hate speech words list

hatespeechdata

Catalog of abusive language data (plos 2020), hate speech dataset catalogue.

This page catalogues datasets annotated for hate speech, online abuse, and offensive language. They may be useful for e.g. training a natural language processing system to detect this language.

The list is maintained by Leon Derczynski , Bertie Vidgen , Hannah Rose Kirk , Pica Johansson, Yi-Ling Chung , Mads Guldborg Kjeldgaard Kongsbak, Laila Sprejer , and Philine Zeinert.

We provide a list of datasets and keywords . If you would like to contribute to our catalogue or add your dataset, please see the instructions for contributing .

If you use these resources, please cite (and read!) our paper: Directions in Abusive Language Training Data: Garbage In, Garbage Out . And if you would like to find other resources for researching online hate, visit The Alan Turing Institute’s Online Hate Research Hub or read The Alan Turing Institute’s Reading List on Online Hate and Abuse Research .

If you’re looking for a good paper on online hate training datasets (beyond our paper, of course!) then have a look at ‘Resources and benchmark corpora for hate speech detection: a systematic review’ by Poletto et al. in Language Resources and Evaluation .

Please send contributions via github pull request. You can do this by visiting the source code on github and clicking the edit icon (a pencil, above the text, on the right) - more details below . There’s a commented-out markdown template at the top of this file. Accompanying data statements preferred for all corpora.

Datasets Table of Contents

List of datasets, detecting abusive albanian.

Link to publication: https://arxiv.org/abs/2107.13592
Link to data: https://doi.org/10.6084/m9.figshare.19333298.v1
Task description: Hierarchical (offensive/not; untargeted/targeted; person/group/other)
Details of task: Detect and categorise abusive language in social media data
Size of dataset: 11 874
Percentage abusive: 13.2%
Language: Albanian
Level of annotation: Posts
Platform: Instagram, Youtube
Medium: Text
Reference: Nurce, E., Keci, J., Derczynski, L., 2021. Detecting Abusive Albanian. arXiv:2107.13592
Dataset reader: 🤗 strombergnlp/shaj

Let-Mi: An Arabic Levantine Twitter Dataset for Misogynistic Language

Link to publication: https://arxiv.org/abs/2103.10195
Link to data: https://drive.google.com/file/d/1mM2vnjsy7QfUmdVUpKqHRJjZyQobhTrW/view
Task description: Binary (misogyny/none) and Multi-class (none, discredit, derailing, dominance, stereotyping & objectiﬁcation, threat of violence, sexual harassment, damning)
Details of task: Introducing an Arabic Levantine Twitter dataset for Misogynistic language
Size of dataset: 6,603 direct tweet replies
Percentage abusive: 48.76%
Language: Arabic
Platform: Twitter
Reference: Hala Mulki and Bilal Ghanem. 2021. Let-Mi: An Arabic Levantine Twitter Dataset for Misogynistic Language. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 154–163, Kyiv, Ukraine (Virtual). Association for Computational Linguistics

Are They our Brothers? Analysis and Detection of Religious Hate Speech in the Arabic Twittersphere

Link to publication: https://ieeexplore.ieee.org/document/8508247
Link to data: https://github.com/nuhaalbadi/Arabic_hatespeech
Task description: Binary (Hate, Not)
Details of task: Religious subcategories
Size of dataset: 6,136
Percentage abusive: 0.45
Reference: Albadi, N., Kurdi, M. and Mishra, S., 2018. Are they Our Brothers? Analysis and Detection of Religious Hate Speech in the Arabic Twittersphere. In: International Conference on Advances in Social Networks Analysis and Mining. Barcelona, Spain: IEEE, pp.69-76.

Multilingual and Multi-Aspect Hate Speech Analysis (Arabic)

Link to publication: https://arxiv.org/abs/1908.11049
Link to data: https://github.com/HKUST-KnowComp/MLMA_hate_speech
Task description: Detailed taxonomy with cross-cutting attributes: Hostility, Directness, Target Attribute, Target Group, How annotators felt on seeing the tweet.
Details of task: Gender, Sexual orientation, Religion, Disability
Size of dataset: 3,353
Percentage abusive: 0.64
Reference: Ousidhoum, N., Lin, Z., Zhang, H., Song, Y. and Yeung, D., 2019. Multilingual and Multi-Aspect Hate Speech Analysis. ArXiv,.

L-HSAB: A Levantine Twitter Dataset for Hate Speech and Abusive Language

Link to publication: https://www.aclweb.org/anthology/W19-3512
Link to data: https://github.com/Hala-Mulki/L-HSAB-First-Arabic-Levantine-HateSpeech-Dataset
Task description: Ternary (Hate, Abusive, Normal)
Details of task: Group-directed + Person-directed
Size of dataset: 5,846
Percentage abusive: 0.38
Reference: Mulki, H., Haddad, H., Bechikh, C. and Alshabani, H., 2019. L-HSAB: A Levantine Twitter Dataset for Hate Speech and Abusive Language. In: Proceedings of the Third Workshop on Abusive Language Online. Florence, Italy: Association for Computational Linguistics, pp.111-118.

Abusive Language Detection on Arabic Social Media (Twitter)

Link to publication: https://www.aclweb.org/anthology/W17-3008
Link to data: http://alt.qcri.org/~hmubarak/offensive/TweetClassification-Summary.xlsx
Task description: Ternary (Obscene, Offensive but not obscene, Clean)
Details of task: Incivility
Size of dataset: 1,100
Percentage abusive: 0.59
Reference: Mubarak, H., Darwish, K. and Magdy, W., 2017. Abusive Language Detection on Arabic Social Media. In: Proceedings of the First Workshop on Abusive Language Online. Vancouver, Canada: Association for Computational Linguistics, pp.52-56.
Dataset reader: 🤗 strombergnlp/offenseval_2020

Abusive Language Detection on Arabic Social Media (Al Jazeera)

Link to data: http://alt.qcri.org/~hmubarak/offensive/AJCommentsClassification-CF.xlsx
Size of dataset: 32,000
Percentage abusive: 0.81
Platform: AlJazeera

Dataset Construction for the Detection of Anti-Social Behaviour in Online Communication in Arabic

Link to publication: https://www.sciencedirect.com/science/article/pii/S1877050918321756
Link to data: https://onedrive.live.com/?authkey=!ACDXj_ZNcZPqzy0&id=6EF6951FBF8217F9!105&cid=6EF6951FBF8217F9
Task description: Binary (Offensive, Not)
Size of dataset: 15,050
Percentage abusive: 0.39
Platform: YouTube
Reference: Alakrot, A., Murray, L. and Nikolov, N., 2018. Dataset Construction for the Detection of Anti-Social Behaviour in Online Communication in Arabic. Procedia Computer Science, 142, pp.174-181.

Hate Speech Detection in the Bengali language: A Dataset and its Baseline Evaluation

Link to publication: https://arxiv.org/pdf/2012.09686.pdf
Link to data: https://www.kaggle.com/naurosromim/bengali-hate-speech-dataset
Task description: Binary (hateful, not)
Details of task: Several categories: sports, entertainment, crime, religion, politics, celebrity and meme
Size of dataset: 30,000
Percentage abusive: 0.33
Language: Bengali
Platform: Youtube and Facebook
Reference: Romim, N., Ahmed, M., Talukder, H., & Islam, M. S. (2021). Hate speech detection in the bengali language: A dataset and its baseline evaluation. In Proceedings of International Joint Conference on Advances in Computational Intelligence (pp. 457-468). Springer, Singapore.

SWSR: A Chinese Dataset and Lexicon for Online Sexism Detection

Link to publication: https://www.sciencedirect.com/science/article/abs/pii/S2468696421000604#fn1
Link to data: https://doi.org/10.5281/zenodo.4773875
Task description: Binary (Sexist, Non-sexist), Categories of sexism (Stereotype based on Appearance, Stereotype based on Cultural Background, MicroAggression, and Sexual Offense), Target of sexism (Individual or Generic)
Details of task: Sexism detection on social media in Chinese
Size of dataset: 8,969 comments from 1,527 weibos
Percentage abusive: 34.5%
Language: Chinese
Platform: Sina Weibo
Reference: Aiqi Jiang, Xiaohan Yang, Yang Liu, Arkaitz Zubiaga, SWSR: A Chinese dataset and lexicon for online sexism detection, Online Social Networks and Media, Volume 27, 2022, 100182, ISSN 2468-6964.

CoRAL: a Context-aware Croatian Abusive Language Dataset

Link to publication: https://aclanthology.org/2022.findings-aacl.21/
Link to data: https://github.com/shekharRavi/CoRAL-dataset-Findings-of-the-ACL-AACL-IJCNLP-2022
Task description: Multi-class based on context dependency categories (CDC)
Details of task: Detectioning CDC from abusive comments
Size of dataset: 2,240
Percentage abusive: 100%
Language: Croatian
Platform: Newspaper comments
Reference: Ravi Shekhar, Mladen Karan and Matthew Purver (2022). CoRAL: a Context-aware Croatian Abusive Language Dataset. Findings of the ACL: AACL-IJCNLP.

Datasets of Slovene and Croatian Moderated News Comments

Link to publication: https://www.aclweb.org/anthology/W18-5116
Link to data: http://hdl.handle.net/11356/1202
Task description: Binary (Deleted, Not)
Details of task: Flagged content
Size of dataset: 17,000,000
Percentage abusive: 0.02
Platform: 24sata website
Reference: Ljubešić, N., Erjavec, T. and Fišer, D., 2018. Datasets of Slovene and Croatian Moderated News Comments. In: Proceedings of the 2nd Workshop on Abusive Language Online (ALW2). Brussels, Belgium: Association for Computational Linguistics, pp.124-131.

Automating News Comment Moderation with Limited Resources: Benchmarking in Croatian and Estonian

Link to publication: https://jlcl.org/content/2-allissues/1-heft1-2020/jlcl_2020-1_3.pdf
Link to data: https://www.clarin.si/repository/xmlui/handle/11356/1399
Task description: Multi-class based on Different rules
Details of task: Flagged content performmed by the real newspaper moderators
Size of dataset: 21M
Percentage abusive: 7.8%
Reference: Ravi Shekhar, Marko Pranjić, Senja Pollak, Andraž Pelicon, Matthew Purver (2020). Automating News Comment Moderation with Limited Resources: Benchmarking in Croatian and Estonian. Journal for Language Technology and Computational Linguistics (JLCL).

Offensive Language and Hate Speech Detection for Danish

Link to publication: http://www.derczynski.com/papers/danish_hsd.pdf
Link to data: https://figshare.com/articles/Danish_Hate_Speech_Abusive_Language_data/12220805
Task description: Branching structure of tasks: Binary (Offensive, Not), Within Offensive (Target, Not), Within Target (Individual, Group, Other)
Size of dataset: 3,600
Percentage abusive: 0.12
Language: Danish
Platform: Twitter, Reddit, newspaper comments
Reference: Sigurbergsson, G. and Derczynski, L., 2019. Offensive Language and Hate Speech Detection for Danish. ArXiv.
Dataset reader: 🤗 DDSC/dkhate

BAJER: Misogyny in Danish

Link to publication: https://aclanthology.org/2021.acl-long.247/
Link to data: request here
Task description: Hierarchy of abusive content labels including subcategories of misogyny
Details of task: “Misogyny detection on social media in Danish”
Size of dataset: 27.9K comments
Percentage abusive: 7% misogynistic, 27% abusive (i.e. 20% abusive but not misogyny)
Level of annotation: Social media post / comment
Platform: Twitter, Facebook, Reddit
Medium: text
Reference: Zeinert, Inie, & Derczynski, 2021. “Annotating Online Misogyny”. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL
Dataset reader: 🤗 strombergnlp/bajer_danish_misogyny

The Dutch Abusive Language Corpus v1.0 (DALC v1.0)

Link to publication: https://aclanthology.org/2021.woah-1.6.pdf - link to the documentation and/or a data statement about the data
Link to data: https://github.com/tommasoc80/DALC
Task description: Multilayered (explicitness and target) for abusive language
Details of task: Abusive language detection in social media in Dutch
Size of dataset: 8,156 tweets
Percentage abusive: 15.06% explicitly abusive; 8.09% implicitly abusive
Language: Dutch
Level of annotation: tweets
Reference: Caselli, T., Schelhaas, A., Weultjes, M., Leistra, F., van der Veen, H., Timmerman, G., and Nissim, M. 2021. “DALC: the Dutch Abusive Language Corpus”. Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), ACL.

Not All Counterhate Tweets Elicit the Same Replies: A Fine-Grained Analysis

Link to publication: https://aclanthology.org/2023.starsem-1.8/
Link to data: https://github.com/albanyan/counterhate_reply
Task description: Four binary classification tasks to investigate replies to counterhate tweets (1) Binary (Agree, Not), (2) Binary (Support_Hateful-tweet, Not), (3) Binary (Attack_Author, Not), and (4) Binary (Additional_Counterhate, Not)
Details of task: Three levels of tweets are considered: a hateful tweet, a counterhate tweet (a reply to a hateful tweet), and all replies to the counterhate tweet. Indicate whether the reply to a counterhate tweet (a) agrees with the counterhate tweet, (b) supports the hateful tweet, (c) attacks the author of the counterhate tweet, and (d) adds additional counterhate
Size of dataset: 2,621 (hateful tweet, counterhate tweet, reply) triples
Percentage abusive: 100% (All main tweets are hateful tweets)
Language: English
Level of annotation: Tweets
Reference: Abdullah Albanyan, Ahmed Hassan, and Eduardo Blanco. 2023. Not All Counterhate Tweets Elicit the Same Replies: A Fine-Grained Analysis. In Proceedings of the 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 71–88, Toronto, Canada. Association for Computational Linguistics.

Pinpointing Fine-Grained Relationships between Hateful Tweets and Replies

Link to publication: https://ojs.aaai.org/index.php/AAAI/article/view/21284
Link to data: https://github.com/albanyan/hateful-tweets-replies
Task description: Four binary classification tasks (1) Binary (Counterhate, Not), (2) Binary (Counterhate_with_Justification, Not), (3) Binary (Attack_Author, Not), and (4) Binary (Additional_Hate, Not)
Details of task: Indicate whether the reply to a hateful tweet (a) is counter hate speech, (b) provides a justification, (c) attacks the author of the tweet, and (d) adds additional hate
Size of dataset: 5,652 hateful tweets and replies
Reference: Abdullah Albanyan and Eduardo Blanco. 2022. Pinpointing Fine-Grained Relationships Between Hateful Tweets and Replies. Proceedings of the AAAI Conference on Artificial Intelligence 36 (10):10418-26.

Large-Scale Hate Speech Detection with Cross-Domain Transfer

Link to publication: https://aclanthology.org/2022.lrec-1.238/
Link to data: https://github.com/avaapm/hatespeech
Task description: Three-class (Hate speech, Offensive language, None)
Details of task: Hate speech detection on social media (Twitter) including 5 target groups (gender, race, religion, politics, sports)
Size of dataset: 100k English (27593 hate, 30747 offensive, 41660 none)
Percentage abusive: 58.3%
Medium: Text and image
Reference: Cagri Toraman, Furkan Şahinuç, Eyup Yilmaz. 2022. Large-Scale Hate Speech Detection with Cross-Domain Transfer. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2215–2225, Marseille, France. European Language Resources Association.
#### Online-Abusive-Attacks-OAA-Dataset
Link to publication: https://ieeexplore.ieee.org/abstract/document/10160004
Link to data: https://github.com/RaneemAlharthi/Online-Abusive-Attacks-OAA-Dataset
Task description: “Binary (abusive, Notabusive)”, “Hierarchical”, “six-class (toxicity, severe toxicity, identity attack,insult, profanity, and threat)”
Details of task: “the first benchmark dataset providing a holistic view of online abusive attacks, including social media profile data and metadata for both targets and perpetrators, in addition to context. The dataset contains 2.3K Twitter accounts, 5M tweets, and 106.9K categorised conversations.”
Size of dataset: 2.3K Twitter accounts, 5M tweets, and 106.9K categorised conversations.
Percentage abusive: online abusive attacks motivated by the targets’ identities (97%), and motivated by the targets’ behavioural attacks (3%).
Language: e.g. English
Level of annotation: What is an “instance”, in this dataset? e.g. Conversation
Platform: e.g. twitter
Medium: text /metadata
Reference: @article{alharthi2023target, title={Target-Oriented Investigation of Online Abusive Attacks: A Dataset and Analysis}, author={Alharthi, Raneem and Alharthi, Rajwa and Shekhar, Ravi and Zubiaga, Arkaitz}, journal={IEEE Access}, year={2023}, publisher={IEEE} }
Link to publication: https://aclanthology.org/2021.emnlp-main.587/
Link to data: https://github.com/amandacurry/convabuse
Task description: Hierarchical: 1. Abuse binary , Abuse severity 1,0,-1,-2,-3; 2. Directedness explicit, implicit Target group, individual–system, individual–3rd party, Type general, sexist, sexual harassment, homophobic, racist, transphobic, ableist, intellectual
Details of task: Abuse detection in conversational AI
Size of dataset: 4,185
Percentage abusive: c. 20%
Level of annotation: utterance (with conversational context)
Platform: Carbonbot on Facebook Messenger and E.L.I.Z.A. chatbots
Reference: Curry, A. C., Abercrombie, G., & Rieser, V. 2021. ConvAbuse: Data, Analysis, and Benchmarks for Nuanced Detection in Conversational AI. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (pp. 7388-7403).

Measuring Hate Speech

Link to publication: https://arxiv.org/abs/2009.10277
Link to data: https://huggingface.co/datasets/ucberkeley-dlab/measuring-hate-speech
Task description: 10 ordinal labels (sentiment, (dis)respect, insult, humiliation, inferior status, violence, dehumanization, genocide, attack/defense, hate speech), which are debiased and aggregated into a continuous hate speech severity score (hate_speech_score) that includes a region for counterspeech & supportive speeech. Includes 8 target identity groups (race/ethnicity, religion, national origin/citizenship, gender, sexual orientation, age, disability, political ideology) and 42 identity subgroups.
Details of task: Hate speech measurement on social media in English
Size of dataset: 39,565 comments annotated by 7,912 annotators on 10 ordinal labels, for 1,355,560 total labels.
Percentage abusive: 25% - however this dichotomization is not in the spirit of the paper/dataset
Level of annotation: Social media comment
Platform: Twitter, Reddit, YouTube
Reference: Kennedy, C. J., Bacon, G., Sahn, A., & von Vacano, C. (2020). Constructing interval variables via faceted Rasch measurement and multitask deep learning: a hate speech application. arXiv preprint arXiv:2009.10277.

Learning From the Worst (Dynamically generated hate speech dataset)

Link to publication: https://aclanthology.org/2021.acl-long.132/
Link to data: https://github.com/bvidgen/Dynamically-Generated-Hate-Speech-Dataset
Task description: Multi-category hate speech detection
Details of task: Hate detection with fine-grained labels for the type and target of hate. Generated over 4 rounds of human-and-model-in-the-loop adversarial data generation. Collected through Dynabench .
Size of dataset: 41,255
Percentage abusive: 54%
Level of annotation: posts
Platform: Synthetically generated by humans to mimic real-world social media posts
Reference: Vidgen, B., Thurush, T., Waseem, Z., Kiela, D., 2021. Learning from the worst: dynamically generated datasets to improve online hate detection. In Proceedings of the 59th Meeting of the Association for Computational Lingusitics (pp. 1667-1682).

The ‘Call me sexist, but’ sexism dataset

Link to publication: https://ojs.aaai.org/index.php/ICWSM/article/view/18085/17888
Link to data: https://doi.org/10.7802/2251
Task description: Sexism detection based on content and phrasing
Details of task: Sexism detection on English social media data informed by survey items measuring sexist attitudes and adversarial examples
Size of dataset: 6325
Percentage abusive: 28%
Level of annotation: tweets and survey items
Platform: Twitter, Social Psychology scales
Reference: Samory, M., Sen, I., Kohne, J., Flöck, F. and Wagner, C., 2021, May. Call me sexist, but…: Revisiting sexism detection using psychological scales and adversarial samples. In Intl AAAI Conf. Web and Social Media (pp. 573-584).

Hate Towards the Political Opponent: A Twitter Corpus Study of the 2020 US Elections on the Basis of Offensive Speech and Stance Detection__

Link to publication: https://aclanthology.org/2021.wassa-1.18/
Link to data: https://www.ims.uni-stuttgart.de/data/stance_hof_us2020
Task description: Hate / Offensive or neither
Details of task: Data collected to be Twitter by supporters of Trump or Biden
Size of dataset: 3,000
Percentage abusive: 12%
Reference: Lara Grimminger and Roman Klinger (2020): Hate Towards the Political Opponent: A Twitter Corpus Study of the 2020 US Elections on the Basis of Offensive Speech and Stance Detection. 11th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis (collocated with EACL 2021).

AbuseEval v1.0

Link to publication: http://www.lrec-conf.org/proceedings/lrec2020/pdf/2020.lrec-1.760.pdf
Link to data: https://github.com/tommasoc80/AbuseEval
Task description: Explicitness annotation of offensive and abusive content
Details of task: Enriched versions of the OffensEval/OLID dataset with the distinction of explicit/implicit offensive messages and the new dimension for abusive messages. Labels for offensive language: EXPLICIT, IMPLICT, NOT; Labels for abusive language: EXPLICIT, IMPLICT, NOTABU
Size of dataset: 14,100
Percentage abusive: 20.75%
Reference: Caselli, T., Basile, V., Jelena, M., Inga, K., and Michael, G. 2020. “I feel offended, don’t be abusive! implicit/explicit messages in offensive and abusive language”. The 12th Language Resources and Evaluation Conference (pp. 6193-6202). European Language Resources Association.

Do You Really Want to Hurt Me? Predicting Abusive Swearing in Social Media

Link to publication: https://www.aclweb.org/anthology/2020.lrec-1.765.pdf
Link to data: https://github.com/dadangewp/SWAD-Repository
Task description: Binary (abusive swear word, non-abusive swear word)
Details of task: Abusive swearing
Size of dataset: 1,511 swear words (1675 tweets)
Percentage abusive: 0.41% (word level), 0.51% (post level)
Level of annotation: Words
Reference: Pamungkas, E. W., Basile, V., & Patti, V. (2020). Do you really want to hurt me? predicting abusive swearing in social media. In The 12th Language Resources and Evaluation Conference (pp. 6237-6246). European Language Resources Association.

Multimodal Meme Dataset (MultiOFF) for Identifying Offensive Content in Image and Text

Link to publication: https://www.aclweb.org/anthology/2020.trac-1.6.pdf
Link to data: https://github.com/bharathichezhiyan/Multimodal-Meme-Classification-Identifying-Offensive-Content-in-Image-and-Text
Task description: Binary (offensive, non-offensive)
Details of task: Hate per se (related to 2016 U.S. presidential election)
Size of dataset: 743
Percentage abusive: 0.41%
Platform: Kaggle, Reddit, Facebook, Twitter and Instagram
Medium: Text and Images/memes
Reference: Suryawanshi, S., Chakravarthi, B. R., Arcan, M., & Buitelaar, P. (2020, May). Multimodal meme dataset (MultiOFF) for identifying offensive content in image and text. In Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying (pp. 32-41).

Hatemoji: A Test Suite and Adversarially-Generated Dataset for Benchmarking and Detecting Emoji-based Hate

Link to publication: https://arxiv.org/abs/2108.05921
Link to data: https://github.com/HannahKirk/Hatemoji
Task description: Branching structure of tasks: Binary (Hate, Not Hate), Within Hate (Type, Target)
Details of task: Hate speech detection for text statements including emoji, consisting of a checklist-based test suite (HatemojiCheck) and an adversarially-generated dataset (HatemojiBuild)
Size of dataset: HatemojiCheck = 3,930; HatemojiBuild = 5,912.
Percentage abusive: HatemojiCheck = 69%, HatemojiBuild = 50%
Level of annotation: Post
Platform: Synthetically-Generated
Medium: Text with emoji
Reference: Kirk, H. R., Vidgen, B., Röttger, P., Thrush, T., & Hale, S. A. 2021. Hatemoji: A test suite and adversarially-generated dataset for benchmarking and detecting emoji-based hate. arXiv preprint arXiv:2108.05921.

HateCheck: Functional Tests for Hate Speech Detection Models

Link to publication: https://arxiv.org/pdf/2012.15606.pdf
Link to data: https://github.com/paul-rottger/hatecheck-data
Task description: Binary (Hate, Not Hate), 7 Targets Within Hate (Women, Trans people, Black people, Gay people, Disabled people, Muslims, Immigrants)
Details of task: A checklist of functional tests to evaluate hate speech detection models.
Size of dataset: 3,728
Percentage abusive: 68%
Reference: Röttger, P., Vidgen, B., Nguyen, D., Waseem, Z., Margetts, H. and Pierrehumbert, J., 2020. Hatecheck: Functional tests for hate speech detection models. arXiv preprint arXiv:2012.15606.

Semeval-2021 Task 5: Toxic Spans Detection

Link to publication: https://aclanthology.org/2021.semeval-1.6.pdf
Link to data: https://github.com/ipavlopoulos/toxic_spans
Task description: Binary toxic spans (toxic, non-toxic) & reading comprehension
Details of task: Predict the spans of toxic posts that were responsible for the toxic label of the posts.
Size of dataset: 10,629
Percentage abusive: 0.56
Platform: Civil Comments
Reference: Pavlopoulos, J., Sorensen, J., Laugier, L., & Androutsopoulos, I. (2021, August). Semeval-2021 task 5: Toxic spans detection. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021) (pp. 59-69).

ToxiSpanSE: An Explainable Toxicity Detection in Code Review Comments

Link to publication: https://arxiv.org/abs/2307.03386
Link to data and tool: (https://github.com/WSU-SEAL/ToxiSpanSE)
Task description: Binary toxic spans (Toxic, Non-toxic)
Details of task: Toxicity, Context
Size of dataset: 19,651
Percentage of toxic in span level: 13.85
Level of annotation: Code Review Comments
Platform: Open Source Software
Reference: Sarker, Jaydeb, Sultana, Sayma, Wilson, Steven R., and Amiangshu Bosu. “ToxiSpanSE: An Explainable Toxicity Detection in Code Review Comments” The 17th ACM/IEEE International Symposium on Empirical Software Engineering and Measurement (ESEM), 2023.

Human-in-the-Loop for Data Collection: a Multi-Target Counter Narrative Dataset to Fight Online Hate Speech

Link to publication: https://aclanthology.org/2021.acl-long.250.pdf
Link to data: https://github.com/marcoguerini/CONAN
Details of task: race, religion, country of origin, sexual orientation, disability, gender
Size of dataset: 5,003
Percentage abusive: 1
Platform: Semi-synthetic text
Reference: Margherita Fanton, Helena Bonaldi, Serra Sinem Tekiroğlu, Marco Guerini Human-in-the-Loop for Data Collection: a Multi-Target Counter Narrative Dataset to Fight Online Hate Speech In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics: Long Papers.

HateXplain: A Benchmark Dataset for Explainable Hate Speech Detection

Link to publication: https://arxiv.org/abs/2012.10289
Link to data: https://github.com/punyajoy/HateXplain
Task description: Level of hate (hate, offensive or normal), on target groups (race, religion, gender, sexual orientation, miscellaneous), and rationales
Details of task: Hate per se
Size of dataset: 20,148
Percentage abusive: 0.57
Level of annotation: Words, phrases, posts
Platform: Twitter and Gab
Reference: Mathew, B., Saha, P., Yimam, S. M., Biemann, C., Goyal, P., & Mukherjee, A. (2021, May). HateXplain: A Benchmark Dataset for Explainable Hate Speech Detection. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 17, pp. 14867-14875).

ALONE: A Dataset for Toxic Behavior among Adolescents on Twitter

Link to publication: https://arxiv.org/pdf/2008.06465.pdf
Link to data: Data made available upon request, please email Ugur Kursuncu [email protected] and [email protected] [email protected].
Task description: Binary (Toxic, Non-Toxic)
Details of task: Annotates interactions (Tweets and their replies), and assigns keywords describing use of emojis, URL content and images.
Size of dataset: 688
Percentage abusive: 0.17
Medium: Multimodal (text, images, emojis, metadata)
Reference: Wijesiriwardene, T., Inan, H., Kursuncu, U., Gaur, M., Shalin, V., Thirunarayan, K., Sheth, A. and Arpinar, I., 2020, Arxiv.

