Paul/hatecheck

1---2annotations_creators:3- crowdsourced4language_creators:5- expert-generated6language:7- en8license:9- cc-by-4.010multilinguality:11- monolingual12pretty_name: HateCheck13size_categories:14- 1K<n<10K15source_datasets:16- original17task_categories:18- text-classification19task_ids:20- hate-speech-detection21---22 23# Dataset Card for HateCheck24 25## Dataset Description26 27HateCheck is a suite of functional test for hate speech detection models. 28The dataset contains 3,728 validated test cases in 29 functional tests.2919 functional tests correspond to distinct types of hate. The other 11 functional tests cover challenging types of non-hate.30This allows for targeted diagnostic insights into model performance.31 32In our ACL paper, we found critical weaknesses in all commercial and academic hate speech detection model that we tested with HateCheck. 33Please refer to the paper (linked below) for results and further discussion, as well as further information on the dataset and a full data statement.34 35- **Paper:** Röttger et al. (2021) - HateCheck: Functional Tests for Hate Speech Detection Model. https://aclanthology.org/2021.acl-long.4/ or https://arxiv.org/abs/2012.1560636- **Repository:** https://github.com/paul-rottger/hatecheck-data37- **Point of Contact:** paul.rottger@oii.ox.ac.uk38 39 40## Dataset Structure41 42"test.csv" contains all 3,728 validated test cases. Each test case (row) has the following attributes:43 44**functionality**45The shorthand for the functionality tested by the test case.46 47**case_id**48The unique ID of the test case (assigned to each of the 3,901 cases we initially generated)49 50**test_case**51The text of the test case.52 53**label_gold**54The gold standard label (hateful/non-hateful) of the test case. All test cases within a given functionality have the same gold standard label.55 56**target_ident**57Where applicable, the protected group targeted or referenced by the test case. We cover seven protected groups in the test suite: women, trans people, gay people, black people, disabled people, Muslims and immigrants.58 59**direction**60For hateful cases, the binary secondary label indicating whether they are *directed* at an individual as part of a protected group or aimed at the group in *general*.61 62**focus_words**63Where applicable, the key word or phrase in a given test case (e.g. "cut their throats").64 65**focus_lemma**66Where applicable, the corresponding lemma (e.g. "cut sb. throat").67 68**ref_case_id**69For hateful cases, where applicable, the ID of the simpler hateful case which was perturbed to generate them.70For non-hateful cases, where applicable, the ID of the hateful case which is contrasted.71 72**ref_templ_id**73The equivalent, but for template IDs.74 75**templ_id**76The unique ID of the template from which the test case was generated (assigned to each of the 866 cases and templates from which we generated the 3,901 initial cases).77 78 79## Citation Information80 81When using HateCheck, please cite our ACL paper:82 83@inproceedings{rottger-etal-2021-hatecheck,84    title = "{H}ate{C}heck: Functional Tests for Hate Speech Detection Models",85    author = {R{\"o}ttger, Paul  and86      Vidgen, Bertie  and87      Nguyen, Dong  and88      Waseem, Zeerak  and89      Margetts, Helen  and90      Pierrehumbert, Janet},91    booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",92    month = aug,93    year = "2021",94    address = "Online",95    publisher = "Association for Computational Linguistics",96    url = "https://aclanthology.org/2021.acl-long.4",97    doi = "10.18653/v1/2021.acl-long.4",98    pages = "41--58",99    abstract = "Detecting online hate is a difficult task that even state-of-the-art models struggle with. Typically, hate speech detection models are evaluated by measuring their performance on held-out test data using metrics such as accuracy and F1 score. However, this approach makes it difficult to identify specific model weak points. It also risks overestimating generalisable model performance due to increasingly well-evidenced systematic gaps and biases in hate speech datasets. To enable more targeted diagnostic insights, we introduce HateCheck, a suite of functional tests for hate speech detection models. We specify 29 model functionalities motivated by a review of previous research and a series of interviews with civil society stakeholders. We craft test cases for each functionality and validate their quality through a structured annotation process. To illustrate HateCheck{'}s utility, we test near-state-of-the-art transformer models as well as two popular commercial models, revealing critical model weaknesses.",100}101 102 103 104