Elamur Research⌕
☰
home / papers / 2016 / Are You a Racist or Am I Seeing Things? Annotator Influence on Hate Speech Detection on Twitter

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

Summary

Crowdsourcing is the usual way to label hate speech at scale, but it is unclear how annotator expertise affects label quality and the classifiers trained on those labels. Waseem re-annotated 6,909 tweets sampled from the 130k-tweet pool of Waseem and Hovy (2016). Two groups labelled them. The "expert" group was feminist and anti-racism activists. The "amateur" group was unscreened CrowdFlower workers. The label set was racism, sexism, neither, plus a new intersectional "both" label. The resulting data set adds 4,033 tweets to Waseem and Hovy (2016).

Crowd workers flagged hate speech more often than experts did. Their majority-vote labels agreed weakly with the experts, while labels on which all workers agreed matched the experts much more closely. The author trained classifiers on each label set using many feature types, chosen by grid search with 5-fold cross validation. Models trained on expert labels performed better.

The work shows that who annotates hate speech shapes the resulting models. It also suggests a cheaper pipeline in which experts review only the tweets where the crowd disagreed.

Key points

  • Data set: 6,909 tweets labelled by both groups and released on GitHub (zeerakw/hatespeech). Of these, 2,876 overlap with Waseem and Hovy (2016), and agreement with that earlier corpus is very low (mean pairwise κ = 0.14). Most disagreements are tweets the earlier corpus marked as hateful but the new annotators did not.
  • Agreement:
    • Inter-annotator κ among amateurs is 0.57.
    • Amateur majority vote vs. experts gives κ = 0.34.
    • Amateur full-agreement labels vs. experts give κ = 0.70.
  • Label rates: Amateur majority vote marks 5.80% of tweets as racism and 19.00% as sexism. Experts mark 1.41% and 13.08%.
  • Classifier scores: The best single features reach about 91.2–91.5 F1 on expert labels, against about 86.3–86.5 on amateur labels.
    • Brown Clusters and length features are the only ones that predict minority classes under both label sets.
    • Author Historical Salient Terms (AHST) is the weakest feature, scoring 55.40 F1 on expert labels.
    • Gender features help only with amateur labels.
  • Results on the Waseem and Hovy (2016) data:
    • Binary task: on par with the original (70.05 vs. 69.94 F1).
    • Multi-class task: much worse (53.43 vs. 73.89).
    • Most errors are false positives.

Interesting ideas

  • I1. Compare crowd and expert labels on the same items. Labelling an identical set with both groups isolates the effect of expertise on the labels and on the trained models. This setup could be reused to audit any subjective annotation task, such as toxicity, propaganda or sentiment in Ukrainian data.
  • I2. Treat unanimous crowd votes as high-quality labels. Items with full crowd agreement can be accepted directly, and scarce experts can focus on the contested ones. This cuts expert workload and also reduces their exposure to disturbing content.
  • I3. Add an intersectional label. Overlapping categories (racism plus sexism) get their own class instead of being forced into one. This is a useful design choice wherever phenomena combine into something qualitatively different from their parts.
  • I4. Do not trust unweighted F1 on imbalanced data. Some features scored high while never predicting a minority class. The paper therefore recommends a weighted F1 that penalises minority-class errors, along with keeping realistic class proportions rather than simply inflating the rare classes.
  • I5. Test author-history features. AHST summarises a user's past tweets as TF-IDF top terms. Its poor result suggests hate speech is situational rather than tied to the author. Running this kind of check is a cheap way to see whether a phenomenon belongs to the user or to the individual message.