Towards a Comprehensive Taxonomy and Large-Scale Annotated Corpus for Online Slur Usage

Link to publication: https://www.aclweb.org/anthology/2020.alw-1.17.pdf
Link to data: https://github.com/networkdynamics/slur-corpus
Task description: 4 primary categories (derogatory, appropriate, non-derogatory/non-appropriate, homonyms, noise)
Size of dataset: 39,811
Percentage abusive: 0.52
Platform: Reddit
Reference: Kurrek, J., Saleem, H. M., & Ruths, D. (2020, November). Towards a comprehensive taxonomy and large-scale annotated corpus for online slur usage. In Proceedings of the Fourth Workshop on Online Abuse and Harms (pp. 138-149).
Link to data: https://www.aclweb.org/anthology/2020.trac-1.6.pdf
Percentage abusive: 0.41

Predicting the Type and Target of Offensive Posts in Social Media

Link to publication: https://aclanthology.org/N19-1144.pdf
Link to data: https://scholar.harvard.edu/malmasi/olid
Task description: Branching structure of tasks. A: offensive / not, B: targeted insult / untargeted, C: individual, group, other.
Reference: Zampieri, M., Malmasi, S., Nakov, P., Rosenthal, S., Farra, N., & Kumar, R. (2019, June). Predicting the Type and Target of Offensive Posts in Social Media. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) (pp. 1415-1420).

Nuanced metrics for measuring unintended bias with real data for text classification

Link to publication: https://arxiv.org/pdf/1903.04561.pdf
Link to data: https://www.tensorflow.org/datasets/catalog/civil_comments
Task description: Toxicity (severe, obscene, threat, insult, identity attack, sexual explicit), and several identity attributes (e.g., gender, religion and race)
Size of dataset: 1,804,875
Percentage abusive: 0.8
Level of annotation: Comments/posts
Reference: Borkan, D., Dixon, L., Sorensen, J., Thain, N., & Vasserman, L. (2019, May). Nuanced metrics for measuring unintended bias with real data for text classification. In Companion proceedings of the 2019 world wide web conference (pp. 491-500).

Introducing CAD: the Contextual Abuse Dataset

Link to publication: https://aclanthology.org/2021.naacl-main.182.pdf
Link to data: https://zenodo.org/record/4881008#.Ye6OwhP7R6o
Task description: Contextually abusive language, person-directed + group-directed
Details of task: Primary categories (secondary categories): Abusive + Identity-directed (derogation/animosity/threatening/glorification/dehumanization), Abusive + Person-directed (derogation/animosity/threatening/glorification/dehumanization), Abusive + Affiliation directed (abuse to them/abuse about them), Counter Speech (against identity-directed abuse/against affiliation-directed abuse/against person-directed abuse), Non-hateful Slurs and Neutral.
Size of dataset: 25,000
Percentage abusive: Affiliation-directed, 6%; Identity-directed, 13%; Person-directed, 5%
Level of annotation: Conversation thread
Reference: Vidgen, B., Nguyen, D., Margetts, H., Rossini, P., and Troble, R., Introducing CAD: the Contextual Abuse Dataset, 2021, In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp.2289–2303

Automated Hate Speech Detection and the Problem of Offensive Language

Link to publication: [https://ojs.aaai.org/index.php/ICWSM/article/view/14955)
Link to data: https://github.com/t-davidson/hate-speech-and-offensive-language
Task description: Hierarchy (Hate, Offensive, Neither)
Size of dataset: 24,802
Percentage abusive: 0.06
Reference: Davidson, T., Warmsley, D., Macy, M., & Weber, I. 2017. Automated Hate Speech Detection and the Problem of Offensive Language. Proceedings of the International AAAI Conference on Web and Social Media, 11(1), 512-515.

Hate Speech Dataset from a White Supremacy Forum

Link to publication: https://www.aclweb.org/anthology/W18-5102.pdf
Link to data: https://github.com/Vicomtech/hate-speech-dataset
Task description: Ternary (Hate, Relation, Not)
Size of dataset: 9,916
Percentage abusive: 0.11
Level of annotation: Sentence - with context of the converstaional thread taken into account
Platform: Stormfront
Reference: de Gibert, O., Perez, N., García-Pablos, A., and Cuadros, M., 2018. Hate Speech Dataset from a White Supremacy Forum. In: Proceedings of the 2nd Workshop on Abusive Language Online (ALW2). Brussels, Belgium: Association for Computational Linguistics, pp.11-20.

Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter

Link to publication: https://www.aclweb.org/anthology/N16-2013
Link to data: https://github.com/ZeerakW/hatespeech
Task description: 3-topic (Sexist, Racist, Not)
Details of task: Racism, Sexism
Size of dataset: 16,914
Percentage abusive: 0.32
Reference: Waseem, Z. and Horvy, D., 2016. Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter. In: Proceedings of the NAACL Student Research Workshop. San Diego, California: Association for Computational Linguistics, pp.88-93.

Detecting Online Hate Speech Using Context Aware Models

Link to publication: https://arxiv.org/pdf/1710.07395.pdf
Link to data: https://github.com/sjtuprog/fox-news-comments
Task description: Binary (Hate / not)
Size of dataset: 1528
Percentage abusive: 0.28
Platform: Fox News
Reference: Gao, L. and Huang, R., 2018. Detecting Online Hate Speech Using Context Aware Models. ArXiv,.

The Gab Hate Corpus: A collection of 27k posts annotated for hate speech

Link to publication: https://psyarxiv.com/hqjxn/
Link to data: https://osf.io/edua3/
Task description: Binary (Hate vs. Offensive/Vulgarity), Binary (Assault on human Dignity/Call for Violence – sub task on message delivery, binary: explicit/implicit), Multinomial classification: Identity based hate (race/ethnicity, nationality/regionalism/xenophobia, gender, religion/belief system, sexual orientation, ideology, political identification/party, mental/physical health)
Size of dataset: 27,665
Percentage abusive: 0.09 Hate, 0.06 Offensive/Vulgar
Platform: Gab
Reference: Kennedy, B., Araria, M., Mostafazadeh Davani, A., Yeh, L., Omrani, A., Kim, Y., Koombs, K., Havaldar, S., Portillo-Wightman, G., Gonzalez, E., Hoover, J., Azatain, A., Hussain, A., Lara, A., Olmos, G., Omary, A., Park, C., Wang, C., Wang, X., Zhang, Y. and Dehghani, M., 2018, The Gab Hate Corpus: A collection of 27k posts annotated for hate speech. PsyArXiv.

Are You a Racist or Am I Seeing Things? Annotator Influence on Hate Speech Detection on Twitter

Link to publication: https://pdfs.semanticscholar.org/3eeb/b7907a9b94f8d65f969f63b76ff5f643f6d3.pdf
Task description: Multi-topic (Sexist, Racist, Neither, Both)
Size of dataset: 4,033
Percentage abusive: 0.16
Reference: Waseem, Z., 2016. Are You a Racist or Am I Seeing Things? Annotator Influence on Hate Speech Detection on Twitter. In: Proceedings of 2016 EMNLP Workshop on Natural Language Processing and Computational Social Science. Copenhagen, Denmark: Association for Computational Linguistics, pp.138-142.

When Does a Compliment Become Sexist? Analysis and Classification of Ambivalent Sexism Using Twitter Data

Link to publication: https://pdfs.semanticscholar.org/225f/f8a6a562bbb64b22cebfcd3288c6b930d1ef.pdf
Link to data: https://github.com/AkshitaJha/NLP_CSS_2017
Task description: Hierarchy of Sexism (Benevolent sexism, Hostile sexism, None)
Details of task: Sexism
Size of dataset: 712
Reference: Jha, A. and Mamidi, R., 2017. When does a Compliment become Sexist? Analysis and Classification of Ambivalent Sexism using Twitter Data. In: Proceedings of the Second Workshop on Natural Language Processing and Computational Social Science. Vancouver, Canada: Association for Computational Linguistics, pp.7-16.

Overview of the Task on Automatic Misogyny Identification at IberEval 2018 (English)

Link to publication: http://ceur-ws.org/Vol-2150/overview-AMI.pdf
Link to data: https://amiibereval2018.wordpress.com/im nt-dates/data/
Task description: Binary (misogyny / not), 5 categories (stereotype, dominance, derailing, sexual harassment, discredit), target of misogyny (active or passive)
Size of dataset: 3,977
Percentage abusive: 0.47
Reference: Fersini, E., Rosso, P. and Anzovino, M., 2018. Overview of the Task on Automatic Misogyny Identification at IberEval 2018. In: Proceedings of the Third Workshop on Evaluation of Human Language Technologies for Iberian Languages (IberEval 2018).

CONAN - COunter NArratives through Nichesourcing: a Multilingual Dataset of Responses to Fight Online Hate Speech (English)

Link to publication: https://www.aclweb.org/anthology/P19-1271.pdf
Task description: Binary (Islamophobic / not), multi-topic (Culture, Economics, Crimes, Rapism, Terrorism, Women Oppression, History, Other/generic)
Details of task: Islamophobia
Size of dataset: 1,288
Platform: Synthetic / Facebook
Reference: Chung, Y., Kuzmenko, E., Tekiroglu, S. and Guerini, M., 2019. CONAN - COunter NArratives through Nichesourcing: a Multilingual Dataset of Responses to Fight Online Hate Speech. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Florence, Italy: Association for Computational Linguistics, pp.2819-2829.

Characterizing and Detecting Hateful Users on Twitter

Link to publication: https://arxiv.org/pdf/1803.08977.pdf
Link to data: https://github.com/manoelhortaribeiro/HatefulUsersTwitter
Task description: Binary (hateful/not)
Size of dataset: 4,972
Level of annotation: Users
Reference: Ribeiro, M., Calais, P., Santos, Y., Almeida, V. and Meira, W., 2018. Characterizing and Detecting Hateful Users on Twitter. ArXiv,.

A Benchmark Dataset for Learning to Intervene in Online Hate Speech (Gab)

Link to publication: https://arxiv.org/abs/1909.04251
Link to data: https://github.com/jing-qian/A-Benchmark-Dataset-for-Learning-to-Intervene-in-Online-Hate-Speech
Size of dataset: 33,776
Percentage abusive: 0.43
Level of annotation: Posts (in the context of a conversation)
Reference: Qian, J., Bethke, A., Belding, E. and Yang Wang, W., 2019. A Benchmark Dataset for Learning to Intervene in Online Hate Speech. ArXiv,.

A Benchmark Dataset for Learning to Intervene in Online Hate Speech (Reddit)

Size of dataset: 22,324
Percentage abusive: 0.24
Level of annotation: Posts (with context of the converstaional thread taken into account)

Multilingual and Multi-Aspect Hate Speech Analysis (English)

Task description: Detailed taxonomy with cross-cutting attributes: Hostility, Directness, Target attribute and Target group.
Size of dataset: 5,647
Percentage abusive: 0.76

Exploring Hate Speech Detection in Multimodal Publications

Link to publication: https://arxiv.org/pdf/1910.03814.pdf
Link to data: https://drive.google.com/file/d/1S9mMhZFkntNnYdO-1dZXwF_8XIiFcmlF/view
Task description: Multimodal Hate Speech Detection, including six primary categories (No attacks to any community, Racist, Sexist, Homophobic, Religion based attack, Attack to other community)
Details of task: Racism, Sexism, Homophobia, Religion-based attack
Size of dataset: 149,823
Percentage abusive: 0.25
Medium: Text and Images/Memes
Reference: Gomez, R., Gibert, J., Gomez, L. and Karatzas, D., 2020. In Proceedings of the IEEE/CVF winter conference on applications of computer vision (pp. 1470-1478).
Link to publication: https://arxiv.org/pdf/1902.09666.pdf
Link to data: http://competitions.codalab.org/ competitions/20011
Reference: Zampieri, M., Malmasi, S., Nakov, P., Rosenthal, S., Farra, N. and Kumar, R., 2019. SemEval-2019 Task 6: Identifying and Categorizing Offensive Language in Social Media (OffensEval). ArXiv,.

hatEval, SemEval-2019 Task 5: Multilingual Detection of Hate Speech Against Immigrants and Women in Twitter (English)

Link to publication: https://www.aclweb.org/anthology/S19-2007
Link to data: http://competitions.codalab.org/competitions/19935
Task description: Branching structure of tasks: Binary (Hate, Not), Within Hate (Group, Individual), Within Hate (Agressive, Not)
Size of dataset: 13,000
Percentage abusive: 0.4
Reference: Basile, V., Bosco, C., Fersini, E., Nozza, D., Patti, V., Pardo, F., Rosso, P. and Sanguinetti, M., 2019. SemEval-2019 Task 5: Multilingual Detection of Hate Speech Against Immigrants and Women in Twitter. In: Proceedings of the 13th International Workshop on Semantic Evaluation. Minneapolis, Minnesota: Association for Computational Linguistics, pp.54-63.

Peer to Peer Hate: Hate Speech Instigators and Their Targets

Link to publication: https://aaai.org/ocs/index.php/ICWSM/ICWSM18/paper/view/17905/16996
Link to data: https://github.com/mayelsherif/hate_speech_icwsm18
Task description: Binary (Hate/Not), only for tweets which have both a Hate Instigator and Hate Target
Size of dataset: 27,330
Percentage abusive: 0.98
Reference: ElSherief, M., Nilizadeh, S., Nguyen, D., Vigna, G. and Belding, E., 2018. Peer to Peer Hate: Hate Speech Instigators and Their Targets. In: Proceedings of the Twelfth International AAAI Conference on Web and Social Media (ICWSM 2018). Santa Barbara, California: University of California, pp.52-61.

Overview of the HASOC track at FIRE 2019: Hate Speech and Offensive Content Identification in Indo-European Languages

Link to publication: https://dl.acm.org/doi/pdf/10.1145/3368567.3368584?download=true
Link to data: https://hasocfire.github.io/hasoc/2019/dataset.html
Task description: Branching structure of tasks. A: Hate / Offensive or Neither, B: Hatespeech, Offensive, or Profane, C: Targeted or Untargeted
Size of dataset: 7,005
Percentage abusive: 0.36
Platform: Twitter and Facebook
Reference: Mandl, T., Modha, S., Majumder, P., Patel, D., Dave, M., Mandlia, C. and Patel, A., 2019. Overview of the HASOC track at FIRE 2019. In: Proceedings of the 11th Forum for Information Retrieval Evaluation,.

