Language selection

Government of Canada / Gouvernement du Canada

Search

Necessity and sufficiency for explaining text classifiers: a case study in hate speech detection

From National Research Council Canada

Download
  1. (PDF, 407 KiB)
DOIResolve DOI: https://doi.org/10.18653/v1/2022.naacl-main.192
AuthorSearch for: 1; Search for: 1; Search for: 1; Search for: 1
Affiliation
  1. National Research Council of Canada. Digital Technologies
FormatText, Article
ConferenceProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL 2022), July 10-15, 2022, Seattle, Washington (Hybrid Conference)
Abstract
We present a novel feature attribution method for explaining text classifiers, and analyze it in the context of hate speech detection. Although feature attribution models usually provide a single importance score for each token, we instead provide two complementary and theoretically-grounded scores – necessity and sufficiency – resulting in more informative explanations. We propose a transparent method that calculates these values by generating explicit perturbations of the input text, allowing the importance scores themselves to be explainable. We employ our method to explain the predictions of different hate speech detection models on the same set of curated examples from a test suite, and show that different values of necessity and sufficiency for identity terms correspond to different kinds of false positive errors, exposing sources of classifier bias against marginalized groups.
Publication date
PublisherAssociation for Computational Linguistics (ACL)
Licence
In
  • Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL 2022) (15 July 2022): 2672–2686. .
LanguageEnglish
Peer reviewedYes
Export citationExport as RIS
Report a correctionReport a correction (opens in a new tab)
Record identifier9c698f67-b648-4edb-866a-4106252f0b1b
Record created2022-09-09
Record modified2022-09-12
Date modified: