Elamur Research⌕
☰
home / papers / 2026 / A Feature-Fusion Ensemble Approach for Tamil Hate Speech Detection

A Feature-Fusion Ensemble Approach for Tamil Hate Speech Detection

Paper · 2026 · Sathasivam Nerujan and Kengatharaiyer Sarveswaran

Summary

The paper addresses automatic detection of offensive and hateful content in Tamil social media text. The task is difficult because Tamil is low-resource and agglutinative, and online users mix it with English, Romanized spellings and slang. Standard subword tokenizers can split such insults into pieces that weaken their signal.

The authors build a hybrid classifier from two parts:

  • Contextual features: hidden states from the last four layers of L3Cube-TamilBERT are averaged, then summarised with both mean and max pooling.
  • Subword features: FastText character n-gram sentence embeddings.

The two feature sets are concatenated and fed to an MLP. They compare this against single fine-tuned transformers (TamilBERT, MuRIL, mBERT), stacking ensembles and TF-IDF fusion. Training and testing use a deduplicated Kaggle Tamil Offensive Speech Detection dataset of YouTube comments, plus HASOC Code-Mixed Tamil for comparison with earlier systems.

The hybrid reaches a Macro-F1 of 0.7883 with comparatively strong hate recall. On HASOC it scores 0.8955 Macro-F1, ahead of the prior results the authors cite. The work suggests that one language-specific encoder with richer pooling plus subword embeddings can rival heavier multilingual ensembles.

Key points

  • Data: after duplicate removal, the Kaggle dataset has 5,953 hate / 18,942 non-hate training items, 662 / 2,105 validation items and 1,678 / 5,181 test items. Emojis were deliberately kept as affective cues.
  • Main model: TamilBERT + FastText (layer averaging) scores Macro-F1 0.7883, hate precision 0.6375 and hate recall 0.7503.
    • A Hate-BERT + FastText variant scored slightly higher, at 0.7888.
    • A MuRIL + XLM-R stacking ensemble gave the best hate precision (0.9296).
  • Ablations with TamilBERT:
    • Dual pooling scored 0.7755, versus 0.7686 for the CLS token and 0.7356 for mean pooling.
    • Last-4 layer averaging scored 0.7836, versus 0.7356 for the last layer alone.
    • TF-IDF fusion with MuRIL + XLM-R performed worst (0.7004).
  • Comparison with prior work: on HASOC Code-Mixed Tamil (SubTask 2), the hybrid's 0.8955 Macro-F1 beats Siva@HASOC (0.88), MUCIC and PSG@HASOC.
    • The public model deoffxlmr-mono-tamil was set aside because a suspicious 0.98 score on the test set pointed to data leakage.
  • Robustness to input type: results are nearly equal on pure Tamil (0.7790) and code-mixed inputs (0.7809).

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

🎓University of JaffnaJaffna · Sri Lanka7 works · 70 citations

 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

🧭 Topics 2Research subjects and languages the work is about.Sentiment & emotion · 1,035 worksBias, fairness & safety · 587 works

Cite

Cite as Sathasivam Nerujan and Kengatharaiyer Sarveswaran. 2026. A Feature-Fusion Ensemble Approach for Tamil Hate Speech Detection. Proceedings of the Second workshop on Challenges in Processing South Asian Languages (CHiPSAL2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/2us8ubrf4jvi

@inproceedings{nerujan2026feature,
  title = {{A Feature-Fusion Ensemble Approach for Tamil Hate Speech Detection}},
  author = {Sathasivam Nerujan and Kengatharaiyer Sarveswaran},
  booktitle = {Proceedings of the Second workshop on Challenges in Processing South Asian Languages (CHiPSAL2026)},
  year = {2026},
  doi = {10.63317/2us8ubrf4jvi},
  url = {https://doi.org/10.63317/2us8ubrf4jvi}
}