Detecting East Asian Prejudice on Social media

Link to publication: https://www.aclweb.org/anthology/2020.alw-1.19.pdf
Link to data: https://zenodo.org/record/3816667
Task description: Task 1: Thematic annotation (East Asia/Covid-19) Task 2: Primary category annotation: 1) Hostility against an East Asian (EA) entity 2) Criticism of an East Asian entity 3) Counter speech 5) Discussion of East Asian prejudice 5) Non-related. Task 3: Secondary category annotation (if (1) or (2) - identifying what East Asian entity was targeted + if (1) interpersonal abuse/threatening language/dehumanization).
Details of task: Detecting East Asian prejudice
Size of dataset: 20,000
Percentage abusive: 27% (Hostility, 19.5%; Criticism, 7.2%)
Reference: Vidgen, B., Botelho, A., Broniatowski, D., Guest, E., Hall, M., Margetts, H., Tromble, R., Waseem, Z. and Hale, S., Detecting East Asian Prejudice on Social media, 2020, In: Proceedings of the Fourth Workshop on Online Abuse and Harms, pp.162–172

Large Scale Crowdsourcing and Characterization of Twitter Abusive Behavior

Link to publication: https://arxiv.org/pdf/1802.00393.pdf
Link to data: https://dataverse.mpi-sws.org/dataset.xhtml?persistentId=doi:10.5072/FK2/ZDTEMN
Task description: Multi-thematic (Abusive, Hateful, Normal, Spam)
Size of dataset: 80,000
Percentage abusive: 0.18
Annotation process: Very detailed information is given: multiple rounds, using a smaller 300 tweet dataset for testing the schema. For the final 80k, 5 judgements per tweet. CrowdFlower
Annotation agreement: 55.9% = 4/5, 36.6% = 3/5, 7.5% = 2/5
Reference: Founta, A., Djouvas, C., Chatzakou, D., Leontiadis, I., Blackburn, J., Stringhini, G., Vakali, A., Sirivianos, M. and Kourtellis, N., 2018. Large Scale Crowdsourcing and Characterization of Twitter Abusive Behavior. ArXiv,.

A Large Labeled Corpus for Online Harassment Research

Link to publication: http://www.cs.umd.edu/~golbeck/papers/trolling.pdf
Link to data: [email protected]
Task description: Binary (Harassment, Not)
Details of task: Person-directed
Size of dataset: 35,000
Reference: Golbeck, J., Ashktorab, Z., Banjo, R., Berlinger, A., Bhagwan, S., Buntain, C., Cheakalos, P., Geller, A., Gergory, Q., Gnanasekaran, R., Gnanasekaran, R., Hoffman, K., Hottle, J., Jienjitlert, V., Khare, S., Lau, R., Martindale, M., Naik, S., Nixon, H., Ramachandran, P., Rogers, K., Rogers, L., Sarin, M., Shahane, G., Thanki, J., Vengataraman, P., Wan, Z. and Wu, D., 2017. A Large Labeled Corpus for Online Harassment Research. In: Proceedings of the 2017 ACM on Web Science Conference. New York: Association for Computing Machinery, pp.229-233.

Ex Machina: Personal Attacks Seen at Scale, Personal attacks

Link to publication: https://arxiv.org/pdf/1610.08914
Link to data: https://github.com/ewulczyn/wiki-detox
Task description: Binary (Personal attack, Not)
Size of dataset: 115,737
Platform: Wikipedia
Reference: Wulczyn, E., Thain, N. and Dixon, L., 2017. Ex Machina: Personal Attacks Seen at Scale. ArXiv,.

Ex Machina: Personal Attacks Seen at Scale, Toxicity

Task description: Toxicity/healthiness judgement (-2 == very toxic, 0 == neutral, 2 == very healthy)
Size of dataset: 100,000
Percentage abusive: NA

Detecting cyberbullying in online communities (World of Warcraft)

Link to publication: http://aisel.aisnet.org/ecis2016_rp/61/
Link to data: http://ub-web.de/research/
Size of dataset: 16,975
Percentage abusive: 0.01
Platform: World of Warcraft
Reference: Bretschneider, U. and Peters, R., 2016. Detecting Cyberbullying in Online Communities. Research Papers, 61.

Detecting cyberbullying in online communities (League of Legends)

Size of dataset: 17,354
Platform: League of Legends

A Quality Type-aware Annotated Corpus and Lexicon for Harassment Research

Link to publication: https://arxiv.org/pdf/1802.09416.pdf
Link to data: https://github.com/Mrezvan94/Harassment-Corpus
Task description: Multi-topic harassment detection
Details of task: Racism, Sexism, Appearance-related, Intellectual, Political
Size of dataset: 24,189
Percentage abusive: 0.13
Reference: Rezvan, M., Shekarpour, S., Balasuriya, L., Thirunarayan, K., Shalin, V. and Sheth, A., 2018. A Quality Type-aware Annotated Corpus and Lexicon for Harassment Research. ArXiv,.

Ex Machina: Personal Attacks Seen at Scale, Aggression and Friendliness

Task description: Aggression/friendliness judgement on a 5 point scale. (-2 == very aggressive, 0 == neutral, 3 == very friendly).
Details of task: Person-Directed + Group-Directed
Size of dataset: 160,000

Are Chess Discussions Racist? An Adversarial Hate Speech Data Set

Link to publication: https://arxiv.org/pdf/2011.10280.pdf
Link to data: https://www.cs.cmu.edu/~akhudabu/Chess.html
Task description: Not Labeled
Details of task: Racism, Misclassification
Size of dataset: 1,000
Percentage abusive: 0.0
Platform: Youtube
Reference: Rupak Sarkar and Ashiqur R. KhudaBukhsh, Nov. 2020. Are Chess Discussions Racist? An Adversarial Hate Speech Data Set. In: The Thirty-Fifth {AAAI} Conference on Artificial Intelligence, {AAAI} 2021

ETHOS: an Online Hate Speech Detection Dataset (Binary)

Link to publication: https://arxiv.org/pdf/2006.08328.pdf
Link to data: https://github.com/intelligence-csd-auth-gr/Ethos-Hate-Speech-Dataset
Details of task: Gender, Race, National Origin, Disability, Religion, Sexual Orientation
Size of dataset: 998
Platform: Youtube, Reddit
Reference: Mollas, I., Chrysopoulou, Z., Karlos, S., and Tsoumakas, G., 2021. ETHOS: an Online Hate Speech Detection Dataset. Complex & Intelligent Systems, Jan. 2022

ETHOS: an Online Hate Speech Detection Dataset (Multi label)

Task description: 8 Categories (Violence, Directed/Undirected, Gender, Race, National Origin, Disability, Sexual Orientation, Religion)
Size of dataset: 433

Twitter Sentiment Analysis

Link to publication: NA
Link to data: https://www.kaggle.com/arkhoshghalb/twitter-sentiment-analysis-hatred-speech
Size of dataset: 31,961
Percentage abusive: 0.07
Reference: Ali Toosi, Jan 2019. Twitter Sentiment Analysis

Toxicity Detection in Software Engineering: Automated Identification of Toxic Code Reviews Using ToxiCR

Link to publication: https://dl.acm.org/doi/abs/10.1145/3583562
Link to data: https://github.com/WSU-SEAL/ToxiCR
Task description: Binary (Toxic, Non-toxic)
Percentage of toxic: 19
Reference: Sarker, Jaydeb, Asif Kamal Turzo, Ming Dong, and Amiangshu Bosu. “Automated Identification of Toxic Code Reviews Using ToxiCR.” ACM Transactions on Software Engineering and Methodology (2023).

Toxicity Detection: Does Context Really Matter? CAT-LARGE (No Context)

Link to publication: https://arxiv.org/pdf/2006.00998.pdf
Link to data: https://github.com/ipavlopoulos/context_toxicity
Size of dataset: 10,000
Percentage abusive: 0.006
Platform: Wikipedia Talk Pages
Reference: Pavlopoulos, J., Sorensen, J., Dixon, L., Thain, N., & Androutsopoulos, I. (2020). Toxicity Detection: Does Context Really Matter? ArXiv:2006.00998 [Cs].

Toxicity Detection: Does Context Really Matter? CAT-LARGE (With Context)

Anatomy of online hate: developing a taxonomy and machine learning models for identifying and classifying hate in online news media.

Link to publication: https://www.aaai.org/ocs/index.php/ICWSM/ICWSM18/paper/viewFile/17885/17024
Link to data: https://www.dropbox.com/s/21wtzy9arc5skr8/ICWSM18%20-%20SALMINEN%20ET%20AL.xlsx?dl=0
Task description: Binary (Hate, Not), Multinomial classification (21 categories divided into ‘hateful language’, ‘hate targets’ and ‘hate sub-targets’)
Size of dataset: 5,143
Percentage abusive: 82%
Level of annotation: Comment
Platform: YouTube and Facebook
Reference: Salminen, J., Almerekhi, H., Milenković, M., Jung, S., An, J., Kwak, H. and Jansen, B., 2018, Anatomy of Online Hate: Developing a Taxonomy and Machine Learning Models for Identifying and Classifying Hate in Online News Media, In: Proceedings of the Twelfth International AAAI Conference on Web and Social Media (ICWSM 2018), pp.330-339
Link to data: http://hdl.handle.net/11356/1401
Size of dataset: 31.5M
Percentage abusive: 12.5%
Language: Estonian (some in Russian also)
Platform: Eesti Ekspress (www.ekspress.ee) website
Link to publication: https://arxiv.org/pdf/2012.10289.pdf
Task description: Binary (Hate, Not) and Three-class (Hate speech, Offensive language, None)
Details of task: Hatespeech detection on social media in English, including 10 categories: African, Islam, Jewish, LGBTQ, Women, Refugee, Arab, Caucasian, Hispanic, Asian
Size of dataset: 20148
Percentage abusive: 57%
Reference: Mathew, B., Saha, P., Yimam, S. M., Biemann, C., Goyal, P., & Mukherjee, A. (2020). Hatexplain: A benchmark dataset for explainable hate speech detection. arXiv preprint arXiv:2012.10289.

CONAN - COunter NArratives through Nichesourcing: a Multilingual Dataset of Responses to Fight Online Hate Speech (French)

Task description: Binary (Islamophobic / not), Multi-topic (Culture, Economics, Crimes, Rapism, Terrorism, Women Oppression, History, Other/generic)
Size of dataset: 1,719
Language: French

Multilingual and Multi-Aspect Hate Speech Analysis (French)

Size of dataset: 4,014
Percentage abusive: 0.72

CyberAgressionAdo-v1

Link to publication: ( url ) - link to the documentation and/or a data statement about the data
Link to data: ( url ) - direct download is preferred, e.g. a link straight to a .zip file
Task description: The collected conversations have been annotated using a fine-grained tagset including information related to the participant roles, the presence of hate speech, the type of verbal abuse present in the message, and whether utterances use different humour figurative devices (e.g., sarcasm or irony).
Details of task: This dataset allows to perform several subtasks related to the task of online hate detection in a conversational setting (hate speech detection, bullying participant role detection, verbal abuse detection, etc.)
Size of dataset: 19 conversations
Level of annotation: exchanged messages
Platform: collected from role playing games mimicking cyberagression situations occuring on private instant messaging platforms.
Medium: text (csv)
Reference: Anaïs Ollagnier, Elena Cabrio, Serena Villata, Catherine Blaya. CyberAgressionAdo-v1: a Dataset of Annotated Online Aggressions in French Collected through a Role-playing Game. Language Resources and Evaluation Conference, Jun 2022, Marseille, France. ⟨hal-03765860⟩

DeTox: A Comprehensive Dataset for German Offensive Language and Conversation Analysis

Link to publication: https://aclanthology.org/2022.woah-1.14.pdf
Link to data: https://github.com/hdaSprachtechnologie/detox
Task description: Comprehensive annotation schema (including sentiment, hate speech, type of discrimination, criminal relevance, expression, toxicity, extremism, target, threat)
Details of task: About half of the comments are from coherent comment threads which allows simple conversation analyses. Every comment was annotated by three annotators.
Size of dataset: 10,278
Percentage abusive: 10.85%
Language: German
Level of annotation: Comments
Reference: Demus, C., Pitz, P., Schütz, M., Probol, N., Siegel, M., and Labudde, L. 2022. DeTox: A Comprehensive Dataset for German Offensive Language and Conversation Analysis. In Proceedings of the Sixth Workshop on Online Abuse and Harms (WOAH), pages 143–153, Seattle, Washington (Hybrid). Association for Computational Linguistics.

RP-Mod & RP-Crowd: Moderator- and Crowd-Annotated German News Comment Datasets

Link to publication: https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/file/c9e1074f5b3f9fc8ea15d152add07294-Paper-round2.pdf
Link to data: https://zenodo.org/record/5291339#.Ybr_9VkxkUE
Task description: Binary (Offensive or Not), Multi-class/-label (sexism, racism, threats, insults, profane language, meta, advertisement).
Details of task: The comments originate from a large German newspaper and are annotated by professional moderators (community managers). Additionally, each comment was further annotated by five different crowd-workers.
Size of dataset: 85,000
Percentage abusive: 8.4%
Platform: German Newspaper (Rheinische Post)
Reference: Assenmacher, D., Niemann, M., Müller, K., Seiler, M., Riehle, D. M., & Trautmann, H. (2021). RP-Mod & RP-Crowd: Moderator- and crowd-annotated german news comment datasets. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmark.

Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis

Link to publication: https://arxiv.org/pdf/1701.08118.pdf
Link to data: https://github.com/UCSM-DUE/IWG_hatespeech_public
Task description: Binary (Anti-refugee hate, None)
Details of task: Refugees
Size of dataset: 469
Reference: Ross, B., Rist, M., Carbonell, G., Cabrera, B., Kurowsky, N. and Wojatzki, M., 2017. Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis. ArXiv,.

