Elamur Research⌕
☰
home / papers / 2019 / ARHNet - Leveraging Community Interaction for Detection of Religious Hate Speech in Arabic

ARHNet - Leveraging Community Interaction for Detection of Religious Hate Speech in Arabic

Paper · 2019 · Arijit Ghosh Chowdhury, Aniket Didolkar, Ramit Sawhney et al.

Summary

This paper tackles automatic detection of religious hate speech in Arabic tweets. Earlier Arabic systems relied only on what the text says. The authors add information about who wrote a tweet and how that author is connected to others. They use the Arabic religious hate dataset of Albadi et al. (2018) and could still download 3950 of its tweets. From the authors of these tweets they build two social graphs, one from follower links and one from retweet links. Node2Vec turns each author's place in a graph into a 64-dimensional vector. That vector is appended to a sentence representation from a neural text encoder (LSTM, GRU, CNN hybrids, or self-attention variants). The text encoders use 600-dimensional word vectors.

The resulting family of models, ARHNet, slightly beats the AraHate baselines on all five reported metrics. The best F1 rises from 0.77 to 0.78. The work suggests that homophily, meaning that people prone to hate tend to cluster together, is a usable signal alongside lexical cues. This matters for low-resource languages with rich morphology.

Key points

  • Data: The released dataset is a list of 5570 tweet IDs with labels; only 3950 could be fetched with Twarc. Of these, 1,685 (42.6%) are labelled Hate, close to the original 45.1%. Annotator agreement was 81%.
  • Preprocessing: Arabic-specific normalisation of Hamza/alef, alef maksura and teh marbuta, plus removal of diacritics and tatweel. Links, mentions and numbers become placeholder tokens. Text is stemmed with the ISRI stemmer from NLTK. A 356-word stopword list is applied, with negation words deliberately excluded.
  • Graph features: Follower and retweet graphs are undirected and unweighted. Node2Vec runs with default p = q = 1, giving 64-dimensional embeddings. Authors with no connections get a shared zero vector.
  • Text side: Self-trained Gensim vectors are concatenated with AraVec Twitter-CBOW 300-dimensional vectors. Training uses dropout 0.5, 240 hidden features, batch size 32, Adam and a sigmoid output.
  • Results: Baselines are AraHate-LR, AraHate-SVM (character 1–4-grams) and AraHate-GRU.
    • The best F1 of 0.78 comes from an LSTM+CNN model with graph features.
    • A GRU+Node2Vec variant has the top precision, and Bi-GRU-CNN-Node2Vec has the top accuracy.
    • AUROC peaks around 0.86.
    • The gains over the baselines are small.

Interesting ideas

  • I1. Each text architecture is trained twice, once with the extra author embedding appended and once without, so the effect of the feature is isolated per model rather than claimed overall.
  • I2. Information that is missing gets an explicit shared "no community" zero vector instead of dropping the example, so the model sees all inputs uniformly.
  • I3. After recovering only part of a dataset, the authors check that the label proportions still match the original release before trusting comparisons.
  • I4. The stopword list deliberately keeps negation words, because removing them would flip or erase the stance of a sentence.
  • I5. URLs, mentions and numbers are rewritten to fixed placeholder tokens, which shrinks the vocabulary while preserving that something of that type was present.

For Mova

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

  • I1 (learn): Evaluate every new knowledge source (e.g. a WordNet feature for the parser) as a matched with/without pair on the same architecture, alongside negative controls.
  • I2 (world): Entities that are absent from Wikidata/WordNet could carry an explicit "unknown" marker, so downstream reasoning can tell missing knowledge apart from a negative fact.
  • I3 (learn): When silver or re-fetched treebank data shrinks, routinely compare label and relation distributions with the source to catch selection drift.
  • I4 (frames): Any token filtering before frame building must protect negators and polarity items so speech acts and implicature stay correct.
  • I5 (tokens): The tokenizer can map such spans to typed placeholders with a recoverable original, keeping the tree clean and the step explainable.

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

🎓Indraprastha Institute of Information Technology DelhiDelhi · India9 works · 1,294 citations🎓Netaji Subhas University of TechnologyNew Delhi · India3 works · 94 citations🎓Manipal Academy of Higher EducationManipal · India5 works · 57 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 3Research subjects and languages the work is about.Sentiment & emotion · 972 worksSemitic languages · 2,908 worksBias, fairness & safety · 517 works

Cite

Cite as Arijit Ghosh Chowdhury, Aniket Didolkar, Ramit Sawhney and Rajiv Ratn Shah. 2019. ARHNet - Leveraging Community Interaction for Detection of Religious Hate Speech in Arabic. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/p19-2038

@inproceedings{chowdhury2019arhnet,
  title = {{ARHNet - Leveraging Community Interaction for Detection of Religious Hate Speech in Arabic}},
  author = {Arijit Ghosh Chowdhury, Aniket Didolkar, Ramit Sawhney and Rajiv Ratn Shah},
  booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop},
  year = {2019},
  doi = {10.18653/v1/p19-2038},
  url = {https://doi.org/10.18653/v1/p19-2038}
}

