Elamur Research⌕
☰
home / papers / 2026 / FLAICOL: Flip-Point-Led Augmentation for Imbalanced Code-Mixed Offensive Language Detection

FLAICOL: Flip-Point-Led Augmentation for Imbalanced Code-Mixed Offensive Language Detection

Paper · 2026 · Danish Mohammed and Vidhya Kamakshi

Summary

Detecting offensive content in Dravidian code-mixed social media posts is difficult. Users switch scripts and languages, spell words inconsistently, and offensive examples are a small minority of the labeled data. The authors propose FLAICOL, which addresses this imbalance by creating extra offensive-class training examples close to the classifier's decision boundary. It adapts the closest-flip-point and homotopy idea of Yousefzadeh and O'Leary (2020) to Transformers. First, the classification-head bias is adjusted until an input can be flipped. Then the bias is gradually returned to its original value while small token-embedding perturbations are optimized. The resulting continuous embeddings are mapped back to real tokens by nearest-neighbour lookup, checked with the original classifier, and added to the training set before retraining. Experiments covered the Tamil–English, Malayalam–English and Kannada–English parts of DravidianCodeMix with XLM-RoBERTa-base and MuRIL-base. FLAICOL raised Offensive-F1 consistently, though not significantly. The approach also serves as an interpretability probe that shows where model decisions are fragile.

Key points

  • Backbones are XLM-RoBERTa-base and MuRIL-base, fine-tuned for binary Not-Offensive vs Offensive classification. Training used PyTorch and HuggingFace on a Colab NVIDIA T4 GPU with FP16.
  • The Kannada–English corpus contains 5874 examples: 4397 Not-Offensive and only 1477 Offensive.
  • Flip search settings were a maximum sequence length of 64, penalty weight α ∈ {1, 10, 100} and a learning rate of 10⁻².
  • In LaBSE-based diagnostics, XLM-derived flips stayed closer to the original data distribution and were more semantically faithful. MuRIL-derived flips were noisier in style and likely need stricter filtering, such as script or similarity constraints.
  • Offensive-F1 gains were small and not statistically significant, which the authors attribute partly to using few random seeds. Proposed future work includes a Bengali–Hindi–English trilingual corpus and combining the method with back-translation.

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

Also on the byline: National Institute of Technology Calicut.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Sentiment & emotion · 972 worksMultilingual & cross-lingual NLP · 828 works

Cite

Cite as Danish Mohammed and Vidhya Kamakshi. 2026. FLAICOL: Flip-Point-Led Augmentation for Imbalanced Code-Mixed Offensive Language Detection. 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-1.4

@inproceedings{mohammed2026flaicol,
  title = {{FLAICOL: Flip-Point-Led Augmentation for Imbalanced Code-Mixed Offensive Language Detection}},
  author = {Danish Mohammed and Vidhya Kamakshi},
  booktitle = {Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages},
  year = {2026},
  doi = {10.18653/v1/2026.dravidianlangtech-1.4},
  url = {https://doi.org/10.18653/v1/2026.dravidianlangtech-1.4}
}