Detecting Offensive Statements Towards Foreigners in Social Media

Link to publication: https://pdfs.semanticscholar.org/23dc/df7c7e82807445afd9f19474fc0a3d8169fe.pdf
Task description: Hierarchical (Anti-foreigner prejudice, split into (1) slightly offensive/offensive and (2) explicitly/substantially offensive). 6 targets (Foreigner, Government, Press, Community, Other, Unknown)
Details of task: Anti-foreigner prejudice
Size of dataset: 5,836
Platform: Facebook
Reference: Bretschneider, U. and Peters, R., 2017. Detecting Offensive Statements towards Foreigners in Social Media. In: Proceedings of the 50th Hawaii International Conference on System Sciences.

GermEval 2018

Link to publication: https://www.researchgate.net/publication/327914386_Overview_of_the_GermEval_2018_Shared_Task_on_the_Identification_of_Offensive_Language
Link to data: https://github.com/uds-lsv/GermEval-2018-Data
Task description: Branching structure: Binary (Offense, Other), 3 levels within Offense (Abuse, Insult, Profanity)
Details of task: Group-directed + Incivility
Size of dataset: 8,541
Percentage abusive: 0.34
Reference: Wiegand, M., Siegel, M. and Ruppenhofer, J., 2018. Overview of the GermEval 2018 Shared Task on the Identification of Offensive Language. In: Proceedings of GermEval 2018, 14th Conference on Natural Language Processing (KONVENS 2018). Vienna, Austria: Research Gate.
Task description: A: Hate / Offensive or neither, B: Hatespeech, Offensive, or Profane
Size of dataset: 4,669

GAHD: A German Adversarial Hate Speech Dataset

Link to publication: https://arxiv.org/abs/2403.19559
Link to data: https://github.com/jagol/gahd
Task description: Binary hate speech detection (“hate speech”, “not-hate speech”)
Details of task: Consists of adversarial and contrastive examples
Size of dataset: 10,996 texts
Percentage abusive: 42.4%
Level of annotation: Post/Sentence
Platform: Synthetic data and news sentences
Reference: Goldzycher, J., Röttger, P., and Schneider, G., 2024. Improving Adversarial Data Collection by Supporting Annotators: Lessons from GAHD, a German Hate Speech Dataset. To appear in the Proceedings of the North American Chapter of the Association for Computational Linguistics (NAACL 2024), Mexico, Mexico City, June 17–19.

Deep Learning for User Comment Moderation, Flagged Comments

Link to publication: https://www.aclweb.org/anthology/W17-3004
Link to data: http://www.straintek.com/data/
Task description: Binary (Flagged, Not)
Size of dataset: 1,450,000
Language: Greek
Platform: Gazetta
Reference: Pavlopoulos, J., Malakasiotis, P. and Androutsopoulos, I., 2017. Deep Learning for User Comment Moderation. In: Proceedings of the First Workshop on Abusive Language Online. Vancouver, Canada: Association for Computational Linguistics, pp.25-35.

Deep Learning for User Comment Moderation, Moderated Comments

Size of dataset: 1,500
Percentage abusive: 0.22

Offensive Language Identification in Greek

Link to publication: https://arxiv.org/pdf/2003.07459v1.pdf
Link to data: https://sites.google.com/site/offensevalsharedtask/home
Size of dataset: 4779
Percentage abusive: 0.29
Reference: Pitenis, Z., Zampieri, M. and Ranasinghe, T., 2020. Offensive Language Identification in Greek. ArXiv.

Hindi / Hindi-English

Hostility detection dataset in hindi.

Link to publication: https://arxiv.org/pdf/2011.03588.pdf
Link to data: https://competitions.codalab.org/competitions/26654
Task description: Branching structure of tasks: Binary (Hostile, Not Hostile), Multi-tags within Hostile (Fake News, Hate, Offense, Defame)
Details of task: Hostility detection
Size of dataset: 8,192
Percentage abusive: 47%
Language: Hindi
Platform: Twitter, Facebook, WhatsApp
Reference: Bhardwaj, M., Akhtar, M.S., Ekbal, A., Das, A. and Chakraborty, T., 2020. Hostility detection dataset in hindi. arXiv preprint arXiv:2011.03588.

Aggression-annotated Corpus of Hindi-English Code-mixed Data

Link to publication: https://arxiv.org/pdf/1803.09402
Link to data: https://github.com/kraiyani/Facebook-Post-Aggression-Identification
Task description: 3 part hierachy for hate (None, Covert Aggression, Overt Aggression), 4 part target categorisation (Physical threat, Sexual threat, Identity threat, Non-threatening aggression), 3-part discursive role categorisation (Attack, Defend, Abet)
Details of task: Numerous sub-categorizations
Size of dataset: 18,000
Language: Hindi-English
Reference: Kumar, R., Reganti, A., Bhatia, A. and Maheshwari, T., 2018. Aggression-annotated Corpus of Hindi-English Code-mixed Data. ArXiv,.
Size of dataset: 21,000
Percentage abusive: 0.27

Did You Offend Me? Classification of Offensive Tweets in Hinglish Language

Link to publication: https://www.aclweb.org/anthology/W18-5118
Link to data: https://github.com/pmathur5k10/Hinglish-Offensive-Text-Classification
Task description: Hierarchy (Not Offensive, Abusive, Hate)
Size of dataset: 3,189
Percentage abusive: 0.65
Reference: Mathur, P., Sawhney, R., Ayyar, M. and Shah, R., 2018. Did you offend me? Classification of Offensive Tweets in Hinglish Language. In: Proceedings of the 2nd Workshop on Abusive Language Online (ALW2). Brussels, Belgium: Association for Computational Linguistics, pp.138-148.

A Dataset of Hindi-English Code-Mixed Social Media Text for Hate Speech Detection

Link to publication: https://www.aclweb.org/anthology/W18-1105
Link to data: https://github.com/deepanshu1995/HateSpeech-Hindi-English-Code-Mixed-Social-Media-Text
Size of dataset: 4,575
Reference: Bohra, A., Vijay, D., Singh, V., Sarfaraz Akhtar, S. and Shrivastava, M., 2018. A Dataset of Hindi-English Code-Mixed Social Media Text for Hate Speech Detection. In: Proceedings of the Second Workshop on Computational Modeling of People’s Opinions, Personality, and Emotions in Social Media. New Orleans, Louisiana: Association for Computational Linguistics, pp.36-41.
Link to data: https://hasocfire.github.io/hasoc/2019/dataset.htm
Task description: A: Hate, Offensive or Neither, B: Hatespeech, Offensive, or Profane, C: Targeted or Untargeted
Size of dataset: 5,983
Percentage abusive: 0.51

Hate Speech Detection in the Indonesian Language: A Dataset and Preliminary Study

Link to publication: https://ieeexplore.ieee.org/document/8355039
Link to data: https://github.com/ialfina/id-hatespeech-detection
Size of dataset: 713
Language: Indonesian
Reference: Alfina, I., Mulia, R., Fanany, M. and Ekanata, Y., 2017. Hate Speech Detection in the Indonesian Language: A Dataset and Preliminary Study. In: International Conference on Advanced Computer Science and Information Systems. pp.233-238.

Multi-Label Hate Speech and Abusive Language Detection in Indonesian Twitter

Link to publication: https://www.aclweb.org/anthology/W19-3506
Link to data: https://github.com/okkyibrohim/id-multi-label-hate-speech-and-abusive-language-detection
Task description: (No hate speech, No hate speech but abusive, Hate speech but no abuse, Hate speech and abuse), within hate, category (Religion/creed, Race/ethnicity, Physical/disability, Gender/sexual orientation, Other invective/slander), within hate, strength (Weak, Moderate and Strong)
Details of task: Religion, Race, Disability, Gender
Size of dataset: 13,169
Percentage abusive: 0.42
Reference: Okky Ibrohim, M. and Budi, I., 2019. Multi-label Hate Speech and Abusive Language Detection in Indonesian Twitter. In: Proceedings of the Third Workshop on Abusive Language Online. Florence, Italy: Association for Computational Linguistics, pp.46-57.

A Dataset and Preliminaries Study for Abusive Language Detection in Indonesian Social Media

Link to publication: https://www.sciencedirect.com/science/article/pii/S1877050918314583
Link to data: https://github.com/okkyibrohim/id-abusive-language-detection
Task description: Hierarchical (Not abusive, Abusive but not offensive, Offensive)
Size of dataset: 2,016
Percentage abusive: 0.54
Reference: Ibrohim, M. and Budi, I., 2018. A Dataset and Preliminaries Study for Abusive Language Detection in Indonesian Social Media. Procedia Computer Science, 135, pp.222-229.

BEEP! Korean Corpus of Online News Comments for Toxic Speech Detection__

Link to publication: https://www.aclweb.org/anthology/2020.socialnlp-1.4
Link to data: https://github.com/kocohub/korean-hate-speech
Task description: Binary (Gender bias, No gender bias), Ternary (Gender bias, Other biases, None), Ternary (Hate, Offensive, None)
Details of task: Person/Group-directed, Gender/Sexual orientation, Sexism, Harmfulness/Toxicity
Size of dataset: 9,381
Percentage abusive: 33.87 (Bias), 57.77 (Toxicity)
Language: Korean
Platform: NAVER entertainment news
Reference: Moon, J., Cho, W. I., and Lee, J., 2020. BEEP! Korean Corpus of Online News Comments for Toxic Speech Detection. In: Proceedings of the Eighth International Workshop on Natural Language Processing for Social Media Month: July. Online: Association for Computational Linguistics, pp.25-31.

Latvian newspaper user comment dataset

Link to publication: https://aclanthology.org/2021.hackashop-1.14.pdf
Link to data: https://www.clarin.si/repository/xmlui/handle/11356/1407
Size of dataset: 12M
Percentage abusive: ~10%
Language: Latvian
Reference: Senja Pollak, Marko Robnik-Šikonja, Matthew Purver, Michele Boggia, Ravi Shekhar, Marko Pranjić, Salla Salmela, Ivar Krustok, Tarmo Paju, Carl-Gustav Linden, Leo Leppänen, Elaine Zosa, Matej Ulčar, Linda Freiental, Silver Traat, Luis Adrián Cabrera-Diego, Matej Martinc, Nada Lavrač, Blaž Škrlj, Martin Žnidaršič, Andraž Pelicon, Boshko Koloski, Vid Podečan, Janez Kranjc, Shane Sheehan, Emanuela Boros, Jose Moreno, Antoine Doucet, Hannu Toivonen (2021). EMBEDDIA Tools, Datasets and Challenges: Resources and Hackathon Contributions. Proceedings of the Hackashop on News Media Content Analysis and Automated Report Generation (EACL).

An Italian Twitter Corpus of Hate Speech against Immigrants

Link to publication: https://www.aclweb.org/anthology/L18-1443
Link to data: https://github.com/msang/hate-speech-corpus
Task description: Binary (Immigrants/Roma/Muslims, Not), additional categories. Within Hate, Intensity measurement (Aggressiveness: No, Weak, Strong, Offensiveness: No, Weak, Strong, Irony: No, Yes, Stereotype: No, Yes, Incitement degree: 0-4)
Details of task: Immigrants, Roma and Muslims + numerous sub-categorizations
Size of dataset: 1,827
Language: Italian
Reference: Sanguinetti, M., Poletto, F., Bosco, C., Patti, V. and Stranisci, M., 2018. An Italian Twitter Corpus of Hate Speech against Immigrants. In: Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018). Miyazaki, Japan: European Language Resources Association (ELRA).

Overview of the EVALITA 2018 Hate Speech Detection Task (Facebook)

Link to publication: http://ceur-ws.org/Vol-2263/paper010.pdf
Link to data: http://www.di.unito.it/~tutreeb/haspeede-evalita18/data.html
Task description: Binary (Hate, Not), Within hate for Facebook only, strength (No hate, Weak hate, Strong hate) and theme ((1) religion, (2) physical and/or mental handicap, (3) socio-economic status, (4) politics, (5) race, (6) sex and gender, (7) Other)
Details of task: Religion, physical and/or mental handicap, socio-economic status, politics, race, sex and gender
Size of dataset: 4,000
Reference: Bosco, C., Dell’Orletta, F. and Poletto, F., 2018. Overview of the EVALITA 2018 Hate Speech Detection Task. In: EVALITA 2018-Sixth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. CEUR, pp.1-9.

Overview of the EVALITA 2018 Hate Speech Detection Task (Twitter)

Task description: Binary (Hate, Not), Within Hate For Twitter only Intensity (1-4 rating), Aggressiveness (No, Weak, Strong), Offensiveness (No, Weak, Strong), Irony (Yes, No)
Details of task: Group-directed

Automatic Misogyny Identification (AMI) at Evalita 2020

Link to publication: http://ceur-ws.org/Vol-2765/paper161.pdf
Link to data: https://github.com/dnozza/ami2020
Task description: Binary (misogyny / not), Binary (aggressive / not), Binary on synthetic fairness test (misogyny / not)
Size of dataset: 6,000 and 1,961 (synthetic fairness test)
Percentage abusive: 47% and 50% (synthetic fairness test)
Reference: Fersini, E., Nozza, D., and Rosso, P., 2020. AMI @ EVALITA2020: Automatic Misogyny Identification. In: Proceedings of the 7th evaluation campaign of Natural Language Processing and Speech tools for Italian (EVALITA 2020).

CONAN - COunter NArratives through Nichesourcing: a Multilingual Dataset of Responses to Fight Online Hate Speech (Italian)

Task description: Binary (Islamophobic, Not), Multi-topic (Culture, Economics, Crimes, Rapism, Terrorism, Women Oppression, History, Other/generic)
Size of dataset: 1,071

Creating a WhatsApp Dataset to Study Pre-teen Cyberbullying

Link to publication: https://www.aclweb.org/anthology/W18-5107
Link to data: https://github.com/dhfbk/WhatsApp-Dataset
Task description: Binary (Cyberbullying, Not)
Size of dataset: 14,600
Percentage abusive: 0.08
Level of annotation: Posts, structured into 10 chats, with token level information
Platform: Synthetic / Whatsapp
Reference: Sprugnoli, R., Menini, S., Tonelli, S., Oncini, F. and Piras, E., 2018. Creating a WhatsApp Dataset to Study Pre-teen Cyberbullying. In: Proceedings of the 2nd Workshop on Abusive Language Online (ALW2) Month: October. Brussels, Belgium: Association for Computational Linguistics, pp.51-59.

