Elamur Research⌕
☰
home / papers / 2022 / Emojis as Anchors to Detect Arabic Offensive Language and Hate Speech

Emojis as Anchors to Detect Arabic Offensive Language and Hate Speech

Summary

Abusive content is scarce on Twitter. Under 2% of random Arabic tweets are offensive, so building labelled corpora by annotating random samples is costly. Keyword lists struggle with dialect spellings, rich morphology and context-dependent insults. The authors propose using emojis as a language-independent signal for collection. They derive offensive-leaning emojis automatically from existing datasets, gather tweets containing them over 18 months, and have crowd workers annotate the result. Labels cover offensiveness, vulgarity, violence and fine-grained hate speech (gender, race, ideology, social class, religion, disability). The resulting corpus has 12,698 tweets, of which 35% are offensive and 11% hate speech. That is roughly double the yield of the earlier pattern-based method. They benchmark transformer classifiers (best: QARiB) and test them on OffensEval 2020 and the multi-platform MPOLD set, with competitive results. Ablations show the models do not depend on the emojis themselves. They also analyse emoji use over time, compare cultural differences with English, and study hate-speech targets and errors with LIME. This makes a large, rich Arabic resource available and offers a cheap collection recipe that carries over to other languages.

Key points

  • Dataset: 12,698 Arabic tweets (MSA and dialects), labelled for OFF, Vulgar, Violence and six hate-speech types. The authors call it the largest such Arabic set, and it is to be released as text without tweet ids.
  • Collection yield compared with earlier methods:
    • random tweets: under 2% offensive;
    • the "yA .. yA" vocative pattern: 19% offensive / 5% HS;
    • emoji anchors: 35% offensive / 11% HS.
    • The method was also shown to work for Bengali.
  • Cross-platform test on MPOLD (4,000 news comments): QARiB fine-tuned on this data scored F1 68.2 on YouTube and 66.95 on Facebook. Previously reported cross-domain figures were 0.60 and 0.62.
  • Emoji ablation:
    • with emojis stripped from all data, F1 was 85.03 on SemEval and 83.05 on the in-house test set;
    • keeping emojis in training but testing only on emoji-free SemEval tweets gave F1 83.94.
  • Annotation: over 190 Appen workers, filtered by at least 80% accuracy on 200 gold test questions. Acknowledged biases include an annotator pool skewed toward Egypt and toward men, while Arabic Twitter is dominated by Gulf users.

Interesting ideas

  • I1. Use a non-lexical signal (here emojis) to harvest rare phenomena before annotation, avoiding brittle keyword lists that miss creative spellings and inflected forms. — For Mova (learn): Mova could gather candidates for rare phenomena such as sarcasm or insults through surface cues like emojis or punctuation runs, then send only that enriched pool to the LLM teacher tiers.
  • I2. After collecting data with an anchor, remove the anchor and retrain or retest to check the model has not simply learned the collection cue. — For Mova (learn): any silver set selected by a cue should get a matching negative control with that cue masked, to confirm the learned rules generalise beyond the selection trigger.
  • I3. The same word can flip from insult to neutral depending on structure, e.g. calling someone a dog versus owning a dog, which defeats plain string matching. — For Mova (frames): insult detection fits naturally as a speech-act rule over the tree, firing on copular predication of an animal noun of a person but not on possession frames.
  • I4. Validate generalisation on external sets collected by a completely different method and on other platforms, not only on a held-out split. — For Mova (answer): reader evaluations should include test sets drawn from other sources (another fable collection or textbook) so that scores are not tied to how the training material was chosen.
  • I5. Select gold quality-control questions in proportion to the expected class distribution, so annotator screening reflects the real label mix. — For Mova (learn): when scoring LLM teacher tiers or witness parsers against gold items, sample those items in proportion to construction frequencies so rare-case accuracy is neither over- nor under-weighted.

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: Linguistics & languages

Cited by

40 papers of our library cite it (of the 140884 whose bibliographies we read); 5 of them have a page here, 35 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2201.06723 · arXiv 2201.06723

@misc{mubarak2022emojis,
  title = {{Emojis as Anchors to Detect Arabic Offensive Language and Hate Speech}},
  author = {Mubarak, Hamdy and Hassan, Sabit and Chowdhury, Shammur Absar},
  year = {2022},
  doi = {10.48550/arxiv.2201.06723},
  eprint = {2201.06723},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2201.06723}
}

References

46 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abdelali, A., Hassan, S., Mubarak, H., Darwish, K., and Samih, Y. 2021a. Pre-training BERT on arabic tweets: Practical considerations. CoRR, abs/2102.10684. ↗ our card
  2. Abdelali, A., Mubarak, H., Samih, Y., Hassan, S., and Darwish, K. 2021b. QADI: Arabic dialect identification in the wild. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pp. 1–10, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. find our card
  3. Alami, H., Ouatik El Alaoui, S., Benlahbib, A., and En-nahnahi, N. 2020. LISAC FSDM-USMBA team at SemEval-2020 task 12: Overcoming AraBERT’s pretrain-finetune discrepancy for Arabic offensive language identification. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pp. 2080–2085, Barcelona (online). International Committee for Computational Linguistics. find
  4. Albadi, N., Kurdi, M., and Mishra, S. 2018. Are they our brothers? analysis and detection of religious hate speech in the arabic twittersphere. In 2018 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pp. 69–76. IEEE. find
  5. Alhazbi, S. 2020. Behavior-based machine learning approaches to identify state-sponsored trolls on twitter. IEEE Access, 8:195132–195141. find
  6. Alshaalan, R. and Al-Khalifa, H. 2020. Hate speech detection in saudi twittersphere: A deep learning approach. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pp. 12–23, Barcelona, Spain (Online). Association for Computational Linguistics. find
  7. Alshehri, A., El Moatez Billah Nagoudi, H. A., and Abdul-Mageed, M. 2018. Think before your click: Data and models for adult content in arabic twitter. In TA-COS 2018: 2nd Workshop on Text Analytics for Cybersecurity and Online Safety, 15. find
  8. Antoun, W., Baly, F., and Hajj, H. 2020. Arabert: Transformer-based model for arabic language understanding. In Proceedings of The 4th Workshop on Open-Source Arabic Corpora and Processing Tools, pp. 9–15. find our card
  9. Belcastro, L., Cantini, R., Marozzo, F., Talia, D., and Trunfio, P. 2020. Learning political polarization on social media using neural networks. IEEE Access, 8:47177–47187. find
  10. Cheng, H., Xing, X., Liu, X., and Lv, Q. 2014. Isc: An iterative social based classifier for adult account detection on twitter. IEEE Transactions on Knowledge and Data Engineering, 27(4):1045–1056. find
  11. Chowdhury, S. A., Abdelali, A., Darwish, K., Soon-Gyo, J., Salminen, J., and Jansen, B. J. 2020a. Improving arabic text categorization using transformer training diversification. In Proceedings of the Fifth Arabic Natural Language Processing Workshop, pp. 226–236. find
  12. Chowdhury, S. A., Mubarak, H., Abdelali, A., Jung, S.-g., Jansen, B. J., and Salminen, J. 2020b. A multi-platform arabic news comment dataset for offensive language detection. In Proceedings of The 12th Language Resources and Evaluation Conference, pp. 6203–6212. find our card
  13. Chung, Y.-L., Kuzmenko, E., Tekiroglu, S. S., and Guerini, M. 2019. Conan–counter narratives through nichesourcing: a multilingual dataset of responses to fight online hate speech. arXiv preprint arXiv:1910.03270. ↗ our card
  14. Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., and Stoyanov, V. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116. ↗
  15. Conover, M., Ratkiewicz, J., Francisco, M. R., Gonçalves, B., Menczer, F., and Flammini, A. 2011. Political polarization on twitter. ICWSM, 133:89–96. find
  16. Darwish, K., Alexandrov, D., Nakov, P., and Mejova, Y. 2017. Seminar users in the arabic twitter sphere. In Ciampaglia, G. L., Mashhadi, A., and Yasseri, T., editors, Social Informatics, pp. 91–108, Cham. Springer International Publishing. find
  17. Davidson, T., Warmsley, D., Macy, M., and Weber, I. 2017. Automated hate speech detection and the problem of offensive language. In Proceedings of the International AAAI Conference on Web and Social Media, volume 11. find our card
  18. Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. 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), pp. 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. find
  19. Dimitrov, D., Ali, B. B., Shaar, S., Alam, F., Silvestri, F., Firooz, H., Nakov, P., and Da San Martino, G. 2021a. Detecting propaganda techniques in memes. In Proceedings of the Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (ACL-IJCNLP’21). find
  20. Dimitrov, D., Ali, B. B., Shaar, S., Alam, F., Silvestri, F., Firooz, H., Nakov, P., and Da San Martino, G. 2021b. Task 6 at semeval-2021: Detection of persuasion techniques in texts and images. In Proceedings of the 15th International Workshop on Semantic Evaluation, SemEval, volume 21. find
  21. Donato, G. and Paggio, P. 2017. Investigating redundancy in emoji use: Study on a twitter based corpus. In Proceedings of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pp. 118–126. find
  22. Dürscheid, C. and Siever, C. M. 2017. Beyond the alphabet–communcataion of emojis. Kurzfassung eines (auf Deutsch) zur Publikation eingereichten Manuskripts. find
