Evaluating NLP Models via Contrast Sets

By Matt Gardner, Yoav Artzi, Victoria Basmova, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala and 16 more. 2020

85 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

In Evaluation & Benchmarks

Citations in the corpus

UnifiedQA: Crossing Format Boundaries With a Single QA System (2020, methodology): “Additionally, we use contrastsets (Gardner et al., 2020) corresponding to several of our datasets (indicated by “CS” tag): BoolQ-CS, ROPES-CS, Quoref-CS, DROP-CS.”
The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes (2020, background): “This strategy is compatible with recent work arguing that tasks should include counterfactual or contrastive examples [38, 22, 26].”
Aligning AI With Shared Human Values (2020, methodology): “For most tasks we use “counterfactual augmentations” (Kaushik et al., 2020) or “contrast set” examples (Gardner et al., 2020), for which examples with different labels are collected simultaneously while enforcing that the scenarios are similar.”
Learning from Task Descriptions (2020, background): “Attempts to generalize semantic parsing to more open domains are still nascent (Chen et al., 2020; Gupta et al., 2020).”
Text Modular Networks: Learning to Decompose Tasks in the Language of Existing Models (2021, methodology): “We evaluate on the DROP contrast set (Gardner et al., 2020), a suite of testonly examples created for assessing robustness via minimally perturbed examples.”
Formalizing Trust in Artificial Intelligence: Prerequisites, Causes and Goals of Human Trust in AI (2020, methodology): “ Factsheets (security) Adversarial attacks and defenses [20] Fallback plan and general safety - - A high level of accuracy Model cards (metrics) - Reliability Factsheets (concept drift) Contrast sets [17], behavioral testing [58] Reproducibility Reproducibility checklists “Show your work” [14] Priva”
Compositional Explanations of Neurons (2020, background): “…of the image-patch attacks of [9], adversarial NLI inputs [1, 37], and the data collection process for recent counterfactual NLI datasets [13, 22], but instead of searching among neuron visualizations or using black-box optimization to synthesize examples, we use explanations as a…”
Directions in abusive language training data, a systematic review: Garbage in, garbage out (2020, background): “Synthetic data can lead to better understanding of model limitations and robustness, such as by ‘perturbing’ or ‘contrasting’ abuse to produce subtly different varieties or to adjust the content just enough to change the primary label [111].”
Customizing Triggers with Concealed Data Poisoning (2020, unknown): “For instance, work that exposes annotation biases in datasets (Gururangan et al., 2018) has lead to new data collection processes and training algorithms (Gardner et al., 2020; Clark et al., 2019).”
Linguistically-Informed Transformations (LIT): A Method for Automatically Generating Contrast Sets (2020, background): “Contrast Sets Contrast sets (Gardner et al., 2020) serve to evaluate a models’ true capabilities by evaluating on out-of-distribution data since previous in-distribution test sets often have systematic gaps, which inﬂate models’ performance on a task (Gururangan et al., 2018; Geva et al., 2019).”
Robustness Gym: Unifying the NLP Evaluation Landscape (2021, methodology): “Predeﬁned out-of-distribution data splits may be used to evaluate a model’s ability to generalize outside of the speciﬁc dataset set on which it was trained [Gardner et • Detecting bias.”
♫ MuSiQue: Multihop Questions via Single-hop Question Composition (2021, background): “Intuitively, because the answerable and unanswerable instances are very similar but have different labels, it’s difficult for models to do well on both instances if they learn to rely on shortcuts (Kaushik et al., 2019; Gardner et al., 2020).”
Asking without Telling: Exploring Latent Ontologies in Contextual Representations (2020, background): “Recent work also advocates for counterfactual learning and evaluation (Qin et al., 2019; Kaushik et al., 2020) to mitigate confounds, or contrastive evaluation sets (Gardner et al., 2020) to rigorously test local decision boundaries.”
Are We Learning Yet? A Meta Review of Evaluation Failures Across Machine Learning (2021, background): “Landmark studies found that language models regularly exploit such “spurious patterns” across a wide range of NLP tasks [46, 72].”
TORQUE: A Reading Comprehension Dataset of Temporal Ordering Questions (2020, methodology): “We report the standard macro F 1 and exact-match (EM) metrics in question answering, and also EM consistency, the percentage of contrast question sets for which a model’s predictions match exactly to all questions in a group (Gardner et al., 2020).”