Results of the PolEval 2019 Shared Task 6:First Dataset and Open Shared Task for Automatic Cyberbullying Detection in Polish Twitter

Link to publication: http://poleval.pl/files/poleval2019.pdf
Link to data: http://poleval.pl/tasks/task6
Task description: Harmfulness score (three values), Multilabel from seven phenomena
Size of dataset: 10,041
Percentage abusive: 0.09
Language: Polish
Reference: Ogrodniczuk, M. and Kobyliński, L., 2019. Results of the PolEval 2019 Shared Task 6: First Dataset and Open Shared Task for Automatic Cyberbullying Detection in Polish Twitter. In: Proceedings of the PolEval 2019 Workshop. Warszawa: Institute of Computer Science, Polish Academy of Sciences.

Toxic Language Dataset for Brazilian Portuguese (ToLD-Br)

Link to publication: https://arxiv.org/abs/2010.04543
Link to data: https://github.com/JAugusto97/ToLD-Br
Task description: Multiclass (LGBTQ+phobia, Insult, Xenophobia, Misogyny, Obscene, Racism)
Details of task: Three annotators per example, demographically diverse selected annotators.
Size of dataset: 21.000
Percentage abusive: 44%
Language: Portuguese
Reference: João A. Leite, Diego F. Silva, Kalina Bontcheva, Carolina Scarton (2020): Toxic Language Detection in Social Media for Brazilian Portuguese: New Dataset and Multilingual Analysis. AACL-IJCNLP 2020

A Hierarchically-Labeled Portuguese Hate Speech Dataset

Link to publication: https://www.aclweb.org/anthology/W19-3510
Link to data: https://b2share.eudat.eu/records/9005efe2d6be4293b63c3cffd4cf193e
Task description: Binary (Hate, Not), Multi-level (81 categories, identified inductively; categories have different granularities and content can be assigned to multiple categories at once)
Details of task: Multiple identities inductively categorized
Size of dataset: 3,059
Reference: Fortuna, P., Rocha da Silva, J., Soler-Company, J., Warner, L. and Nunes, S., 2019. A Hierarchically-Labeled Portuguese Hate Speech Dataset. In: Proceedings of the Third Workshop on Abusive Language Online. Florence, Italy: Association for Computational Linguistics, pp.94-104.

Offensive Comments in the Brazilian Web: A Dataset and Baseline Results

Link to publication: http://www.each.usp.br/digiampietri/BraSNAM/2017/p04.pdf
Link to data: https://github.com/rogersdepelle/OffComBR
Task description: Binary (Offensive, Not), Target (Xenophobia, homophobia, sexism, racism, cursing, religious intolerance)
Details of task: Religion/creed, Race/ethnicity, Physical/disability, Gender/sexual orientation
Size of dataset: 1,250
Platform: g1.globo.com
Reference: de Pelle, R. and Moreira, V., 2017. Offensive Comments in the Brazilian Web: A Dataset and Baseline Results. In: VI Brazilian Workshop on Social Network Analysis and Mining. SBC.

Offensive Language Identification Dataset for Brazilian Portuguese (OLID-BR)

Link to publication: https://link.springer.com/article/10.1007/s10579-023-09657-0
Link to data: https://huggingface.co/datasets/dougtrajano/olid-br
Task description: Binary (Offensive, Not), Multi-label (11 labels), Targeted or Not, Target Type (individual, group, other), Offensive Spans Identification
Details of task: Religion/creed, Race/ethnicity, Physical/disability, Gender/sexual orientation, Profanity/obscene.
Size of dataset: 7,943
Percentage abusive: 0.85
Platform: Twitter, YouTube, and other Portuguese Datasets
Reference: Trajano, D., Bordini, R.H. & Vieira, R. OLID-BR: offensive language identification dataset for Brazilian Portuguese. Lang Resources & Evaluation (2023).

Automatic Toxic Comment Detection in Social Media for Russian

Link to publication: https://github.com/alla-g/toxicity-detection-thesis/blob/main/toxicity_corpus/DATASTATEMENT.md
Link to data: https://github.com/alla-g/toxicity-detection-thesis/blob/main/toxicity_corpus/russian_distorted_toxicity.tsv
Task description: Toxicity - binary (1 == toxic, 0 == not toxic), Distortion - binary (1 == has distortion, 0 == does not have distortion),
Details of task: 1) multitask Russian toxicity detection with distortion detection as an auxiliary task; 2) testing toxicity classifiers on parallel distorted and manually corrected data
Size of dataset: 3000 texts: 561 toxic, 2439 not toxic; 126 distorted, 2874 not distorted.
Percentage abusive: 18.7%
Language: Russian
Level of annotation: comment
Platform: VKontakte
Reference: Gorbunova, A. (2022). Automatic Toxic Comment Detection in Social Media for Russian [Unpublished bachelor’s thesis]. National Research University Higher School of Economics.

Reducing Unintended Identity Bias in Russian Hate Speech Detection

Link to publication: https://aclanthology.org/2020.alw-1.8.pdf
Link to data: License Required (Last checked 17/01/2022)
Details of task: Toxicity, Harassment, Sexism, Homophobia, Nationalism
Reference: Zueva, Nadezhda, et al, Oct. 2020. Reducing Unintended Identity Bias in Russian Hate Speech Detection. In: Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 65–69

Detection of Abusive Speech for Mixed Sociolects of Russian and Ukrainian Languages

Link to publication: https://nlp.fi.muni.cz/raslan/2018/paper04-Andrusyak.pdf
Link to data: https://github.com/bohdan1/AbusiveLanguageDataset
Task description: Binary (True == Abusive, False == Not)
Details of task: Multilingual, Abusive Words, Political
Size of dataset: 2,000
Language: Surzhyk (Russian & Ukranian)
Reference: Andrusyak, B., Rimel, M. and Kern, R., 2018. Detection of Abusive Speech for Mixed Sociolects of Russian and Ukrainian Languages. In: Proceedings of Recent Advances in Slavonic Natural Language Processing, RASLAN 2018, pp. 77–84, 2018.

Russian South Park

Link to publication: https://aclanthology.org/2021.bsnlp-1.3/
Link to data: https://github.com/Sariellee/Russan-Hate-speech-Recognition
Task description: Binary (abusive, non-abusive)
Details of task: Abusive language in Russian South Park scripts
Size of dataset: 1400
Percentage abusive: 22.2%
Level of annotation: Sentence
Platform: TV Subtitles
Reference: Saitov & Derczynski, 2021. “Abusive Language Recognition in Russian”. Proceedings of the 8th BSNLP Workshop on Balto-Slavic Natural Language Processing, ACL
Link to data: http://hdl.handle.net/11356/1201
Size of dataset: 7,600,000
Language: Slovene
Platform: MMC RTV website

Overview of MEX-A3T at IberEval 2018: Authorship and Aggressiveness Analysis in Mexican Spanish Tweets

Link to publication: http://ceur-ws.org/Vol-2150/overview-mex-a3t.pdf
Link to data: https://mexa3t.wixsite.com/home/aggressive-detection-track
Task description: Binary (Aggressive, Not)
Size of dataset: 11,000
Language: Spanish
Reference: Alvarez-Carmona, M., Guzman-Falcon, E., Montes-y-Gomez, M., Escalante, H., Villasenor-Pineda, L., Reyes-Meza, V. and Rico-Sulayes, A., 2018. Overview of MEX-A3T at IberEval 2018: Authorship and aggressiveness analysis in Mexican Spanish tweets. In: Proceedings of the Third Workshop on Evaluation of Human Language Technologies for Iberian Languages (IberEval 2018).

Overview of the Task on Automatic Misogyny Identification at IberEval 2018 (Spanish)

Link to data: https://amiibereval2018.wordpress.com/important-dates/data/
Task description: Binary (Misogyny, Not), 5 categories (Stereotype, Dominance, Derailing, Sexual harassment, Discredit), Target of misogyny (Active or Passive)
Size of dataset: 4,138
Percentage abusive: 0.5

hatEval, SemEval-2019 Task 5: Multilingual Detection of Hate Speech Against Immigrants and Women in Twitter (Spanish)

Link to data: competitions.codalab.org/competitions/19935
Size of dataset: 6,600
Size of dataset: 100k (7325 hate, 27140 offensive, 65535 none)
Language: Turkish

A Corpus of Turkish Offensive Language on Social Media

Link to publication: https://coltekin.github.io/offensive-turkish/troff.pdf
Size of dataset: 36232
Percentage abusive: 0.19
Reference: Çöltekin, C., 2020. A Corpus of Turkish Offensive Language on Social Media. In: Proceedings of the 12th International Conference on Language Resources and Evaluation.

Hate-Speech and Offensive Language Detection in Roman Urdu

Link to publication: https://www.aclweb.org/anthology/2020.emnlp-main.197/
Link to data: https://github.com/haroonshakeel/roman_urdu_hate_speech
Task description: There are 2 subtasks, Coarse-grained Classification(Hate-Offensive vs Normal) and Fine-grained classification( Abusive/Offensive, Sexism, Religious Hate, Profane, Normal)
Details of task: Binary classification + Hate-Offensive label is further broken down into 4 fine-grained labels
Size of dataset: 10041
Percentage abusive: 0.24%
Language: Urdu-English
Reference: Hammad Rizwan, Muhammad Haroon Shakeel, and Asim Karim. 2020. Hate-speech and offensive language detection in Roman Urdu. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2512–2522, Online. Association for Computational Linguistics.

Lists of abusive keywords

“The Weaponized Word offers several thousand discriminatory, derogatory and threatening terms across 125+ languages, available through a RESTful API. Access is free for most academic researchers and registered humanitarian nonprofits.”
Data link: weaponizedword.org
“HurtLex is a lexicon of offensive, aggressive, and hateful words in over 50 languages. The words are divided into 17 categories, plus a macro-category indicating whether there is stereotype involved.”
Data link: github.com/valeriobasile/hurtlex
Reference: Hurtlex: A Multilingual Lexicon of Words to Hurt , Proc. CLiC-it 2018
Data link: http://staffwww.dcs.shef.ac.uk/people/G.Gorrell/publications-materials/abuse-terms.txt
Reference: Twits, Twats and Twaddle: Trends in Online Abuse towards UK Politicians , Proc. ICWSM
You can also use the GATE abuse tagger, available at https://cloud.gate.ac.uk/shopfront/displayItem/gate-hate
Data link: https://github.com/uds-lsv/lexicon-of-abusive-words
Reference: Inducing a Lexicon of Abusive Words – A Feature-Based Approach , Proc. NAACL-HLT 2018
Data link: Reddit hate lexicon
Reference: You can’t stay here: the efficacy of Reddit’s 2015 ban examined through hate speech , Proc. ACL Hum-Comput Interact.
SexHateLex is a Chinese lexicon of hateful and sexist words.
Data link: SexHateLex
Size of lexicon: 3,016
Reference: SWSR: A Chinese Dataset and Lexicon for Online Sexism Detection , Journal of OSNEM, Vol.27, 2022, 100182, ISSN 2468-6964.

How to Contribute

We accept entries to our catalogue based on pull requests to the README.md file. The dataset must be avaliable for download to be included in the list. If you want to add an entry, follow these steps!

Please send just one dataset addition/edit at a time - edit it in, then save. This will make everyone’s life easier (including yours!)
Go to the README.md file and click the edit button in the top right corner of the file.

Edit the markdown file. Please first go the correct language. The items are then sorted by their publication date (newest first). Add your item by copy and pasting the following template and adding all the details:
Check the “Preview Changes” tab to confirm everything is good to go!
If you’re ready to submit, propose the changes. Make sure you give some brief detail on the proposed change.

Submit the pull request on the next page when prompted.

This page is http://hatespeechdata.com/ .

Saved searches

Use saved searches to filter your results more quickly.

To see all available qualifiers, see our documentation .

Notifications You must be signed in to change notification settings

Datasets for Hate Speech Detection

aymeam/Datasets-for-Hate-Speech-Detection

Folders and files, repository files navigation, datasets from related literature.

In this repository , we present information on datasets that have been used for hate speech detection or related concepts such as cyberbullying , abusive language , online harassment , among others, to make it easier for researchers to obtain datasets.

Even when there are several social media platforms to get data from, the construction of a balanced labeled dataset is a costly task in time and effort, and it is still a problem for the researchers in the area. Although most of the below-listed datasets are not explicitly available, some of them can be obtained from the authors if requested.

Other languages

Multilingual (parallel data), multimodal datasets.

An official website of the United States government

Official websites use .gov A .gov website belongs to an official government organization in the United States.

Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.

Publications
Account settings
Advanced Search
Journal List

A curated dataset for hate speech detection on social media text

Devansh mody, yidong huang, thiago eustaquio alves de oliveira.

Author information
Article notes
Copyright and License information

Corresponding author. [email protected] @ModyDevansh

Received 2022 Oct 18; Revised 2022 Dec 3; Accepted 2022 Dec 12; Collection date 2023 Feb.

This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

Social media platforms have become the most prominent medium for spreading hate speech, primarily through hateful textual content. An extensive dataset containing emoticons, emojis, hashtags, slang, and contractions is required to detect hate speech on social media based on current trends. Therefore, our dataset is curated from various sources like Kaggle, GitHub, and other websites. This dataset contains hate speech sentences in English and is confined into two classes, one representing hateful content and the other representing non-hateful content. It has 451,709 sentences in total. 371,452 of these are hate speech, and 80,250 are non-hate speech. An augmented balanced dataset with 726,120 samples is also generated to create a custom vocabulary of 145,046 words. The total number of contractions considered in the dataset is 6403. The total number of bad words usually used in hateful content is 377. The text in each sentence of the final dataset, which is utilized for training and cross-validation, is limited to 180 words. The generated contractions dataset can be used for any projects in the area of NLP for data preprocessing. The augmented dataset can help to reduce the number of out-of-vocabulary words, and the hate speech dataset can be used as a classifier to detect hate or no hate on social media platforms.

Keywords: Hate speech, Cyberhate, Cyberbullying, Natural language processing, Online Hate

Specifications table

Value of the Data

This dataset is useful for training machine learning models to identify hate speech on social media in text. It reflects current social media trends and the modern ways of writing hateful text, using emojis, emoticons, or slang. It will help social media managers, administrators, or companies develop automatic systems to filter out hateful content on social media by identifying a text and categorizing it as hateful or non-hateful speech.

