metadata

annotations_creators:
  - crowdsourced
language_creators:
  - expert-generated
language:
  - zh
license:
  - cc-by-4.0
multilinguality:
  - monolingual
pretty_name: Mandarin HateCheck
size_categories:
  - 1K<n<10K
source_datasets:
  - original
task_categories:
  - text-classification
task_ids:
  - hate-speech-detection

Dataset Card for Multilingual HateCheck

Dataset Description

Multilingual HateCheck (MHC) is a suite of functional tests for hate speech detection models in 10 different languages: Arabic, Dutch, French, German, Hindi, Italian, Mandarin, Polish, Portuguese and Spanish. For each language, there are 25+ functional tests that correspond to distinct types of hate and challenging non-hate. This allows for targeted diagnostic insights into model performance.

For more details, please refer to our paper about MHC, published at the 2022 Workshop on Online Abuse and Harms (WOAH) at NAACL 2022. If you are using MHC, please cite our work!

Paper: Röttger et al. (2022) - Multilingual HateCheck: Functional Tests for Multilingual Hate Speech Detection Models. https://arxiv.org/abs/2206.09917
Repository: https://github.com/rewire-online/multilingual-hatecheck
Point of Contact: [email protected]

Dataset Structure

The csv format mostly matches the original HateCheck data, with some adjustments for specific languages.

mhc_case_id The test case ID that is unique to each test case across languages (e.g., "mandarin-1305")

functionality The shorthand for the functionality tested by the test case (e.g, "target_obj_nh"). The same functionalities are tested in all languages, except for Mandarin and Arabic, where non-Latin script required adapting the tests for spelling variations.

test_case The test case text.

label_gold The gold standard label ("hateful" or "non-hateful") of the test case. All test cases within a given functionality have the same gold standard label.

target_ident Where applicable, the protected group that is targeted or referenced in the test case. All HateChecks cover seven target groups, but their composition varies across languages.

ref_case_id For hateful cases, where applicable, the ID of the hateful case which was perturbed to generate this test case. For non-hateful cases, where applicable, the ID of the hateful case which is contrasted by this test case.

ref_templ_id The equivalent to ref_case_id, but for template IDs.

templ_id The ID of the template from which the test case was generated.

case_templ The template from which the test case was generated (where applicable).

gender_male and gender_female For gender-inflected languages (French, Spanish, Portuguese, Hindi, Arabic, Italian, Polish, German), only for cases where gender inflection is relevant, separate entries for gender_male and gender_female replace case_templ.

label_annotated A list of labels given by the three annotators who reviewed the test case (e.g., "['hateful', 'hateful', 'hateful']").

label_annotated_maj The majority vote of the three annotators (e.g., "hateful"). In some cases this differs from the gold label given by our language experts.

disagreement_in_case True if label_annotated_maj does not match label_gold for the entry.

disagreement_in_template True if the test case is generated from an IDENT template and there is at least one case with disagreement_in_case generated from the same template. This can be used to exclude entire templates from MHC.