Errudite: Scalable, Reproducible, and Testable Error Analysis

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 747–763 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 747 Errudite: Scalable, Reproducible, and Testable Error Analysis Tongshuang Wu1, Marco Tulio Ribeiro2, Jeffrey Heer1, and Daniel S. Weld1 1Paul G. Allen School of Computer Science & Engineering, University of Washington 2Microsoft Research {wtshuang,jheer,weld}@cs.washington.edu marcotcr@microsoft.com Abstract Though error analysis is crucial to understand- ing 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 cod- ifies model and task agnostic principles for informative error analysis, and presents Er- rudite, an interactive tool for better support- ing this process. First, error groups should be precisely defined for reproducibility; Erru- dite 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 neg- ative examples; Errudite enables systematic grouping of relevant instances with filtering queries. Third, hypotheses about the cause of errors should be explicitly tested; Errudite sup- ports this via automated counterfactual rewrit- ing. 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 substan- tial ambiguities in prior published error analy- ses practices, and (3) enhances the error anal- ysis experience by allowing users to test and revise prior beliefs. 1 Introduction The attempt to analyze when, how, and why mod- els fail (error analysis) is a crucial part of the development cycle. Understanding model short- comings helps NLP developers revise their mod- els, uncover bugs, make deployment decisions, and communicate model performance. Two com- mon forms of error analysis are (1) data grouping, where aggregate metrics are computed for partic- ular slices of interest (e.g., accuracy over question types in machine comprehension, per-label perfor- mance in semantic role labeling) (Liu et al., 2017; He et al., 2017), and (2) counterfactual error anal- ysis, where one modifies the input data to assess if expectations are met, such as adding irrelevant data to see if new errors are introduced (Jia and Liang, 2017; Ribeiro et al., 2018). In practice, however, groupings and counter- factual tests are very coarse or limited. The in- put to most NLP tasks is unstructured text, which makes systematic in-depth error analysis challeng- ing. Even answering simple questions such as “how accurate is my model when person names are involved?” requires extensive coding, and the use of additional tools such as NER or POS tag- gers. Due to such difficulties, a common alter- native is to group a subset of error samples with manual labels on potential error causes. While useful, the high cost of manual labeling limits analyses to small samples. We surveyed 10 papers with error analyses that examine a sam- ple of incorrect predictions, e.g., (Wadhwa et al., 2018; Min et al., 2017)1, and found the sample sizes ranged from 50 …

Content truncated to 500 words