Deep Learning (DL) and Natural Language Processing (NLP) practitioners can be the target beneficiaries as this dataset can be used for detecting hateful speech through DL and NLP techniques. Here the samples are composed of text sentences and labels belonging to two categories “0″ for non-hateful and “1″ for hateful.

Additionally, this data set can be used as a benchmark data set to detect hate speech

The final preprocessed data set is neutralized in such a way that it can be used by anyone as it doesn't include any entities or names which can have an impact or cyber harm on the user that generated the content. Researchers can take advantage of the pre-processed dataset for their projects as it maintains and follows the policy guidelines.

The dataset presented here provides an alternative to smaller, more specialized datasets like the ones in [2 , 3] , and [4] .

1. Objective

Nowadays, social media text comprises slang, emojis, and emoticons. The conceived dataset aims to handle these diverse input types embedded in the text and enable the training of more effective Artificial Intelligence Models to identify hate speech. The previously available datasets from single sources were insufficient to represent the variations in hate speech content, so the new dataset aggregates and annotates data collected from various sources.

2. Data Description

Dataset used to identify hate speech has 451,709 samples in total. 371,452 of these are non-hateful speech, and 80,250 are hateful speech. An augmented data to generate vocabulary has 726,129 samples. The number of words in the vocabulary is 145,046. The number of entries in the contraction dictionary is 6403. The total number of hateful bad words is 377. The length of the text is set to a maximum of 180 words in our final dataset. Table 1 brings a thorough description of the data.

Columns in the dataset and its descriptions.

The dataset made available also contains a contraction dictionary used to process the raw data into the final pre-processed dataset. The contraction dictionary has the format presented in Table 2 .

Sample contractions and their possible expanded forms from the contractions dictionary.

The dataset also contains a list of bad words usually employed in hateful content, e.g., Fuck, cum and others.

In the raw data, the proportion of hateful and not-hateful samples is 17.8% and 82.2%.

3. Experimental Design, Materials and Methods

The final pre-processed dataset containing the aggregated and annotated sentences was acquired by applying the following method:

In the data made available in the online repository, the 0_RawData folder contains data collected from the different sources listed in the specification table to assemble a dataset of sentences.

A dictionary of contractions and a list of profanities commonly used in the English language by internet users is made available in the folder 1_ContractionProfanitiesEnglish .

(a) Emojis and emoticons from the raw data were converted to text and hyperlinks, user mentions, multiple spaces new line characters were removed from the sentences.
(b) After all these elements were removed, contractions found in the text were then expanded. Grammatical errors were corrected using the word mover's distance between sentences generated from the multiple possibilities for expanded contractions using the Google News Word2Vectors [10] and the open-source Gensim model [11] .
(c) The date and time occurrences were removed.
(d) Accented numbers and characters were also removed, along with the ampersand symbols in the beginning of some words.
(e) The remaining numbers were converted to words.
(f) Special characters (_"\-;%()|+&=*%.,!?:#$@[]/) were then removed from the text and an intermediate file was generated ( preprocessed_data_yidong_devansh.csv ) .
(g) From this file, the misspelled profanities were replaced with the correctly spelled versions based on the Profanities.csv file to generate the final preprocessed dataset ( final_preprocessed_data_yidong_devansh.csv ) .
(h) Duplicated entries in the final_preprocessed_data_yidong_devansh.csv were also removed before data augmentation and class balancing procedures.

In the 3_DataAugmentationAndBERTVocab folder, the raw datafile Final_Y_D_data.csv was used to generate a balanced version of the dataset ( YD_aug_data_balanced.csv ) by a ) undersampling the class with the majority of samples (non-hateful class); and b ), augmenting sentences from the hateful class using the contextual word embeddings from BERT models [12] with substitution and insertion methods as well as the synonym augmentation using WordNet embeddings [13] .

In the 4_PretrainedBERT folder, a custom BERT tokenizer, vocabulary, and configuration are made available to NLP practitioners.

The training data and validation folds for a 5-Fold cross-validation scheme can be found in the 5_TrainValidationFolds folder. After the preprocessed dataset is created, the samples are randomly shuffled and partitioned into five different folds. These folds are stratified to contain a roughly balanced number of samples for both classes.

6_MissclassifiedBERT2DataFolds Contains sentences that were misclassified by the HS-Bert Model and their true labels. The HS-BERT is a fine-tuned BERT model for hate speech detection. This model was trained on the 5-Fold Cross Validation data. The samples misclassified during training and validation were selected to compose a new data folder containing sentences that may be hard for such models to classify. Other researchers may use these sentences to improve their models in active learning schemes.

Ethics Statements

The data in the manuscript have been acquired via web scraping. The Terms of service (ToS) for all web resources listed in the specification table and used in the curated dataset allow the scrapping and distribution of data. In terms of copyright, the data curated belongs to web resources and not the users of those resources. No news website was used in the curated dataset. This dataset protects the privacy of individuals. No one's personal information was taken during the data collection. Identities of entities were removed if they occurred in the dataset, and the dataset was neutralized to comply with legal and ethical standards and policy regarding the use of social media data for research purposes. The purpose of the task is to detect hate speech, not to target a particular user or group of users. The dataset collected was curated from publicly available sources. The present work does not contain data scrapped directly from social media platforms (e.g., Twitter and Facebook); thus, it does not violate the scrapping policies of such platforms.

CRediT Author Statement

Devansh Mody: Conceptualization, Methodology, Software, Validation, Investigation, Resources, Data Curation, Writing – original draft, Writing – review & editing; YiDong Huang: Conceptualization, Methodology, Software, Validation, Investigation, Resources, Data Curation, Writing – original draft, Writing – review & editing; Thiago Eustaquio Alves de Oliveira: Methodology, Validation, Investigation, Resources, Writing – original draft, Writing – review & editing, Supervision, Project administration, Funding acquisition.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

Funding: This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) - Discovery Grant [grant number RGPIN-2020-04309], Lakehead University SRC and Start-up Grant, Faculty of Science and Environmental and Start-up Grant.

Data Availability

A Curated Hate Speech Dataset (Original data) (Mendeley Data).

1. Thomas Davidson, Dana Warmsley, Michael Macy and Ingmar Weber, “Automated Hate Speech Detection and the Problem of Offensive Language”, arXiv:1703.04009v1 [cs.CL] 11 Mar 2017.
2. Jing Qian, Anna Bethke, Yinyin Liu, Elizabeth Belding, and William Yang Wang, “A Benchmark Dataset for Learning to Intervene in Online Hate Speech”, arXiv:1909.04251v1 [cs.CL] 10 Sep 2019.
3. Mathew B., Saha P., Yimam S., Biemann C., Goyal P., and Mukherjee A., “HateXplain: A Benchmark Dataset for Explainable Hate Speech Detection”, ArXiv: 2012.10289 18 Dec 2020.
4. Ona de Gibert, Naiara Perez, AitorGarcıa-Pablos and MontseCuadros, “Hate Speech Dataset from a White Supremacy Forum”, arXiv:1809.04444 [cs.CL] 12 Sep 2018.
5. Nikola Ljubeˇsi´, Darja Fiˇser, and Tomaˇz Erjavec, “The FRENK Datasets of Socially Unacceptable Discourse in Slovene and English”, arXiv:1906.02045v2 [cs.CL] 13 Jun 2019.
6. Amir PouranBen Veyseh, FranckDernoncourt, Quan HungTran, and Thien HuuNguyen, “What Does This Acronym Mean? Introducing a New Dataset for Acronym Identification and Disambiguation”, arXiv:2010.14678v1 [cs.CL] 28 Oct 2020.
7. Mandl Thomas, Modha Sandip, Majumder Prasenjit, Patel Daksh, Dave Mo-hana, Mandlia Chintak, Patel Aditya. FIRE ’19: Proceedings of the 11th Forum for Information Retrieval Evaluation December. 2019. Overview of the HASOC track at FIRE 2019: hate speech and offensive content identification In Indo-European languages; pp. 14–17. Pages. [ Google Scholar ]
8. Francesco Barbieri, Jose Camacho-Collados, Leonardo Neves, Luis Espinosa-Anke, “TweetEval: Unified Benchmark and Comparative Evaluation for Tweet Classification”, arXiv:2010.12421v2 [cs.CL] 26 Oct 2020.
9. Nedjma Ousidhoum, Zizheng Lin, Hongming Zhang, Yangqiu Song, Dit-Yan Yeung, “Multilingual and Multi-Aspect Hate Speech Analysis”, arXiv:1908.11049v1 [cs.CL] 29 Aug 2019.
10. Google Code Archive - Long-term storage for Google Code Project Hosting. (n.d.). Retrieved October 8, 2022, from https://code.google.com/archive/p/word2vec/
11. Rehurek Radim, Sojka Petr. Proceedings of the LREC 2010 workshop on new challenges for NLP frameworks. 2010. Software framework for topic modeling with large corpora. [ Google Scholar ]
12. Geetha M.P., Karthika Renuka D. Improving the performance of aspect based sentiment analysis using fine-tuned Bert Base Uncased model. Int. J. Intell. Netw. 2021;2:64–69. [ Google Scholar ]
13. Fellbaum Christiane. Theory and Applications of Ontology: Computer Applications. Springer; Dordrecht: 2010. WordNet; pp. 231–243. [ Google Scholar ]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

View on publisher site
PDF (397.4 KB)
Collections

Add to Collections

News Tracker
Top 100 Kenyans 2024
Feel Good Stories

NCIC Classifies These 23 Words as Hate Speech [LIST]

by Geoffrey Lutta on Friday, 8 April 2022 - 1:45 pm

The National Cohesion and Integration Commission (NCIC) has released a list of lexicon words it construes as hate speech or bordering incitement to violence.

NCIC chairperson, Rev. Samuel Kobia on Friday, April 8, explained that the lexicon words singled out by the Commission tend to provoke violence among different communities when used.

The words listed are in different dialects including English, Swahili, Sheng, and a variety of local languages.

The words banned for use in English include fumigation, eliminate and kill.

Words banned in Swahili include Kaffir (derived from the Arabic term Kafir which means disbeliever or one who conceals the truth), madoadoa (dots), chunga kura (secure the vote), mende (cockroach ), watu wa kurusha mawe (people who throw stones), watajua hawajui (they will know that they do not know), wabara waende kwao (people from off the coast should go back to their homes), wakuja (those that come), Chinja Kafir (kill the infidel) and kwekwe (weeds).

Sheng words which have been flagged include Kama noma noma, Kama mbaya mbaya (If it is bad, then it is bad), Hatupangwingwi (No one can arrange us) and operation Linda Kura (secure the vote).

In Kikuyu, the words flagged include Kihii (uncircumcised man), Uthamaki ni witu (the Kingdom is ours), and Mwiji in Kimeru (uncircumcised man).

Others include Kimurkeldet (brown teeth), Otutu labotonik (uproot the weed), and Ngetiik (Uncircumcised).

"The NCIC has classified terms, which have been regularly used in Kenya’s political landscape with the intent to provoke violence among various communities of diverse political viewpoints," Kobia stated.

"We have classified these terms as coded messages which can be used to activate hatred, and consciously eliminate other communities. These terms are in various languages including English, Kiswahili, Sheng, Kikuyu, Kalenjin and Non-verbal nods."

NCIC flagged different social media platforms as the main outlets where hate speech is mostly propagated. Kobia named Kenya Kwanza Alliance and Azimio La Umoja One Kenya as the main perpetrators of inciteful words.

He highlighted 39 social forums which are under active investigation and other 49 cases which relevant government authorities are probing.

He, however, emphasised that the commission will focus on social monitoring to ensure peaceful elections.

"We monitored Facebook, Twitter, YouTube among other social media platforms. The highest number that we identified to be spreading hate/incitement was Facebook followed by Twitter. In addition, we also received emails from the general public reporting complaint," Kobia remarked.

Angry Residents Attack Police, Chop Off Officer’s Hand

A collage of former Gender CS Aisha Jumwa and Millicent Omanga, January 17, 2025.

Aisha Jumwa, Millicent Omanga Score Big in State Appointments

Why UAE Backed SGR Deal Is Unlike China’s

UDA Merges With Another Party

President William Ruto speaking during the swearing of CSs Lee Kinyanjui, William Kabogo and Mutahi Kagwe at State House Nairobi, January 17,2025.

UDA Senator Calls for Immediate Resignation of Ruto

Multi-Trillion Chinese Company to Create Jobs for Kenyans

(Left to Right) General Service Unit Commandant Ranson Lolmodooni, DCI Director Amin Mohammed and Inspector Genral of Police Douglas Kanja during a press breifing in Nairobi on December 4 at Harrambee house.

Why IG Kanja and DCI Amin Skipped Court in Abduction Case

Former President Uhuru Kenyatta during an address

Fight for Your Rights: Uhuru's Bold Call to Gen Zs

Kenya Power Sends Message to Kenyans Over Charging Phones Overnight

Residents queue to submit bursary application forms at the NG-CDF Likoni offices on January 13, 2021.

Council of Governors Clarifies Issuance of Bursaries After Restriction

Health Cabinet Secretary Deborah Barasa, with Kenyatta National Hospital CEO Dr Evanson Kamuri, interacting with a mother at the hospital, January 16, 2025.

Health Crisis to Hit Hospitals Next Week

Ruto's First 4 Instructions to Mutahi Kagwe

PeaceTech Lab | Putting the Right Tools in the Right Hands to Build Peace

Our Hate Speech Lexicons

PeaceTech Lab’s hate speech Lexicons identify and explain inflammatory language on social media while offering alternative words and phrases that can be used to combat the spread of hate speech. Our Lexicons serve as a pivotal resource for local activists and organizations working to stop and prevent hate speech worldwide.

Click on an image below to read a Lexicon.

Hate Speech Dataset Catalogue

This page catalogues datasets annotated for hate speech, online abuse, and offensive language. They may be useful for e.g. training a natural language processing system to detect this language.

The list is maintained by Leon Derczynski , Bertie Vidgen , Hannah Rose Kirk , Pica Johansson, Yi-Ling Chung , Mads Guldborg Kjeldgaard Kongsbak, Laila Sprejer , and Philine Zeinert.