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

191 papers of our library cite it (of the 140884 whose bibliographies we read); 22 of them have a page here, 169 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/w16-5618

@inproceedings{waseem2016are,
  title = {{Are You a Racist or Am I Seeing Things? Annotator Influence on Hate Speech Detection on Twitter}},
  author = {Waseem, Zeerak},
  booktitle = {Proceedings of the First Workshop on NLP and Computational Social Science},
  year = {2016},
  doi = {10.18653/v1/w16-5618},
  url = {https://doi.org/10.18653/v1/w16-5618}
}

References

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

  1. Results Running our system on the Waseem and Hovy (2016) data set, we find that our best performing system does not substantially outperform on the binary classification task Waseem and Hovy (2016) (F1 ours : 70.05, F1 W H : 69.94). We find that our system performs significantly worse than Waseem an… find
  2. Waseem and Hovy (2016) may suffer from personal bias, as the only the authors annotated, and only the annotations positive for hate speech were reviewed by one other person. It is our contention that hate speech corpora should reflect real life, in that hate speech is a rare occurrence comparatively… find
  3. While Sood et al. (2012) incorporate edit distances to find variants of slurs, they are not able to find terms that do not occur in these lists. find
  4. Nobata et al. (2016) address this, by using comprehensive lists of slurs obtained from Hatebase 4 . Waseem and Hovy (2016) and Ross et al. (2016) focus on building corpora which they annotate for containing hate speech. Our work closely resembles Waseem and Hovy (2016), as they also run classificati… find
  5. Waseem and Hovy (2016) obtain an F1-score of 73.91 on their data set, using character n-grams and gender information. find
  6. Nobata et al. (2016) employ a wide array of features for abusive language detection, including but not limited to POS tags, the number of blacklisted words in a document, n-gram features including token and character n-grams and length features. The primary challenge this paper presents, is the need… find
  7. Geneva, Switzerland. International World Wide Web Conferences Steering Committee. Björn Ross, Michael Rist, Guillermo Carbonell, Benjamin Cabrera, Nils Kurowsky, and Michael Wojatzki. 2016. Measuring the reliability of hate speech annotations: The case of the european refugee crisis. Future Work We… find
  8. Association David Bamman and Noah Smith. 2015. Contextualized for Computational Linguistics. sarcasm detection on twitter. Sara Owsley Sood, Judd Antin, and Elizabeth F. Peter F. Brown, Peter V. deSouza, Robert L. Mercer, Vin-Churchill. 2012. Using crowdsourcing to improve cent J. Della Pietra, and… find
  9. AAAI. Ying Chen, Yilu Zhou, Sencun Zhu, and Heng Xu. 2012. Derald Wing Sue, Christina M Capodilupo, Gina C Detecting offensive language in social media to pro-Torino, Jennifer M Bucceri, Aisha Holder, Kevin L tect adolescent online safety. In Privacy, Security, Risk Nadal, and Marta Esquilin. 2007.… find
  10. IEEE, September. William Warner and Julia Hirschberg. 2012. find
  11. Detecting Adrian Chen. 2012. Inside facebook’s out-hate speech on the world wide web. In Proceedings of sourced anti-porn and gore brigade, where the Second Workshop on Language in Social Media, ’camel toes’ are more offensive than ’crushed LSM ’12, pages 19–26, Stroudsburg, PA, USA. Asso-heads’. ht… find
  12. Kimberle Crenshaw. 1989. Demarginalizing the in-Association for Computational Linguistics. tersection of race and sex: A black feminist critique of antidiscrimination doctrine, feminist eory and antiracist politics. University of Chicago Legal Forum, 1989(1). find
  13. Kevin Gimpel, Nathan Schneider, Brendan O’Connor, Dipanjan Das, Daniel Mills, Jacob Eisenstein, Michael Heilman, Dani Yogatama, Jeffrey Flanigan, and Noah A. Smith. 2011. Part-of-speech tagging for twitter: Annotation, features, and experiments. In Proceedings of the 49th Annual Meeting of the Assoc… find our card
  14. Chikashi Nobata, Joel Tetreault, Achint Thomas, Yashar Mehdad, and Yi Chang. 2016. Abusive language detection in online user content. In Proceedings of the 25th International Conference on World Wide Web, WWW ’16, pages 145–153, Republic and Canton of find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/waseem-2016-are-you-racist-or-am/