Elamur Research⌕
☰
home / papers / 2026 / PrimeLine@DravidianLangTech 2026: Abusive Tamil Comment Detection Using MuRIL

PrimeLine@DravidianLangTech 2026: Abusive Tamil Comment Detection Using MuRIL

Paper · 2026 · Rithikaa V, S.Sumathi, Nithya Varshini C N R et al.

Summary

The paper addresses binary detection of abusive comments in Tamil social media. Keyword lists and classical classifiers struggle here because Tamil is agglutinative and users mix Tamil script, romanised Tamil and English within one comment. For the DravidianLangTech@ACL 2026 shared task, the authors fine-tune MuRIL, a BERT-style encoder developed by Google Research whose pre-training covered 17 Indian languages, written both in their own scripts and in transliterated form. A single linear layer on top of the [CLS] vector makes the prediction. The organisers' data has 3,652 labelled comments, split 90/10 with stratification, plus 912 unlabelled test comments. Preprocessing is minimal: only Unicode normalisation, with no transliteration or language identification, so that code-switching signals are preserved. On the validation split, MuRIL reaches macro-F1 0.83, ahead of XLM-R (0.79) and mBERT (0.77) trained under identical settings. The authors attribute the gain to transliterated text in MuRIL's pre-training. The text reports no test-set score. Most errors come from implicit abuse and heavily mixed comments. The work provides a simple, reproducible baseline with released code for low-resource, code-mixed abuse detection.

Key points

  • Data: 3,652 training comments, about 64% Non-Abusive (2,338) and 36% Abusive (1,314). These were split into 3,286 train and 366 validation examples, with a 912-comment blind test set. Mean comment length is about 18 tokens.
  • Setup: MuRIL-base (12 layers, 768 dims) with WordPiece tokenisation, max length 128, and the Hugging Face Trainer. Training used AdamW, learning rate 2e-5, batch size 16, weight decay 0.01, and 4 epochs. Training took about 25 minutes on one GPU.
  • Validation macro-F1: MuRIL 0.83 (P 0.84, R 0.83), XLM-R 0.79, mBERT 0.77. Only validation results appear in the paper, not official test scores.
  • Per-class results for MuRIL:
    • Non-Abusive: precision 0.87, recall 0.91.
    • Abusive: precision 0.81, recall 0.75, so roughly a quarter of abusive comments are missed.
  • Limitations: Errors come mainly from sarcasm or euphemism and from dense code-mixing, and false negatives dominate. Class imbalance was left unaddressed. Suggested next steps include focal or weighted loss and back-translation augmentation. Code: github.com/Nithya-svg/primeline_tamil_abusive.

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, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

Cite

Cite as Rithikaa V, S.Sumathi, Nithya Varshini C N R and Sanjay Krishnan K. 2026. PrimeLine@DravidianLangTech 2026: Abusive Tamil Comment Detection Using MuRIL. Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.dravidianlangtech-1.51

@inproceedings{v2026primeline,
  title = {{PrimeLine@DravidianLangTech 2026: Abusive Tamil Comment Detection Using MuRIL}},
  author = {Rithikaa V, S.Sumathi, Nithya Varshini C N R and Sanjay Krishnan K},
  booktitle = {Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages},
  year = {2026},
  doi = {10.18653/v1/2026.dravidianlangtech-1.51},
  url = {https://doi.org/10.18653/v1/2026.dravidianlangtech-1.51}
}

References

12 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. S. Anbukkarasi and S. Varadhaganapathy. Deep learning-based hate speech detection in code-mixed Tamil text. IETE Journal of Research, 69(11):7893– 7898, 2023. URL https://doi.org/10.1080/ 03772063.2021.1992853. find
  2. Bharathi Raja Chakravarthi, Ruba Priyadharshini, Subalalitha Chinnaudayar Navaneethakrishnan, Navaneethan Rajasekaran, Sajeetha Thavareesan, and Dhivya Chinnappa. Findings of the shared task on offensive language identification in Tamil, Malayalam, and Kannada. In Proceedings of the First Workshop o…
  3. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Comp…
  4. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 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, pages…
  5. Simran Khanuja, Diksha Bansal, Sarvesh Mehtani, Savya Khosla, Atreyee Dey, Balaji Gopalan, Dilip Kumar Margam, Pooja Aggarwal, Rajiv Teja Nagipogu, Shachi Dave, Shruti Gupta, Subhash Chandra Bose Gali, Vish Subramanian, and Partha Talukdar. MuRIL: Multilingual representations for Indian languages. a…
  6. R. Krishna et al. Abusive comment detection in Tamil code-mixed data by adjusting data distribution. In Proceedings of the International Conference on Computational Intelligence and Data Engineering (ICCIDE 2024), 2024. URL https://doi.org/10.1007/ 978-981-97-6714-4. find
  7. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692, 2019. URL https://arxiv.org/abs/1907.11692. R, Ratnavel Rajalakshmi, Shun… ↗
  8. Chikashi Nobata, Joel Tetreault, Achint Thomas, Yashar Mehdad, and Yi Chang. Abusive language detection in online user content. In Proceedings of the 25th International Conference on World Wide Web, pages 145– 153, 2016. URL https://doi.org/10.1145/ 2872427.2883062. find
  9. Tharindu Ranasinghe and Marcos Zampieri. Multilingual offensive language identification with cross-lingual embeddings. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5838–5847, 2021. URL https://aclanthology.org/2021.emnlp-main.470. ↗
  10. Nishanth S, Shruthi Rengarajan, S Anantha-sivan, Burugu Rahul, and Sachin Kumar S. ANSR@DravidianLangTech 2025: Detection of abusive Tamil and Malayalam text targeting women on social media using RoBERTa and XGBoost. In Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies f… find
  11. Siva Sai and Yashvardhan Sharma. Towards offensive language identification for Dravidian languages. In Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Languages, pages 18– 27, 2021. URL https://aclanthology.org/ 2021.dravidianlangtech-1.3. ↗
  12. Bhuvaneswari Sivagnanam, Kathiravan Pannerselvam, Jananayagan V, Charmathi Rajkumar, Ramesh Kannan find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/v-2026-primeline-dravidianlangtech-2026-abusive-tamil/