Evaluating Models' Local Decision Boundaries via Contrast Sets

Abstract

Standard test sets for supervised learning evaluate in-distribution generalization. Unfortunately, when a dataset has systematic gaps (e.g., annotation artifacts), these evaluations are misleading: a model can learn simple decision rules that perform well on the test set but do not capture a dataset's intended capabilities. We propose a new annotation paradigm for NLP that helps to close systematic gaps in the test data. In particular, after a dataset is constructed, we recommend that the dataset authors manually perturb the test instances in small but meaningful ways that (typically) change the gold label, creating contrast sets. Contrast sets provide a local view of a model's decision boundary, which can be used to more accurately evaluate a model's true linguistic capabilities. We demonstrate the efficacy of contrast sets by creating them for 10 diverse NLP datasets (e.g., DROP reading comprehension, UD parsing, IMDb sentiment analysis). Although our contrast sets are not explicitly adversarial, model performance is significantly lower on them than on the original test sets---up to 25\% in some cases. We release our contrast sets as new evaluation benchmarks and encourage future dataset construction efforts to follow similar annotation processes.

References in corpus (5)

• Pathologies of Neural Models Make Interpretations Difficult
• A Multi-Axis Annotation Scheme for Event Temporal Relations
• Cold Case: The Lost MNIST Digits
• HellaSwag: Can a Machine Really Finish Your Sentence?
• NLVR2 Visual Bias Analysis

Cited by in corpus (13)

• AI Safety for Everyone
• HateCheck: Functional Tests for Hate Speech Detection Models
• CREAK: A Dataset for Commonsense Reasoning over Entity Knowledge
• Better Robustness by More Coverage: Adversarial Training with Mixup Augmentation for Robust Fine-tuning
• Automatic Construction of Evaluation Suites for Natural Language Generation Datasets
• TORQUE: A Reading Comprehension Dataset of Temporal Ordering Questions
• Temporal Reasoning on Implicit Events from Distant Supervision
• Exploring the Efficacy of Automatically Generated Counterfactuals for Sentiment Analysis
• Ethical-Advice Taker: Do Language Models Understand Natural Language Interventions?
• Interpreting Language Models Through Knowledge Graph Extraction
• Tribrid: Stance Classification with Neural Inconsistency Detection
• NOPE: A Corpus of Naturally-Occurring Presuppositions in English
• COM2SENSE: A Commonsense Reasoning Benchmark with Complementary Sentences