.
Key insights
- Expert-driven, model-agnostic manual perturbation is crucial for creating meaningful contrast sets that accurately target intended linguistic phenomena without introducing biases towards specific model failures.
- NLP models often rely on superficial correlations and dataset artifacts rather than genuine understanding, leading to a lack of robustness when faced with subtly altered but semantically distinct inputs.
- Evaluating models on local decision boundaries via contrast sets offers a more rigorous diagnostic tool, exposing systematic gaps in model understanding that standard i.i.d. evaluations fail to reveal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.