Errudite: Scalable, Reproducible, and Testable Error Analysis

By Tongshuang Sherry Wu, Marco Tulio Ribeiro, Jeffrey Heer, Daniel S. Weld. 2019

165 citations recorded by the source corpus.

Though error analysis is crucial to understanding and improving NLP models, the common practice of manual, subjective categorization of a small sample of errors can yield biased and incomplete conclusions. This paper codifies model and task agnostic principles for informative error analysis, and presents Errudite, an interactive tool for better supporting this process. First, error groups should be precisely defined for reproducibility; Errudite supports this with an expressive domain-specific language. Second, to avoid spurious conclusions, a large set of instances should be analyzed, including both positive and negative examples; Errudite enables systematic grouping of relevant instances with filtering queries. Third, hypotheses about the cause of errors should be explicitly tested; Errudite supports this via automated counterfactual rewriting. We validate our approach with a user study, finding that Errudite (1) enables users to perform high quality and reproducible error analyses with less effort, (2) reveals substantial ambiguities in prior published error analyses practices, and (3) enhances the error analysis experience by allowing users to test and revise prior beliefs.

Subfields: Artificial Intelligence, Information Systems, 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 Information Systems

In Natural Language Processing

In Evaluation & Benchmarks

Citations in the corpus

HateCheck: Functional Tests for Hate Speech Detection Models (2020, background): “Aggregate performance metrics offer limited actionable insight into specific model weak points (Wu et al., 2019).”
Tailor: Generating and Perturbing Text with Semantic Controls (2021, methodology): “…related are the creation of minimally edited datasets, either through manual rewriting (Gard-ner et al., 2020; Kaushik et al., 2020), or creating perturbation templates (Andreas, 2020; Li et al., 2020; Ribeiro et al., 2020; Wu et al., 2019); T ailor reduces the human e ﬀ orts these studies require.”
Polyjuice: Generating Counterfactuals for Explaining, Evaluating, and Improving Models (2021, methodology): “provided by control codes and [BLANK]s allow for analyses that would be non-trivial to do manually (Wu et al., 2019) or with masked language models (e.”
Beyond Accuracy: Behavioral Testing of NLP Models with CheckList (2020, background): “Further, by summarizing the performance as a single aggregate statistic, it becomes di ﬃ cult to ﬁgure out where the model is failing, and how to ﬁx it (Wu et al., 2019).”
Evaluating Models’ Local Decision Boundaries via Contrast Sets (2020, methodology): “We recommend authors categorize their perturbations up front in order to simplify future analyses and bypass some of the pitfalls of post-hoc error categorization (Wu et al., 2019).”
Evaluating NLP Models via Contrast Sets (2020, methodology): “We recommend authors categorize their perturbations up front in order to simplify future analyses and bypass some of the pitfalls of post-hoc error categorization (Wu et al., 2019).”
Evaluation Examples are not Equally Informative: How should that change NLP Leaderboards? (2021, background): “The first ameliorates the unavoidable randomness of finite evaluations while the second enables error analysis (Wu et al., 2019) and model probing (Belinkov and Glass, 2019; Zhang et al., 2019).”
TACRED Revisited: A Thorough Evaluation of the TACRED Relation Extraction Task (2020, unknown): “More similar to our approach is rewriting of instances (Jia and Liang, 2017; Ribeiro et al., 2018) but instead of evaluating model robustness we use rewriting to test explicit error hypotheses, similar to Wu et al. (2019).”
TX-Ray: Quantifying and Explaining Model-Knowledge Transfer in (Un-)Supervised NLP (2019, methodology): “Knowledge transfer in NLP is currently focused on ‘decision understanding’ [10] by analyzing supervised probing tasks [4] using either performance metrics [7, 37] or laborious perinstance explainability [3, 39].”
RECAST: Interactive Auditing of Automatic Toxicity Detection Models (2020, methodology): “eractive exploration of attention mechanisms in transformer models [12,8]. Other tools, like Errudite, allow users to view hypotheses with respect to the true error distribution on the entire dataset [20]. Some techniques attempt to adversarially perturb language input to a model [21,16], with some ”
Interpreting Predictions of NLP Models (2020, methodology): “, error analysis, Errudite (Wu et al., 2019), diagnostic “challenge” test sets (Naik et al.”
Tempura: Query Analysis with Structural Templates (2020, background): “Alternatively, they may start with a speciﬁc example that represents an experience, and try to understand if the example is part of a general trend [41].”
Automatic Construction of Evaluation Suites for Natural Language Generation Datasets (2021, methodology): “Another approach is to delegate the slicing agency to those working with a dataset through interactive systems like Errudite [22] or the Language Interpretability Tool [23].”
Robustness Gym: Unifying the NLP Evaluation Landscape (2021, background): “Erru-dite [Wu et al., 2019] supports users in exploring subpopulations of their data, while CrossCheck [Arendt et al., 2020] and Manifold [Zhang et al., 2018] focus on visualization and analysis for model comparison.”
The Language Interpretability Tool: Extensible, Interactive Visualizations and Analysis for NLP Models (2020, methodology): “In this, it is most similar to Errudite (Wu et al., 2019), which provides an integrated UI for NLP error analysis, including a custom DSL for text transformations and the ability to evaluate over a corpus.”
AI Transparency in the Age of LLMs: A Human-Centered Research Roadmap (2023, methodology): “While there has been emerging HCI work developing tools for behavioral evaluation of models [29, 200], how to extend this work to LLMs is a non-trivial question.”
SEAL : Interactive Tool for Systematic Error Analysis and Labeling (2022, background): “Identifying systematic model failures requires practitioners to think creatively about model evaluation (Ribeiro et al., 2020; Wu et al., 2019; Goel et al., 2021b; Kiela et al., 2021; Yuan et al., 2022).”