KEC'S CODE CRAFTERS@DravidianLangTech 2026: Abusive Tamil Text Detection Targeting Women on Social Media

Paper · 2026 · Nivetha, Nethrasri S and Malliga Subramanian

Summary

This paper describes a system submitted to the DravidianLangTech-2026 shared task on identifying abusive Tamil comments aimed at women. The task is binary classification of YouTube comments as Abusive or Non-Abusive. Tamil is hard to process here because its morphology is complex and users often mix it with English or write it in transliterated form. The organisers supplied 25,948 training comments and 915 test comments, all manually annotated by fluent Tamil speakers.

The authors built a cleaning pipeline for noisy social-media text and represented comments with TF-IDF features. They then evaluated four scikit-learn classifiers against each other: Logistic Regression, a linear SVM, a Multinomial Naive Bayes model and a Random Forest. Hyperparameters and decision thresholds were tuned on an 80/20 train–validation split. Logistic Regression with a tuned threshold performed best, reaching 0.86 for both accuracy and F1.

The work offers a light, interpretable baseline for gender-targeted abuse detection in a low-resource language. Its error analysis shows that sarcasm, implied harassment and code-mixing remain open problems for lexical models.

Key points

Data: 26,863 Tamil YouTube comments in total, split into 25,948 for training and 915 for testing. The two classes are roughly balanced. Label 1 covers sexist, derogatory, threatening or sexually abusive remarks directed at women.
Preprocessing and features:
- Regex removal of URLs, @mentions, hashtags and special symbols.
- Conversion of emojis to text descriptions with the emoji library.
- Whitespace normalisation and lowercasing.
- TF-IDF with unigrams and bigrams, min_df = 2 and a cap of 10,000 features.
Models:
- LinearSVC with default settings.
- Logistic Regression with its decision threshold moved to 0.59.
- Random Forest selected by grid search (100 trees, unlimited depth, threshold 0.55).
- MultinomialNB.
Results (F1/accuracy): LR 0.86, SVM 0.84, RF 0.83, NB 0.80. RF's recall on the abusive class was only 0.75, which the authors attribute to overfitting on sparse features.
Limitations and future work: The models mostly failed on sarcasm, implicit abuse, slang and phonetic transliteration. The authors plan to try MuRIL, XLM-RoBERTa, ensembles and data augmentation. Code is released at github.com/NivethaVallarasu/dravidian.

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 2Computer Science · 16,667 works Artificial Intelligence · 16,194 works

Cite

Cite as Nivetha, Nethrasri S and Malliga Subramanian. 2026. KEC'S CODE CRAFTERS@DravidianLangTech 2026: Abusive Tamil Text Detection 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-1.43

@inproceedings{nivetha2026kec,
  title = {{KEC'S CODE CRAFTERS@DravidianLangTech 2026: Abusive Tamil Text Detection Targeting Women on Social Media}},
  author = {Nivetha, Nethrasri S and Malliga Subramanian},
  booktitle = {Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages},
  year = {2026},
  doi = {10.18653/v1/2026.dravidianlangtech-1.43},
  url = {https://doi.org/10.18653/v1/2026.dravidianlangtech-1.43}
}

References

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

[2] Subramanian, M., Shanmugavadivel, K., Indhuja, V.S., Kowshik, P., & Jayasurya, S. (2025). KECEm-power@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 Language… find
[10] Chakravarthi, B.R., Priyadharshini, R., Madasamy, A.K., Thavareesan, S., Sherly, E., Rajiakodi, S., Palani, B., Subramanian, M., Cn, S., & Chinnappa, D. (Eds.). (2025). Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages. Albuquerque, New Mexic… find
[3] Harini, B.N.S., Meghana, K.V., Supriya, K., Samiksha, T., & Premjith, B. (2025). HTMS@DravidianLangTech 2025: Fusing TF-IDF and BERT with Dimensionality Reduction for Abusive Language Detection in Tamil and Malayalam. In Proceedings of the Fifth Workshop on Speech, Vision, and Language Technolog… find
[5] G, D., et al. (2025). Abusive Text Detection Targeting Women in Tamil and Malayalam using Machine Learning Classifiers. In Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages (pp. XXX-XXX). Association for Computational Linguistics. find
G.Y., et al. (2025). [6] Bade, GS DravidianLangTech@2025: Women Targeted Abusive Texts Detection on Social Media. arXiv preprint, arXiv:2504.02863. https://arxiv.org/abs/2504.02863 — [arXiv:2504.02863](https://arxiv.org/abs/2504.02863) ↗
[7] Shanmugavadivel, K., Subramanian, M., Pooja Sree, M., Palanimurugan, V., & Roshini Priya, K. (2025). InnovationEngineers@DravidianLangTech 2025: Enhanced CNN Models for Detecting Misogyny in Tamil Memes Using Image and Text Classification. In Proceedings of the Fifth Workshop on Speech, Vision,… find
[8] Vaidyanathan, V.K., Srihari, V.K., & Durairaj, T. (2025). NLP goats@DravidianLangTech 2025: Towards Safer Social Media: Detecting Abusive Language Directed at Women in Dravidian Languages. In Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages.… find
[9] Thirumoorthy, S., Durairaj, T., & Rajalakshmi, R. (2025). Hydrangea@DravidianLanTech2025: Abusive language Identification from Tamil and Malay- find

Updated 1 November 2026

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nivetha-2026-kec-s-code-crafters-dravidianlangtech/