References

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

  1. Koyel Ghosh and Apurbalal Senapati. 2022. Hate speech detection: a comparison of mono and multilingual transformer model with cross-language evalu-Abdullah Al Nahian, Mst Rafia Islam, Azmine Toushik Wasi, and Md Manjurul Ahsan. 2025. Nlpopsciol@ ation. In Proceedings of the 36th Pacific Asia Confere… find
  2. Jiaxiang Liu, Xuyi Chen, Shikun Feng, Shuohuan Wang, Xuan Ouyang, Yu Sun, Zhengjie Huang, and Weiyue Su. 2020. Kk2018 at SemEval-2020 task 9: Adversarial training for code-mixing sentiment classification. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 817–823, Barcelona (onl… find
  3. Dhiman Goswami, Nishat Raihan, Antara Mahmud, Antonios Anastasopoulos, and Marcos Zampieri. 2023. Offmix-3l: A novel code-mixed test dataset in bangla-english-hindi for offensive language identification. In Proceedings of the 11th International Workshop on Natural Language Processing for Social Medi… find
  4. Endang Wahyu Pamungkas and Viviana Patti. 2019. Cross-domain and cross-lingual abusive language detection: A hybrid approach with deep learning and a multilingual lexicon. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 363… find
  5. Shenson Joseph and Herat Joshi. 2024. Ulmfit: universal language model fine-tuning for text classification. International Journal of Advanced Medical Sciences and Technology, 4(6):10–54105. find
  6. Pallabi Kakati and Devendra Dandotiya. 2024. Automatic detection of hate speech in code-mixed indian languages in twitter social media interaction using dconvblstm-muril ensemble method. Social Network Analysis and Mining, 14(1):108. find
  7. Md Nishat Raihan, Dhiman Goswami, Antara Mahmud, Antonios Anstasopoulos, and Marcos Zampieri. 2023. Sentmix-3l: A bangla-english-hindi code-mixed dataset for sentiment analysis. In Proceedings of the First Workshop in South East Asian Language Processing, pages 79–84. find
  8. Vidhya Kamakshi and Narayanan C Krishnan. 2023. Explainable image classification: The journey so far and the road ahead. AI, 4(3):620–651. find
  9. Nishat Raihan, Dhiman Goswami, Antara Mahmud, Antonios Anastasopoulos, and Marcos Zampieri. 2024. Emomix-3l: a code-mixed dataset for bangla-english-hindi for emotion detection. In Proceedings of the 7th Workshop on Indian Language Data: Resources and Evaluation, pages 11–16. find
  10. Sanjana Kavatagi and Rashmi Rachh. 2025. Hastika: hate speech and target identification in kannada-english code-mixed text. Language Resources and Evaluation, pages 1–46. find
  11. Priya Rani, Shardul Suryawanshi, Koustava Goswami, Bharathi Raja Chakravarthi, Theodorus Fransen, and John Philip McCrae. 2020. A comparative study of different state-of-the-art hate speech detection methods in hindi-english code-mixed data. In Proceedings of the Second Workshop on Trolling, Aggress… find
  12. Simran Khanuja, Diksha Bansal, Sarvesh Mehtani, Savya Khosla, Atreyee Dey, Balaji Gopalan, Dilip Kumar Margam, Pooja Aggarwal, Rajiv Teja Nagipogu, Shachi Dave, and 1 others. 2021. Muril: Multilingual representations for indian languages. arXiv preprint arXiv:2103.10730.
  13. Pradeep Kumar Roy and Abhinav Kumar. 2025. Ensuring safety in digital spaces: Detecting code-mixed hate speech in social media posts. Data & Knowledge Engineering, page 102409. find
  14. Diederik P Kingma and Jimmy L Ba. 2015. Adam : A method for stochastic optimization. International Conference on Learning Representations, 7.
  15. V Sanh. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In Proceedings of Thirty-third Conference on Neural Information Processing Systems (NIPS2019). find
  16. Rohith Gowtham Kodali, Durga Prasad Manukonda, and Daniel Iglesias. 2025. bytesizedllm@ nlu of devanagari script languages 2025: Hate speech detection and target identification using customized attention bilstm and xlm-roberta base embeddings. In Proceedings of the First Workshop on Challenges in Pr… find
  17. Happy Khairunnisa Sariyanto, Diclehan Ulucan, Oguzhan Ulucan, and Marc Ebner. 2025. Towards explainable hate speech detection. In Findings of the Association for Computational Linguistics: ACL 2025, pages 12883–12893. find
  18. Niranjan Kumar, Pranav Gupta, Billodal Roy, and Souvik Bhattacharyya. 2025. Lexilogic@ dravidianlangtech 2025: Detecting misogynistic memes and abusive tamil and malayalam text targeting women on social media. In Proceedings of the Fifth Workshop on Speech, Vision, and Language Technologies for Drav… find
  19. Sunil Saumya, Abhinav Kumar, and Jyoti Prakash Singh. 2021. Offensive language identification in dravidian code mixed social media text. In Proceedings of the First Workshop on Speech and Language Technologies for Dravidian Languages, pages 36–45. find
  20. Rajat Sharma, Nikhil Reddy, Vidhya Kamakshi, Narayanan C Krishnan, and Shweta Jain. 2021. Maire-a model-agnostic interpretable rule extraction procedure for explaining classifiers. In International Cross-Domain Conference for Machine Learning and Knowledge Extraction, pages 329–349. Springer. find
  21. Ritesh Kumar, Bornini Lahiri, Atul Kr Ojha, and Akanksha Bansal. 2020. Comma@ fire 2020: Exploring multilingual joint training across different classification tasks. In FIRE (Working Notes), pages 823–828. find
  22. Rizwana Kallooravi Thandil, Muneer Vk, and Sabique Pv. 2025. A novel deep learning framework with advanced feature engineering for hate speech detection in accented malayalam speech. Humanities and Social Sciences Communications, 13(1):10. find
  23. P Deepasree Varma, P Vinod, M Nandakumar, K Akshay, and Akhil Madhu. 2022. Hate speech detection in english and malayalam code-mixed text using bert embedding. In 2022 International Conference on Computing, Communication, Security and Intelligent Systems (IC3SIS), pages 1–6. find
  24. Roozbeh Yousefzadeh and Dianne P. O’Leary. 2020. Deep learning interpretation: Flip points and homotopy methods. In Proceedings of The First Mathematical and Scientific Machine Learning Conference, volume 107 of Proceedings of Machine Learning Research, pages 1–26. PMLR. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mohammed-2026-flaicol-flip-point-led-augmentation/