Elamur Research⌕
☰
home / papers / 2021 / A Large-Scale English Multi-Label Twitter Dataset for Cyberbullying and Online Abuse Detection

A Large-Scale English Multi-Label Twitter Dataset for Cyberbullying and Online Abuse Detection

Paper · 2021 · Semiu Salawu, Jo Lumsden and Yulan He

Summary

Existing cyberbullying corpora tend to be outdated, small, or dominated by harmless posts. Many also assign only one label per document, which hides the fact that a single message can be abusive in several ways at once. The authors build a new English Twitter corpus of 62,587 tweets. They collected it with deliberately targeted searches (profane terms, the #sarcasm hashtag, and violence-related hashtags found by snowballing), so that offensive material makes up about 82% of the data. Each tweet is annotated with a multi-label scheme of eight classes that covers both direct abuse and subtler relational bullying such as exclusion and ridicule through sarcasm. Four transformer classifiers (BERT, DistilBERT, RoBERTa, XLNet) were trained on the corpus, and RoBERTa performed best. In cross-domain tests, a RoBERTa model trained on this corpus scored better on the Davidson and Kaggle Toxic Comments datasets than models trained on those datasets themselves. This suggests the resource captures abuse signals that carry over to other platforms. The dataset is publicly released.

Key points

  • Size and labels: 62,587 tweets with eight labels: Bullying, Insult, Profanity, Sarcasm, Threat, Exclusion, Porn and Spam. A tweet may carry any combination of them.
  • Collection:
    • Searches used frequent profane terms from Wang et al. (2014) and the #sarcasm hashtag.
    • Threat hashtags were seeded from a 5,000-tweet Streaming API sample and then expanded by snowballing.
    • Exclusion queries had to be derived from only six example tweets.
  • Model results: the best macro ROC-AUC was about 0.97 (RoBERTa, pre-trained). The further fine-tuned variants scored slightly lower overall.
  • Cross-domain results:
    • On Davidson, RoBERTa C→D reached macro F1 0.880, against 0.843 for D→D.
    • On Kaggle, C→K reached 0.567, against 0.503 for K→K.
    • The reverse transfers (D→C, K→C) were weaker.
  • Known limits and availability: Sarcasm, Exclusion and Threat are sparse, so these classes are predicted less reliably. The data is hosted at bitbucket.org/ssalawu/cyberbullying-twitter.

Interesting ideas

  • I1. The authors fill rare classes by searching specifically for likely positives rather than oversampling a natural-distribution corpus, which gives the model many more real examples of the minority phenomenon.
  • I2. New search terms are discovered by snowballing: a few seed hashtags fetch tweets, and those tweets reveal further related hashtags, so a small seed grows into a broad set of queries.
  • I3. They chose a flat multi-label scheme over a hierarchical one, arguing that a coarse first-level decision can silently drop valid finer labels.
  • I4. The taxonomy separates direct abuse from indirect, relational abuse such as exclusion and sarcastic ridicule, recognising that harm can occur without hostile vocabulary.
  • I5. Cross-dataset transfer is always reported next to in-domain baselines trained and tested on the target set, so a generalisation claim has a fair reference point.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): When silver data lacks rare constructions such as threats or ironic statements, targeted retrieval could gather candidates for the witness parsers and teachers to label, instead of duplicating the few cases already found.
  • I2 (learn): The same seed-and-expand loop could grow lists of cue words for rare speech acts or verb senses, with each expansion step logged so the provenance stays explainable.
  • I3 (frames): A single utterance should be allowed several simultaneous speech-act or pragmatic tags (for example insult plus sarcasm) rather than one forced choice through a decision tree.
  • I4 (frames): This argues for detecting irony and exclusion from implicature and from who addresses whom, not from lexical cues like swear words.
  • I5 (learn): Evaluations of Mova components on foreign treebanks or test suites should pair every transfer score with a native-trained control, which works as a built-in sanity check.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of WarwickCoventry · United Kingdom7 works · 1,291 citations🎓Aston UniversityBirmingham · United Kingdom4 works · 223 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksSocial Sciences · 1,772 worksSociology and Political Science · 577 works

🧭 Topics 2Research subjects and languages the work is about.English · 2,263 worksBias, fairness & safety · 517 works

Cite

Cite as Semiu Salawu, Jo Lumsden and Yulan He. 2021. A Large-Scale English Multi-Label Twitter Dataset for Cyberbullying and Online Abuse Detection. Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2021.woah-1.16

@inproceedings{salawu2021large,
  title = {{A Large-Scale English Multi-Label Twitter Dataset for Cyberbullying and Online Abuse Detection}},
  author = {Semiu Salawu, Jo Lumsden and Yulan He},
  booktitle = {Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021)},
  year = {2021},
  doi = {10.18653/v1/2021.woah-1.16},
  url = {https://doi.org/10.18653/v1/2021.woah-1.16}
}