References

19 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Junita Amalia, Sarah Rosdiana Tambunan, Susi Eva Maria Purba, and Walker Valentinus Simanjuntak. 2025. Enhancing hate speech detection: Leveraging emoji preprocessing with bilstm model. Journal of Information Systems and Informatics, 7(2). find
  2. Fazlourrahman Balouchzahi, Sepideh Bashang, Grigori Sidorov, and H. Shashirekha. 2022. Co-mata oli-code-mixed malayalam and tamil offensive language identification. find
  3. Sean Benhur and Kanchana Sivanraju. 2021. Psg@hasoc-dravidian codemixfire2021: Pre-trained transformers for offensive language identification in tanglish. CoRR, abs/2110.02852. — [arXiv:2110.02852](https://arxiv.org/abs/2110.02852) find
  4. Paula Fortuna and Sergio Nunes. 2018. A survey on automatic detection of hate speech in text. ACM Computing Surveys. find
  5. A. Iftikhar et al. 2025. Beyond words: A hybrid transformer-ensemble approach for detecting hate speech and offensive language on social media. PeerJ Computer Science. find
  6. S. F. Karim et al. 2025. Cuet_blitz_aces@lt-edi-2025: Leveraging transformer ensembles and majority voting for hate speech detection. In Proceedings of the 5th Workshop on Speech, Vision, and Language Technologies for Dravidian Languages (ACL Anthology). find
  7. Thomas Mandl, Sandip Modha, Anand Kumar M, and Bharathi Raja Chakravarthi. 2021. Overview of the hasoc track at fire 2020: Hate speech and offensive language identification in tamil, malayalam, hindi, english and german. In Proceedings of the 12th Annual Meeting of the Forum for Information Retrieva… find
  8. C Maria Nancy, N Radha, and R Swathika. 2025. SSN_IT_HATE@LT-EDI-2025: Caste and migration hate speech detection. In Proceedings of the 5th Conference on Language, Data and Knowledge: Fifth Workshop on Language Technology for Equality, Diversity, Inclusion, pages 84–89, Naples, Italy. Unior Press. find
  9. Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992,…
  10. B. Roy et al. 2025. Lexilogic@dravidianlangtech 2025: Multimodal hate speech detection in dravidian languages. In Proceedings of the DravidianLangTech@NAACL 2025 (ACL Anthology). find
  11. Ganesh Sundhar S, Durai Singh K, Gnanasabesan G, Hari Krishnan N, and Mc Dhanush. 2025. Wise@LT-EDI-2025: Combining classical and neural representations with multi-scale ensemble learning for code-mixed hate speech detection. In Proceedings of the 5th Conference on Language, Data and Knowledge: Fift… find
  12. Debjoy Saha, Naman Paharia, Deba-jit Chakraborty, Punyajoy Saha, and Animesh Mukherjee. 2021. Hate-alert@DravidianLangTech-EACL2021: Ensembling strategies for transformer-based offensive language detection. In Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Langua… find
  13. Siva Sai and Yashvardhan Sharma. 2020. Siva@ hasoc-dravidian-codemix-fire-2020: Multilingual offensive speech detection in code-mixed and romanized text. In FIRE (Working Notes), pages 336–343. find
  14. S. Sangeetham et al. 2024. Enhanced detection of hate speech in dravidian languages in social media using ensemble transformers. Interdisciplinary Journal of Information, Knowledge, and Management. find
  15. Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. CoRR, abs/1905.05950.
  16. Ning Xu. 2026. Tokenization and morphological fidelity in Uralic NLP. arXiv preprint. ArXiv:2602.04241. (Mandl et al., 2021) 7. find
  17. Language Resource References Bharathi Raja Chakravarthi, Ruba Priyadharshini, Vigneshwaran Muralidaran, Shardul Suryawanshi, Navya Jose, Elizabeth Sherly, and John P. McCrae. 2021. Overview of the track on sentiment analysis for dravidian languages in code-mixed text. In Proceedings of the 12th Annu… find
  18. Hata, Eshika Nahata. 2020. Tamil Offensive Speech Detection. Kaggle. Kaggle dataset; accessed 2026-02-27. find
  19. Asrita Venkata Mandalam, Yashvardhan Sharma, Bharathi Raja Chakravarthi, et al. 2020. Overview of the hasoc-dravidiancodemix shared task on offensive language identification in code-mixed text. In Proceedings of the 12th Forum for Information Retrieval Evaluation (FIRE 2020), pages 55–59. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nerujan-2026-feature-fusion-ensemble-approach-tamil/