References

26 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. N. Albadi, M. Kurdi, and S. Mishra. 2018. Are they our brothers? analysis and detection of religious hate speech in the arabic twittersphere. In 2018 IEEE/ACM International Conference on Advances radimrehurek.com/gensim/models/word2vec.html Debanjan Mahata, John R Talburt, and Vivek Kumar Singh. 2015. From chirps to whistles: discovering event-specific informative content from twitter. In Proceedings of the ACM web science conference, page 17. ACM. in Social Networks Analysis and Mining (ASONAM), pages 69–76. find
  2. Silvio Amir, Byron C. Wallace, Hao Lyu, Paula Carvalho, and Mário J. Silva. 2016. Modelling context with user embeddings for sarcasm detection in social media. CoRR, abs/1607.00976. ↗
  3. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Distributed representations of words and phrases and their compositionality. CoRR, abs/1310.4546. ↗
  4. Pinkesh Badjatiya, Shashank Gupta, Manish Gupta, and Vasudeva Varma. 2017. Deep learning for hate speech detection in tweets. CoRR, abs/1706.00188. ↗
  5. Pushkar Mishra, Marco Del Tredici, Helen Yannakoudakis, and Ekaterina Shutova. 2018. Author profiling for abuse detection. In COLING. find
  6. Matteo Bonotti. 2017. Religion, hate speech and non-domination. Ethnicities, 17:259–274. ↗
  7. Rajiv Ratn Shah, Anupam Samanta, Deepak Gupta, Yi Yu, Suhua Tang, and Roger Zimmermann. 2016a. Prompt: Personalized user tag recommendation for social media photos leveraging personal and social contexts. In 2016 IEEE International Symposium on Multimedia (ISM), pages 486–492. IEEE. find
  8. Wei-Fan Chen and Lun-Wei Ku. 2016. UTCNN: a deep learning model of stance classification on social media text. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1635–1645, Osaka, Japan. find
  9. The COLING 2016 Organizing Committee. find
  10. Thomas Davidson, Dana Warmsley, Michael W. Macy, and Ingmar Weber. 2017. Automated hate speech detection and the problem of offensive language. CoRR, abs/1703.04009. ↗
  11. Rajiv Ratn Shah, Yi Yu, Suhua Tang, Shin’ichi Satoh, Akshay Verma, and Roger Zimmermann. 2016b. Concept-level multimodal ranking of flickr photo tags via recall based weighting. In Proceedings of the 2016 ACM Workshop on Multimedia COM-MONS, pages 19–26. ACM. find
  12. Nemanja Djuric, Jing Zhou, Robin Morris, Mihajlo Grbovic, Vladan Radosavljevic, and Narayan Bhamidipati. 2015. Hate speech detection with comment embeddings. In Proceedings of the 24th International Conference on World Wide Web, WWW ’15 Companion, pages 29–30, New York, NY, USA. ACM. find
  13. Rajiv Ratn Shah, Yi Yu, Akshay Verma, Suhua Tang, Anwar Dilawar Shaikh, and Roger Zimmermann. 2016c. Leveraging multimodal information for event summarization and concept-level sentiment analysis. Knowledge-Based Systems, 108:102–109. ↗
  14. Abu Bakr Soliman, Kareem Eissa, and Samhaa R. El-Beltagy. 2017. Aravec: A set of arabic word embedding models for use in arabic nlp. Procedia Computer Science, 117:256 – 265. Arabic Computational Linguistics. find
  15. Teona Gelashvili. 2018. Hate speech on social media: Implications of private regulation and governance gaps. Student Paper. find
  16. Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable feature learning for networks. CoRR, abs/1607.00653. ↗
  17. C. Tlmcel and F. Leon. 2017. Predicting political opinions in social networks with user embeddings. In 2017 13th IEEE International Conference on Intelligent Computer Communication and Processing (ICCP), pages 213–219. find
  18. Dirk Hovy. 2015. Demographic factors improve classification performance. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 752–762, Beijing, China. Association for Computational Linguistics.
  19. Zeerak Waseem and Dirk Hovy. 2016. Hateful symbols or hateful people? predictive features for hate speech detection on twitter. In Proceedings of the NAACL Student Research Workshop, pages 88–93, San Diego, California. Association for Computational Linguistics.
  20. Pengfei Liu, Xipeng Qiu, and Xuanjing Huang. 2016. Recurrent neural network for text classification with multi-task learning. CoRR, abs/1605.05101. ↗
  21. Yi Yang, Ming-Wei Chang, and Jacob Eisenstein. 2016. Toward socially-infused information extraction: Embedding authors, mentions, and entities. CoRR, abs/1609.08084. ↗
  22. Walid Magdy, Kareem Darwish, and Ingmar Weber. 2015. #failedrevolutions: Using twitter to study the antecedents of ISIS support. CoRR, abs/1503.02401. ↗
  23. Yi Yang and Jacob Eisenstein. 2015. Putting things in context: Community-specific embedding projections for sentiment analysis. CoRR, abs/1511.06052. ↗
  24. Debanjan Mahata, Jasper Friedrichs, Rajiv Ratn Shah, and Jing Jiang. 2018. Detecting personal intake of medicine from twitter. IEEE Intelligent Systems, 33(4):87–95. ↗
  25. Dawei Yin and Brian D. Davison. 2009. Detection of harassment on web 2.0. find
  26. Chunting Zhou, Chonglin Sun, Zhiyuan Liu, and Francis C. M. Lau. 2015. A C-LSTM neural network for text classification. CoRR, abs/1511.08630. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/chowdhury-2019-arhnet-leveraging-community-interaction-detection/