Elamur Research⌕
☰
home / papers / 2026 / DLRG@DravidianLangTech 2026: Explainable Transformer-Based Detection of Abusive Tamil Text Targeting Women on Social Media

DLRG@DravidianLangTech 2026: Explainable Transformer-Based Detection of Abusive Tamil Text Targeting Women on Social Media

Paper · 2026 · Mirudhula Sankar and Ratnavel Rajalakshmi

Summary

The paper addresses automatic detection of abusive Tamil social media comments aimed at women. This is difficult because Tamil is low-resource and online posts are informal, with phonetic spellings, slang, dialect variation and implicit abuse such as sarcasm. Using the binary abusive/non-abusive dataset from the DravidianLangTech 2026 shared task, the authors fine-tune MuRIL, a multilingual transformer pretrained on Indian languages. They apply almost no preprocessing and classify from the [CLS] representation. They also add a post-hoc explainability step: each word is deleted in turn, and the change in log-odds of the prediction serves as that word's importance. The best validation results were about 0.814 accuracy and 0.813 macro-F1. These beat a TF-IDF + SVM baseline, a plain MuRIL baseline and a TF-IDF + MuRIL hybrid. All reported figures are validation scores, and the paper does not clearly explain why the proposed MuRIL setup outperforms its MuRIL baseline. The work matters because it shows a single, interpretable transformer can be competitive for moderating abusive content in Tamil.

Key points

  • Data: DravidianLangTech 2026 shared-task Tamil comments (Rajiakodi et al., 2026), labelled abusive vs. non-abusive. Text was kept raw, without stopword removal or heavy normalization.
  • Model: fine-tuned MuRIL with a fully connected layer on the [CLS] token, trained with cross-entropy loss. Settings were learning rate 2×10⁻⁵ and batch size 16 over 10 epochs, and the best checkpoint (epoch 7) was chosen by validation metrics.
  • Results (validation accuracy / macro-F1):
    • Proposed MuRIL + XAI: 0.8140 / 0.8126
    • TF-IDF + SVM: 0.7975 / 0.7975
    • Hybrid TF-IDF + MuRIL: 0.7900 / 0.7934
    • MuRIL baseline: 0.7800 / 0.7809
  • Explainability: a leave-one-word-out perturbation scores each word by the log-odds difference, without changing the model. Most errors involve implicit or context-dependent abuse.
  • Limitations: binary labels only, a single dataset, and word-removal explanations that miss interactions between words. Code is released at github.com/mirud5173/Abusive-Tamil-Comment-Detection-using-Transformer-Models.

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

🎓Vellore Institute of Technology UniversityVellore · India7 works · 268 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksPsychology · 1,552 worksSocial Psychology · 306 works

Cite

Cite as Mirudhula Sankar and Ratnavel Rajalakshmi. 2026. DLRG@DravidianLangTech 2026: Explainable Transformer-Based Detection of Abusive Tamil Text Targeting Women on Social Media. 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-1.34

@inproceedings{sankar2026dlrg,
  title = {{DLRG@DravidianLangTech 2026: Explainable Transformer-Based Detection of Abusive Tamil Text Targeting Women on Social Media}},
  author = {Mirudhula Sankar and Ratnavel Rajalakshmi},
  booktitle = {Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages},
  year = {2026},
  doi = {10.18653/v1/2026.dravidianlangtech-1.34},
  url = {https://doi.org/10.18653/v1/2026.dravidianlangtech-1.34}
}

References

4 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. G. Dhanyashree and 1 others. 2025. Lin-guaists@dravidianlangtech 2025: Abusive tamil and malayalam text targeting women on social media. In Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages. find
  2. Tareque Md Hanif and Md Rashadur Rahman. 2025. Cuet_agile@dravidianlangtech 2025: Fine-tuning transformers for detecting abusive text targeting women from tamil and malayalam texts. In Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages. find
  3. R Ramesh Kannan, Ratnavel Rajalakshmi, and Lokesh Kumar. 2021. Indicbert based approach for sentiment analysis on code-mixed tamil tweets. In FIRE (Working Notes), pages 729–736. find
  4. Md Mubasshir Naib and 1 others. 2025. cuetrap-tors@dravidianlangtech 2025: Transformer-based approaches for detecting abusive tamil text targeting women on social media. In Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages. find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sankar-2026-dlrg-dravidianlangtech-2026-explainable-transformer/