Elamur Research⌕
☰
home / papers / 2016 / Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter

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

Summary

The paper tackles automatic detection of racist and sexist hate speech on Twitter. The authors note that NLP work in this area has been held back by the lack of a shared definition and by little analysis of which features actually help. They built a corpus of 16,914 tweets by starting from seed slurs and hashtags tied to minorities. They labeled the tweets as sexist, racist or neither, using an 11-item decision list grounded in critical race theory. An outside reviewer checked their labels.

They then tested a logistic regression classifier with 10-fold cross-validation, combining character n-grams with demographic and other proxies: inferred gender, location/time zone, and length statistics. Character n-grams of up to four characters proved a strong baseline, clearly ahead of word n-grams. Of the extra features, only gender gave a gain, and that gain was tiny and not significant. Location and length slightly reduced performance.

The work matters because it supplies a public, realistically imbalanced benchmark (tweet IDs plus labels) and explicit annotation guidelines. It also shows that cheap sub-word features already carry most of the signal.

Key points

  • Corpus: 136,052 tweets were collected over two months, and 16,914 were annotated:
    • 3,383 sexist, from 613 users
    • 1,972 racist, from only 9 users
    • 11,559 neither
  • Agreement: The authors and an outside reviewer reached κ = 0.84. Most disagreements involved the sexism label, and most of the reviewer's changes moved tweets to "neither."
  • Main results (F1):
    • character n-grams: 73.89
      • gender: 73.93 (not significant)
      • gender + location: 73.62 (a significant drop, p = 0.0355)
    • word n-grams: 64.58
  • Gender coverage: Name and pronoun heuristics assigned a gender to only about 52.56% of users. Among those identified, men far outnumbered women in every class.
  • Indicative features: The strongest n-grams were fragments of words like "sexist," "Islam" and "Muslim," plus one gendered slur. Racist tweets mostly repurposed ordinary vocabulary rather than slurs.

Interesting ideas

  • I1. Seed-term collection with innocuous examples: Collecting data from seed terms means the "neither" class contains harmless tweets that use the same loaded words, such as neutral mentions of Muslims. This forces a classifier to learn usage rather than just vocabulary, and it suits any toxicity or propaganda dataset.
  • I2. Explicit annotation checklist: Turning a vague notion like "hateful" into a numbered list of criteria makes labeling more consistent and easier to audit. It is a useful template when building guidelines for subjective categories in new languages.
  • I3. Intuitive bootstrap significance test: Resample the predictions many times and report how often the winning system flips. This gives a simple p-value that is easy to explain, and it works for comparing any two classifiers on the same test set.
  • I4. Feature importance summed across folds: Adding up logistic regression coefficients over all cross-validation folds gives a more stable view of which features matter than inspecting a single model.
  • I5. Checking proxy coverage before adding metadata: Metadata features inferred by proxy, such as gender from names or location from time zones, only help if coverage is high. Measuring how many users the proxy actually labels tells you early whether such features are worth adding.

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

Fields: Natural language processing

Cited by

503 papers of our library cite it (of the 140884 whose bibliographies we read); 40 of them have a page here, 463 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/n16-2013

@inproceedings{waseem2016hateful,
  title = {{Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter}},
  author = {Waseem, Zeerak and Hovy, Dirk},
  booktitle = {Proceedings of the NAACL Student Research Workshop},
  year = {2016},
  doi = {10.18653/v1/n16-2013},
  url = {https://doi.org/10.18653/v1/n16-2013}
}

