Evaluating Models’ Local Decision Boundaries via Contrast Sets

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1307–1323 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1307 Evaluating Models’ Local Decision Boundaries via Contrast Sets Matt Gardner⋆♦ Yoav ArtziΓ Victoria Basmova♦♣ Jonathan Berant♦♠ Ben Bogin♠ Sihao Chen♥ Pradeep Dasigi♦ Dheeru Dua2 Yanai Elazar♦♣ Ananth Gottumukkala2 Nitish Gupta♥ Hanna Hajishirzi♦△ Gabriel Ilharco△ Daniel Khashabi♦ Kevin Lin+ Jiangming Liu♦† Nelson F. Liu¶ Phoebe Mulcaire△ Qiang Ning♦ Sameer Singh2 Noah A. Smith♦△ Sanjay Subramanian♦ Reut Tsarfaty♦♣ Eric Wallace+ Ally ZhangΓ Ben Zhou♥ ♦Allen Institute for AI ΓCornell University ♣Bar-Ilan University ♠Tel-Aviv University ♥University of Pennsylvania △University of Washington 2UC Irvine +UC Berkeley †University of Edinburgh ¶Stanford University mattg@allenai.org Abstract Standard test sets for supervised learning eval- uate in-distribution generalization. Unfortu- nately, when a dataset has systematic gaps (e.g., annotation artifacts), these evaluations are misleading: a model can learn simple deci- sion rules that perform well on the test set but do not capture the abilities a dataset is intended to test. We propose a more rigorous annotation paradigm for NLP that helps to close system- atic gaps in the test data. In particular, after a dataset is constructed, we recommend that the dataset authors manually perturb the test in- stances in small but meaningful ways that (typ- ically) 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 lin- guistic capabilities. We demonstrate the effi- cacy of contrast sets by creating them for 10 di- verse NLP datasets (e.g., DROP reading com- prehension, UD parsing, and IMDb sentiment analysis). Although our contrast sets are not explicitly adversarial, model performance is significantly lower on them than on the origi- nal test sets—up to 25% in some cases. We re- lease our contrast sets as new evaluation bench- marks and encourage future dataset construc- tion efforts to follow similar annotation pro- cesses. 1 Introduction Progress in natural language processing (NLP) has long been measured with standard benchmark datasets (e.g., Marcus et al., 1993). These bench- marks help to provide a uniform evaluation of new modeling developments. However, recent work shows a problem with this standard evaluation paradigm based on i.i.d. test sets: datasets often ⋆ Matt Gardner led the project. All other authors are listed in alphabetical order. Two similarly-colored and similarly-posed chow dogs are face to face in one image. Two similarly-colored and similarly-posed cats are face to face in one image. Three similarly-colored and similarly-posed chow dogs are face to face in one image. Two differently-colored but similarly-posed chow dogs are face to face in one image. Original Example: Example Textual Perturbations: Two similarly-colored and similarly-posed chow dogs are face to face in one image. Example Image Perturbation: Figure 1: An example contrast set for NLVR2 (Suhr and Artzi, 2019). The label for the original example is TRUE and the label for all of the perturbed exam- ples is FALSE. The contrast set allows probing of a model’s decision …

Content truncated to 500 words