  23. Gülaçtı, F. 2010. The effect of perceived social support on subjective well-being. Procedia-Social and Behavioral Sciences, 2(2):3844–3849. find
  24. Hassan, S., Mubarak, H., Abdelali, A., and Darwish, K. 2021a. ASAD: Arabic social media analytics and unDerstanding. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pp. 113–118, Online. Association for Computational Linguistics. find
  25. Hassan, S., Samih, Y., Mubarak, H., and Abdelali, A. 2020. ALT at SemEval-2020 task 12: Arabic and English offensive language identification in social media. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pp. 1891–1897, Barcelona (online). International Committee for Computational Linguistics. find
  26. Hassan, S., Shaar, S., and Darwish, K. 2021b. Cross-lingual emotion detection. CoRR, abs/2106.06017. ↗
  27. Husain, F. 2020. Osact4 shared task on offensive language detection: Intensive preprocessing based approach. OSACT, 4. find
  28. Intapong, P., Charoenpit, S., Achalakul, T., and Ohkura, M. 2017. Assessing symptoms of excessive sns usage based on user behavior and emotion: analysis of data obtained by sns apis. In 9th International Conference on Social Computing and Social Media, SCSM 2017 held as part of the 19th International Conference on Human-Computer Interaction, HCI International 2017, pp. 71–83. Springer Verlag. find
  29. Kiela, D., Firooz, H., Mohan, A., Goswami, V., Singh, A., Ringshia, P., and Testuggine, D. 2020. The hateful memes challenge: Detecting hate speech in multimodal memes. arXiv preprint arXiv:2005.04790. ↗ our card
  30. Landis, J. R. and Koch, G. G. 1977. The measurement of observer agreement for categorical data. Biometrics, 33(1):159–174. find
  31. Mei, Q. 2019. Decoding the new world language: Analyzing the popularity, roles, and utility of emojis. In Companion Proceedings of The 2019 World Wide Web Conference, pp. 417–418. find
  32. Mubarak, H., Abdelali, A., Hassan, S., and Darwish, K. 2020a. Spam detection on arabic twitter. In Social Informatics, pp. 237–251, Cham. Springer International Publishing. find
  33. Mubarak, H. and Darwish, K. 2014. Using twitter to collect a multi-dialectal corpus of arabic. In Proceedings of the EMNLP 2014 Workshop on Arabic Natural Language Processing (ANLP), pp. 1–7. find
  34. Mubarak, H. and Darwish, K. 2016. Demographic surveys of arab annotators on crowdflower. In Weaving Relations of Trust in Crowd Work: Transparency and Reputation across Platforms. find
  35. 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, pp. 52–56. find our card
  36. Mubarak, H., Darwish, K., Magdy, W., Elsayed, T., and Al-Khalifa, H. 2020b. Overview of osact4 arabic offensive language detection shared task. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, with a Shared Task on Offensive Language Detection. find
  37. Mubarak, H., Hassan, S., and Abdelali, A. 2021. Adult content detection on Arabic Twitter: Analysis and experiments. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pp. 136–144, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. find
  38. Mubarak, H., Rashed, A., Darwish, K., Samih, Y., and Abdelali, A. 2020c. Arabic offensive language on twitter: Analysis and experiments. arXiv preprint arXiv:2004.02192. ↗ our card
  39. Nakov, P., Nayak, V., Dent, K., Bhatawdekar, A., Sarwar, S. M., Hardalov, M., Dinkov, Y., Zlatkova, D., Bouchard, G., and Augenstein, I. 2021. Detecting abusive language on online platforms: A critical analysis. arXiv preprint arXiv:2103.00153. ↗
  40. Ousidhoum, N., Lin, Z., Zhang, H., Song, Y., and Yeung, D.-Y. 2019. Multilingual and multi-aspect hate speech analysis. arXiv preprint arXiv:1908.11049. ↗ our card
  41. Platt, J. 1998. Sequential minimal optimization: A fast algorithm for training support vector machines. find
  42. Polignano, M., Basile, P., de Gemmis, M., Semeraro, G., and Basile, V. 2019. Alberto: Italian bert language understanding model for nlp challenging tasks based on tweets. In CLiC-it. find our card
  43. Ribeiro, M. T., Singh, S., and Guestrin, C. 2016. "why should i trust you?": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, 1135–1144, New York, NY, USA. Association for Computing Machinery. find our card
  44. Salminen, J., Hopf, M., Chowdhury, S. A., Jung, S.-g., Almerekhi, H., and Jansen, B. J. 2020. Developing an online hate classifier for multiple social media platforms. Human-centric Computing and Information Sciences, 10(1):1–34. find
  45. Waldron, J. 2012. The harm in hate speech. Harvard University Press. find
  46. Wiegand, M. and Ruppenhofer, J. 2021. Exploiting emojis for abusive language detection. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pp. 369–380. Zampieri, M., Nakov, P., Rosenthal, S., Atanasova, P., Karadzhov, G., Mubarak, H., Derczynski, L., Pitenis, Z., and Çöltekin, c. 2020. SemEval-2020 Task 12: Multilingual Offensive Language Identification in Social Media (OffensEval 2020). In Proceedings of SemEval. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mubarak-2022-emojis-anchors-detect-arabic-offensive/