References

38 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Mariek Vanden Abeele and Rozane De Cock. 2013. Cyberbullying by mobile phone among adolescents: The role of gender and peer group status. Communications, 38:107–118. ↗
  2. Emily M Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587–604. ↗
  3. Uwe Bretschneider, Thomas Wöhner, and Ralf Peters. 2014. Detecting online harassment in social networks. find
  4. David Van Bruwaene, Qianjia Huang, and Diana Inkpen. 2020. A multi-platform dataset for detecting cyberbullying in social media. Language Resources and Evaluation, pages 1–24. ↗
  5. Despoina Chatzakou, Nicolas Kourtellis, Jeremy Blackburn, Emiliano De Cristofaro, Gianluca Stringhini, and Athena Vakali. 2017. Mean birds: Detecting aggression and bullying on twitter. pages 13–22. Association for Computing Machinery, Inc. find
  6. David Cohen. 2015. Facebook changes definition of monthly active users. find
  7. Maral Dadvar and Franciska De Jong. 2012. Cyberbullying detection: A step toward a safer internet yard. pages 121–125. find
  8. Maral Dadvar, Dolf Trieschnigg, and Franciska De Jong. 2013. Expert knowledge for automatic detection of bullies in social networks. pages 57–64. find
  9. Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. 2017. Automated hate speech detection and the problem of offensive language. ↗
  10. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. 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), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. ↗
  11. Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2018. Measuring and mitigating unintended bias in text classification. volume 7, pages 67–73. Association for Computing Machinery, Inc. find
  12. Antigoni-Maria Founta, Constantinos Djouvas, Despoina Chatzakou, Ilias Leontiadis, Jeremy Blackburn, Gianluca Stringhini, Athena Vakali, Michael Sirivianos, and Nicolas Kourtellis. 2018. Large scale crowdsourcing and characterization of twitter abusive behavior. ↗
  13. Patrick J Grother, Mei L Ngan, and Kayee K Hanaoka. 2019. Face recognition vendor test (frvt) part 3: Demographic effects. find
  14. Cynthia Van Hee, Gilles Jacobs, Chris Emmery, Bart Desmet, Els Lefever, Ben Verhoeven, Guy De Pauw, Walter Daelemans, and Véronique Hoste. 2018. Automatic detection of cyberbullying in social media text. PLOS ONE, 13:e0203794. ↗
  15. Homa Hosseinmardi, Sabrina Arredondo Mattson, Rahat Ibn Rafiq, Richard Han, Qin Lv, and Shivakant Mishra. 2015. Poster: Detection of cyberbullying in a mobile social network: Systems issues. page 481. Association for Computing Machinery, Inc. find
  16. Dirk Hovy and Shannon L Spruit. 2016. The social impact of natural language processing. pages 591– 598.
  17. Qianjia Huang, Vivek Kumar Singh, and Pradeep Kumar Atrey. 2014. Cyber bullying detection using social and textual analysis. pages 3–6. ACM. find
  18. Kaggle. 2012. Detecting insults in social commentary. find
  19. Kaggle. 2018. Toxic comment classification challenge. find
  20. April Kontostathis, Kelly Reynolds, Andy Garron, and Lynne Edwards. 2013. Detecting cyberbullying: Query terms and techniques. volume volume, pages 195–204. Association for Computing Machinery. find
  21. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pre-training approach. Computing Research Repository, arXiv:1907.11692. ↗
  22. Vinita Nahar, Sanad Al-Maskari, Xue Li, and Chaoyi Pang. 2014. Semi-supervised learning for cyberbullying detection in social networks. volume 8506 LNCS, pages 160–171. Springer Verlag. find
  23. B. Sri Nandhini and J. I. Sheeba. 2015. Online social network bullying detection using intelligence techniques. volume 45, pages 485–492. Elsevier B.V. find
  24. Ofcom Research. 2019. Online nation. Ofcom Research. find
  25. Nedjma Ousidhoum, Zizheng Lin, Hongming Zhang, Yangqiu Song, and Dit-Yan Yeung. 2019. Multilingual and multi-aspect hate speech analysis. arXiv preprint arXiv:1908.11049. ↗
  26. Evani Radiya-Dixit and Xin Wang. 2020. How fine can fine-tuning be? learning efficient language models. ↗
  27. Ashwin Rajadesingan, Reza Zafarani, and Huan Liu. 2015. Sarcasm detection on twitter:a behavioral modeling approach. pages 97–106. Association for Computing Machinery, Inc. find
  28. Gathika Rathnayake, Thushari Atapattu, Mahen Herath, Georgia Zhang, and Katrina Falkner. 2020. Enhancing the identification of cyberbullying through participant roles. In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 89–94, Online. Association for Computational Linguistics. find
  29. Rahul Sambaraju and Chris McVittie. 2020. Examining abuse in online media. Social and personality psychology compass, 14(3):e12521. ↗
  30. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. Computing Research Repository, arXiv:1910.01108. ↗
  31. Peter K. Smith, Jess Mahdavi, Manuel Carvalho, Sonja Fisher, Shanette Russell, and Neil Tippett. 2008. Cyberbullying: Its nature and impact in secondary school pupils. Journal of Child Psychology and Psychiatry and Allied Disciplines, 49:376–385. ↗
  32. Rachele Sprugnoli, Stefano Menini, Sara Tonelli, Filippo Oncini, and Enrico Piras. 2018. Creating a WhatsApp dataset to study pre-teen cyberbullying. In Proceedings of the 2nd Workshop on Abusive Language Online (ALW2), pages 51–59, Brussels, Belgium. Association for Computational Linguistics. find
  33. Statista. 2019. Global mobile social penetration rate 2019, by region. find
  34. Bertie Vidgen and Leon Derczynski. 2020. Directions in abusive language training data, a systematic review: Garbage in, garbage out. Plos one, 15(12):e0243300. ↗
  35. Wenbo Wang, Lu Chen, Krishnaprasad Thirunarayan, and Amit P Sheth. 2014. Cursing in english on twitter. pages 415–425. find
  36. Zeerak Waseem and Dirk Hovy. 2016. Hateful symbols or hateful people? predictive features for hate speech detection on twitter. pages 88–93.
  37. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. Computing Research Repository, arXiv:1906.08237. ↗
  38. Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. Predicting the type and target of offensive posts in social media. volume 1, pages 1415–1420. Association for Computational Linguistics (ACL).

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/salawu-2021-large-scale-english-multi-label/