References

22 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Diana Abbas. 2015. What’s in a location. https://www.youtube.com/watch? v=GNlDO9Lt8J8, October. Talk at Twitter Flight 2015. Seen on Jan 17th 2016. ↗
  2. BBC. 2015. Facebook, google and twitter agree german hate speech deal. http://www.bbc.com/ news/world-europe-35105003. Accessed on 26/11/2016. ↗
  3. Ying Chen, Yilu Zhou, Sencun Zhu, and Heng Xu. 2012a. Detecting offensive language in social media to protect adolescent online safety. In Privacy, Security, Risk and Trust (PASSAT), 2012 International Conference on and 2012 International Conference on Social Computing (SocialCom), pages 71–80. IEEE… find
  4. Yunfei Chen, Lanbo Zhang, Aaron Michelony, and Yi Zhang. 2012b. 4is of social bully filtering: Identity, inference, influence, and intervention. In Proceedings of the 21st ACM International Conference on Information and Knowledge Management, CIKM ’12, pages 2677–2679, New York, NY, USA. ACM. find
  5. Tori DeAngelis. 2009. Unmasking ’racial micro aggressions’. Monitor on Psychology, 40(2):42. find
  6. Lisa Eadicicco. 2014. This female game developer was harassed so severely on twitter she had to leave her home. http://www.businessinsider.com/brianna-wu-harassed-twitter-2014-10?IR=T, Oct. Seen on Jan. 25th, 2016. ↗
  7. Stephan Gouws, Donald Metzler, Congxing Cai, and Eduard Hovy. 2011. Contextual bearing on linguistic variation in social media. In Proceedings of the Workshop on Languages in Social Media, LSM ’11, pages 20–29, Stroudsburg, PA, USA. Association for Computational Linguistics. find
  8. Mark Kantrowitz. 1994. Name corpus: List of male, female, and pet names. http://www.cs.cmu.edu/afs/cs/project/ai-repository/ ai/areas/nlp/corpora/names/0.html. Last accessed on 29th February 2016. ↗
  9. Heather Hensman Kettrey and Whitney Nicole Laster. 2014. Staking territory in the world white web: An exploration of the roles of overt and color-blind racism in maintaining racial boundaries on a popular web site. Social Currents, 1(3):257–274. find
  10. Natasha Lomas. 2015. Facebook, google, twitter commit to hate speech action in germany. http://techcrunch.com/2015/12/16/ germany-fights-hate-speech-on-social-media/, Dec. Seen on 23rd Jan. 2016. ↗
  11. Alexandra Ma. 2015. Global survey finds nordic countries have the most feminists. http://www.huffingtonpost.com/entry/ global-gender-equality-study-yougov_ us_564604cce4b045bf3deeb96d, November. Seen on Jan 19th. ↗
  12. Black Lives Matter. 2012. Guiding principles. http://blacklivesmatter.com/ guiding-principles/. Accessed on 26/11/2016. ↗
  13. Peggy McIntosh, 2003. Understanding prejudice and discrimination., chapter White privilege: Unpacking the invisible knapsack, pages 191–196. McGraw-Hill. find
  14. Geir Moulson. 2016. Zuckerberg in germany: No place for hate speech on facebook. http://abcnews.go.com/Technology/ wireStory/zuckerberg-place-hate-speech-facebook-37217309. Accessed 10/03/2016. ↗
  15. Colin Roberts, Martin Innes, Matthew Williams, Jasmin Tregidga, and David Gadd. 2013. Understanding who commits hate crime and why they do it. find
  16. Sara Sood, Judd Antin, and Elizabeth Churchill. 2012a. Profanity use in online communities. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 1481–1490. ACM. find
  17. Sara Owsley Sood, Judd Antin, and Elizabeth F. Churchill. 2012b. Using crowdsourcing to improve profanity detection. In AAAI Spring Symposium: Wisdom of the Crowd, volume SS-12-06 of AAAI Technical Report. AAAI. find
  18. Stéphan Tulkens, Lisa Hilte, Elise Lodewyckx, Ben Verhoeven, and Walter Daelemans. 2015. Detecting racism in dutch social media posts, 2015/12/18. find
  19. William Warner and Julia Hirschberg. 2012. Detecting hate speech on the world wide web. In Proceedings of the Second Workshop on Language in Social Media, LSM ’12, pages 19–26, Stroudsburg, PA, USA. Association for Computational Linguistics. find our card
  20. Hate Speech Watch. 2014. Hate crimes: Consequences of hate speech. http://www.nohatespeechmovement.org/hate-speech-watch/focus/ consequences-of-hate-speech, June. Seen on on 23rd Jan. 2016. ↗
  21. Julia Carrie Wong. 2016. Mark Zuckerberg tells Facebook staff to stop defacing Black Lives Matter slogans. http://www.theguardian.com/ technology/2016/feb/25/mark-zuckerberg-facebook-defacing-black-lives-matter-signs. Accessed on 10/03/2016. ↗
  22. Matthew Zook. 2012. Mapping racist tweets in response to president obama’s re-election. http://www.floatingsheep.org/2012/11/ mapping-racist-tweets-in-response-to. html. Accessed on 11/03/2016. ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/waseem-2016-hateful-symbols-or-hateful-people/