We provide a list of datasets and keywords . If you would like to contribute to our catalogue or add your dataset, please see the instructions for contributing .

If you use these resources, please cite (and read!) our paper: Directions in Abusive Language Training Data: Garbage In, Garbage Out . And if you would like to find other resources for researching online hate, visit The Alan Turing Institute's Online Hate Research Hub or read The Alan Turing Institute's Reading List on Online Hate and Abuse Research .

If you're looking for a good paper on online hate training datasets (beyond our paper, of course!) then have a look at 'Resources and benchmark corpora for hate speech detection: a systematic review' by Poletto et al. in Language Resources and Evaluation .

Accompanying data statements preferred for all corpora.

How to contribute

We accept entries to our catalogue based on pull requests to the content folder. The dataset must be avaliable for download to be included in the list. If you want to add an entry, follow these steps!

Please send just one dataset addition/edit at a time - edit it in, then save. This will make everyone’s life easier (including yours!)

Create file

Go to the repo url file and click the "Add file" dropdown and then click on "Create new file".

Choose location

In the following page type content/datasets/<name-of-the-file>.md . if you want to add an entry to the datasets catalog or content/keywords/<name-of-the-file>.md if you want to add an entry to the lists of abusive keywords, if you want to just add an static page you can leave in the root of content it will automatically get assigned an url eg: /content/about.md becomes the /about page

Fill in content

Copy the contents of templates/dataset.md or templates/keywords.md respectively to the camp below, filling out the fields with the correct data format. Everything below the second --- will automatically get rendered into the page, so you may add any standard markdown fields e.g tables, headings, lists...

Commit changes

Click on "Commit changes", on the popup make sure you give some brief detail on the proposed change. and then click on Propose changes

Submit the pull request on the next page when prompted.

Search for datasets

Abusive Language Detection on Arabic Social Media (Al Jazeera)

Link to publication: https://www.aclweb.org/anthology/W17-3008

Link to data: http://alt.qcri.org/~hmubarak/offensive/AJCommentsClassification-CF.xlsx

Task Description: Ternary (Obscene, Offensive but not obscene, Clean)

Details of Task: Incivility

Size of Dataset: 32000

Percentage Abusive: 0.81 %

Language: Arabic

Level of Annotation: Posts

Platform: AlJazeera

Medium: Text

Reference: Mubarak, H., Darwish, K. and Magdy, W., 2017. Abusive Language Detection on Arabic Social Media. In: Proceedings of the First Workshop on Abusive Language Online. Vancouver, Canada: Association for Computational Linguistics, pp.52-56.

Large-Scale Hate Speech Detection with Cross-Domain Transfer

Link to publication: https://aclanthology.org/2022.lrec-1.238/

Link to data: https://github.com/avaapm/hatespeech

Task Description: Three-class (Hate speech, Offensive language, None)

Details of Task: Hate speech detection on social media (Twitter) including 5 target groups (gender, race, religion, politics, sports)

Size of Dataset: 100k English (27593 hate, 30747 offensive, 41660 none)

Percentage Abusive: 58.3 %

Language: English

Platform: Twitter

Medium: Text, Image

Reference: Cagri Toraman, Furkan Şahinuç, Eyup Yilmaz. 2022. Large-Scale Hate Speech Detection with Cross-Domain Transfer. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2215–2225, Marseille, France. European Language Resources Association.

CoRAL: a Context-aware Croatian Abusive Language Dataset

Link to publication: https://aclanthology.org/2022.findings-aacl.21/

Link to data: https://github.com/shekharRavi/CoRAL-dataset-Findings-of-the-ACL-AACL-IJCNLP-2022

Task Description: Multi-class based on context dependency categories (CDC)

Details of Task: Detectioning CDC from abusive comments

Size of Dataset: 2240

Percentage Abusive: 100 %

Language: Croatian

Platform: Posts

Medium: Newspaper Comments

Reference: Ravi Shekhar, Mladen Karan and Matthew Purver (2022). CoRAL: a Context-aware Croatian Abusive Language Dataset. Findings of the ACL: AACL-IJCNLP.

AbuseEval v1.0

Link to publication: http://www.lrec-conf.org/proceedings/lrec2020/pdf/2020.lrec-1.760.pdf

Link to data: https://github.com/tommasoc80/AbuseEval

Task Description: Explicitness annotation of offensive and abusive content

Details of Task: Enriched versions of the OffensEval/OLID dataset with the distinction of explicit/implicit offensive messages and the new dimension for abusive messages. Labels for offensive language: EXPLICIT, IMPLICT, NOT; Labels for abusive language: EXPLICIT, IMPLICT, NOTABU

Size of Dataset: 14100

Percentage Abusive: 20.75 %

Level of Annotation: Tweets

Reference: Caselli, T., Basile, V., Jelena, M., Inga, K., and Michael, G. 2020. "I feel offended, don’t be abusive! implicit/explicit messages in offensive and abusive language". The 12th Language Resources and Evaluation Conference (pp. 6193-6202). European Language Resources Association.

Let-Mi: An Arabic Levantine Twitter Dataset for Misogynistic Language

Link to publication: https://arxiv.org/abs/2103.10195

Link to data: https://drive.google.com/file/d/1mM2vnjsy7QfUmdVUpKqHRJjZyQobhTrW/view

Task Description: Binary (misogyny/none) and Multi-class (none, discredit, derailing, dominance, stereotyping & objectiﬁcation, threat of violence, sexual harassment, damning)

Details of Task: Introducing an Arabic Levantine Twitter dataset for Misogynistic language

Size of Dataset: 6603

Percentage Abusive: 48.76 %

Medium: Text, Images

Reference: Hala Mulki and Bilal Ghanem. 2021. Let-Mi: An Arabic Levantine Twitter Dataset for Misogynistic Language. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 154–163, Kyiv, Ukraine (Virtual). Association for Computational Linguistics

Offensive Language and Hate Speech Detection for Danish

Link to publication: http://www.derczynski.com/papers/danish_hsd.pdf

Link to data: https://figshare.com/articles/Danish_Hate_Speech_Abusive_Language_data/12220805

Task Description: Branching structure of tasks: Binary (Offensive, Not), Within Offensive (Target, Not), Within Target (Individual, Group, Other)

Details of Task: Group-directed + Person-directed

Size of Dataset: 3600

Percentage Abusive: 0.12 %

Language: Danish

Platform: Twitter, Reddit, Newspaper comments

Reference: Sigurbergsson, G. and Derczynski, L., 2019. Offensive Language and Hate Speech Detection for Danish. ArXiv.

Detecting Abusive Albanian

Link to publication: https://arxiv.org/abs/2107.13592

Link to data: https://doi.org/10.6084/m9.figshare.19333298.v1

Task Description: Hierarchical (offensive/not; untargeted/targeted; person/group/other)

Details of Task: Detect and categorise abusive language in social media data

Size of Dataset: 11874

Percentage Abusive: 13.2 %

Language: Albanian

Platform: Instagram, Youtube

Reference: Nurce, E., Keci, J., Derczynski, L., 2021. Detecting Abusive Albanian. arXiv:2107.13592

Hate Speech Detection in the Bengali language: A Dataset and its Baseline Evaluation

Link to publication: https://arxiv.org/pdf/2012.09686.pdf

Link to data: https://www.kaggle.com/naurosromim/bengali-hate-speech-dataset

Task Description: Binary (hateful, not)

Details of Task: Several categories: sports, entertainment, crime, religion, politics, celebrity and meme

Size of Dataset: 30000

Percentage Abusive: 0.33 %

Language: Bengali

Platform: Youtube, Facebook

Reference: Romim, N., Ahmed, M., Talukder, H., & Islam, M. S. (2021). Hate speech detection in the bengali language: A dataset and its baseline evaluation. In Proceedings of International Joint Conference on Advances in Computational Intelligence (pp. 457-468). Springer, Singapore.

Measuring Hate Speech

Link to publication: https://arxiv.org/abs/2009.10277

Link to data: https://huggingface.co/datasets/ucberkeley-dlab/measuring-hate-speech

Task Description: 10 ordinal labels (sentiment, (dis)respect, insult, humiliation, inferior status, violence, dehumanization, genocide, attack/defense, hate speech), which are debiased and aggregated into a continuous hate speech severity score (hate_speech_score) that includes a region for counterspeech & supportive speeech. Includes 8 target identity groups (race/ethnicity, religion, national origin/citizenship, gender, sexual orientation, age, disability, political ideology) and 42 identity subgroups.

Details of Task: Hate speech measurement on social media in English

Size of Dataset: 39,565 comments annotated by 7,912 annotators on 10 ordinal labels, for 1,355,560 total labels.

Percentage Abusive: 25 %

Level of Annotation: Social media comment

Platform: Twitter, Reddit, Youtube

Reference: Kennedy, C. J., Bacon, G., Sahn, A., & von Vacano, C. (2020). Constructing interval variables via faceted Rasch measurement and multitask deep learning: a hate speech application. arXiv preprint arXiv:2009.10277.

Lists of Abusive Keywords

List Description: HurtLex is a lexicon of offensive, aggressive, and hateful words in over 50 languages. The words are divided into 17 categories, plus a macro-category indicating whether there is stereotype involved.

Data Link: https://github.com/valeriobasile/hurtlex

Reference: http://ceur-ws.org/Vol-2253/paper49.pdf, Proc. CLiC-it 2018

SexHateLex is a Chinese lexicon of hateful and sexist words.

Data Link: https://doi.org/10.5281/zenodo.4773875

Reference: http://ceur-ws.org/Vol-2253/paper49.pdf, Journal of OSNEM, Vol.27, 2022, 100182, ISSN 2468-6964.

Subscribe to the PwC Newsletter

Join the community, edit dataset, edit dataset tasks.

Some tasks are inferred based on the benchmarks list.

Add a Data Loader

Remove a data loader.

huggingface/datasets -
t-davidson/hate-speech-and-offensive-language -

Edit Dataset Modalities

Edit dataset languages, edit dataset variants.

The benchmarks section lists all benchmarks using a given dataset or any of its variants. We use variants to distinguish between results evaluated on slightly different versions of the same dataset. For example, ImageNet 32⨉32 and ImageNet 64⨉64 are variants of the ImageNet dataset.

Add a new evaluation result row

Hate speech and offensive language.

HSOL is a dataset for hate speech detection. The authors begun with a hate speech lexicon containing words and phrases identified by internet users as hate speech, compiled by Hatebase.org. Using the Twitter API they searched for tweets containing terms from the lexicon, resulting in a sample of tweets from 33,458 Twitter users. They extracted the time-line for each user, resulting in a set of 85.4 million tweets. From this corpus they took a random sample of 25k tweets containing terms from the lexicon and had them manually coded by CrowdFlower (CF) workers. Workers were asked to label each tweet as one of three categories: hate speech, offensive but not hate speech, or neither offensive nor hate speech.

Benchmarks Edit Add a new result Link an existing benchmark

Dataset loaders edit add remove.

Similar Datasets

License edit, modalities edit, languages edit.

IMAGES

Hate Speech
List of hate speech terms released by NCIC
Is hate speech legal?
GitHub
Hate Speech on Behance
Hate Speech on Behance

COMMENTS

Hate Speech Dataset Catalogue | hatespeechdata">Hate Speech Dataset Catalogue | hatespeechdata
Hate Speech Dataset Catalogue. This page catalogues datasets annotated for hate speech, online abuse, and offensive language. They may be useful for e.g. training a natural language processing system to detect this language.
Datasets-for-Hate-Speech-Detection - GitHub">aymeam/Datasets-for-Hate-Speech-Detection - GitHub
In this repository, we present information on datasets that have been used for hate speech detection or related concepts such as cyberbullying, abusive language, online harassment, among others, to make it easier for researchers to obtain datasets.
Hatebase
Hatebase was built to assist companies, government agencies, NGOs and research organizations moderate online conversations and potentially use hate speech as a predictor for regional violence. (Language-based classification, or symbolization, is one of a handful of quantifiable steps toward genocide.)
Hatebase">Search Results - Hatebase
A collaborative, regionalized repository of multilingual hate speech. Showing 1 - 50 of 3,893 results. Read FAQs
hate speech detection on social media text">A curated dataset for hate speech detection on social media text
Dec 22, 2024 · This dataset contains hate speech sentences in English and is confined into two classes, one representing hateful content and the other representing non-hateful content. It has 451,709 sentences in total. 371,452 of these are hate speech, and 80,250 are non-hate speech.
Words as Hate Speech [LIST]">NCIC Classifies These 23 Words as Hate Speech [LIST]
Apr 8, 2022 · President William Ruto (left) and ODM leader Raila Odinga (right) at rallies in West Pokot and Kajiado, respectively in January 2022. The National Cohesion and Integration Commission (NCIC) has released a list of lexicon words it construes as hate speech or bordering incitement to violence.
Hate Speech Lexicons | PeaceTech Lab">Hate Speech Lexicons | PeaceTech Lab
PeaceTech Lab’s hate speech Lexicons identify and explain inflammatory language on social media while offering alternative words and phrases that can be used to combat the spread of hate speech. Our Lexicons serve as a pivotal resource for local activists and organizations working to stop and prevent hate speech worldwide.
Hate Speech Dataset Catalogue">Hate Speech Dataset Catalogue
Hate Speech Dataset Catalogue. This page catalogues datasets annotated for hate speech, online abuse, and offensive language. They may be useful for e.g. training a natural language processing system to detect this language.
Hate Speech and Offensive Language Dataset - Papers With Code">Hate Speech and Offensive Language Dataset - Papers With Code
HSOL is a dataset for hate speech detection. The authors begun with a hate speech lexicon containing words and phrases identified by internet users as hate speech, compiled by Hatebase.org. Using the Twitter API they searched for tweets containing terms from the lexicon, resulting in a sample of tweets from 33,458 Twitter users.
Hate Speech Dataset - Mendeley Data">A Curated Hate Speech Dataset - Mendeley Data
Oct 3, 2022 · This dataset contains hate speech sentences in English. It has 451709 sentences in total. 371452 of these are hate speech, and 80250 are non-hate speech. The dataset is organized into folders as follows: 0_RawData contains data collected from different sources to assemble a dataset of hate speech sentences.