Elamur Research⌕
☰
home / papers / 2026 / Abusive Content Detection in Telugu-English Code-Mixed Social Media Using Hybrid Transformer Architectures

Abusive Content Detection in Telugu-English Code-Mixed Social Media Using Hybrid Transformer Architectures

Paper · 2026 · Bojja Revanth Reddy and Sivaiah Bellamkonda

Summary

The paper addresses binary detection of abusive versus non-abusive comments in Telugu-English code-mixed social media text. In this setting, Romanized Telugu, inconsistent transliteration, slang and frequent language switching weaken systems built for a single language. The authors build an augmented corpus of 49,910 comments. They expand it with synonym replacement, controlled word insertion and deletion, and generated spelling and dialect variants. They also normalize it by stripping URLs, emojis and stopwords, unifying spelling variants, and converting Telugu script back into Roman form. Their hybrid model passes transformer contextual embeddings through two parallel branches: a multi-channel CNN for short phrase cues and a BiLSTM for sequential context. The branch outputs are concatenated and classified by a sigmoid output layer. Among the mBERT, IndicBERT and DeBERTa backbones, DeBERTa reaches 0.97 on every metric, roughly ten points above the others. The work suggests that strong encoders paired with complementary feature extractors help in noisy, low-resource code-mixed settings. However, the original data source is not described, and it is unclear whether augmentation preceded the train/test split.

Key points

  • Final augmented dataset has 49,910 comments: 23,342 non-abusive (label 0) and 26,568 abusive (label 1), split 75% train / 25% test.
  • Architecture: transformer embeddings feed a parallel multi-channel CNN and a BiLSTM, followed by concatenation and a dense layer with Sigmoid. It was implemented in TensorFlow/Keras with binary cross-entropy, Adam and early stopping.
  • Results (accuracy / precision / recall / F1):
    • DeBERTa: 0.97 on all four metrics.
    • IndicBERT: 0.87 / 0.87 / 0.86 / 0.87.
    • mBERT: 0.86 on all four metrics.
  • The setup section names XLM-RoBERTa among the encoders, but no XLM-RoBERTa results appear in the comparison table, which lists mBERT instead.
  • Proposed future work includes domain-adaptive pretraining on Telugu-English social media corpora, attention-based fusion of CNN/BiLSTM features, lightweight transformers, multi-class labels, cross-lingual transfer to other Indian code-mixed languages, and explainability.

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 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

🧭 Topics 2Research subjects and languages the work is about.English · 2,263 worksMultilingual & cross-lingual NLP · 828 works

Cite

Cite as Bojja Revanth Reddy and Sivaiah Bellamkonda. 2026. Abusive Content Detection in Telugu-English Code-Mixed Social Media Using Hybrid Transformer Architectures. 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-1.1

@inproceedings{reddy2026abusive,
  title = {{Abusive Content Detection in Telugu-English Code-Mixed Social Media Using Hybrid Transformer Architectures}},
  author = {Bojja Revanth Reddy and Sivaiah Bellamkonda},
  booktitle = {Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages},
  year = {2026},
  doi = {10.18653/v1/2026.dravidianlangtech-1.1},
  url = {https://doi.org/10.18653/v1/2026.dravidianlangtech-1.1}
}

References

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

  1. Utsab Barman, Amitava Das, Joachim Wagner, and media text. We proposed a hybrid architecture Jennifer Foster. 2014. Code mixing: A challenge for that combines transformer-based contextual em-language identification in the language of social media. In Proceedings of the First Workshop on Compu-beddin… find
  2. Piotr Bojanowski, Edouard Grave, Armand Joulin, Experimental results demonstrate that the emand Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5:135–146.
  3. Bharathi Raja Chakravarthi and R. Priyadharshini. 2020. A dataset for Tamil-English code-mixed offensive text classification. In Proceedings of the 12th Language Resources and Evaluation Conference (LREC). find
  4. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, et al. 2020. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116. [16] Akarsh Srivastava and Ankush Singh. 2020. PHINC: A parallel Hindi-English code-mixed corpus. In Proceedings of the 12th… find
  5. J. Angel Diaz-Garcia and Joao Paulo Carvalho. 2025. A survey of textual cyber abuse detection using cutting-edge language models and large language models. arXiv preprint arXiv:2501.05443. [18] Kusampudi Varma, Preetham Sathineni, and Radhika Mamidi. 2021. Sentiment analysis in code-mixed Telugu-Eng… — [arXiv:2501.05443](https://arxiv.org/abs/2501.05443) find
  6. P. Jakkula. 2020. Comparative study on Telugu text classification using machine learning and deep learning techniques. SciSpace Link. [19] P. Vijayaraghavan and A. Das. 2022. Abusive speech detection using deep learning models. Unpublished manuscript. find
  7. V. Khanduja, S. Singh, and H. Bansal. 2024. Telugu language hate speech detection using deep learning transformer models: Corpus generation and evaluation. Systems and Soft Computing, 6:200112. find
  8. A. Kumar, A. Ojha, and T. Solorio. 2021. Code-mixed abusive language detection: Challenges and solutions. Unpublished manuscript. find
  9. Scott M. Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems.
  10. S. Maddu and V. R. Sanapala. 2024. A survey on NLP tasks, resources and techniques for low-resource Telugu-English code-mixed text. ACM Digital Library. find
  11. Thomas Mandl, Sandip Modha, Punyajoy Majumder, Diptesh Patel, et al. 2019. Overview of the HASOC track at FIRE 2019. In Proceedings of FIRE, pages 14–17. find
  12. Atul Krushna Puranik, Ritesh Kumar, Navneet Kumar, Bharathi Raja Chakravarthi, and Bornini Lahiri. 2021. Telugu-English code-mixed dataset for sentiment analysis and offensive language detection. In Proceedings of the First Workshop on Language Technology for Equality, Diversity and Inclusion, pages… find
  13. Juan Ramos. 2003. Using tf-idf to determine word relevance in document queries. In Proceedings of the First Instructional Conference on Machine Learning, pages 133–142. find
  14. Tharindu Ranasinghe, Marcos Zampieri, and Constantin Orasan. 2020. MUDES: Multilingual detection of offensive spans. In Proceedings of EMNLP, pages 6222–6231. find
  15. A. Singh and S. Rajput. 2021. Code-mixed sentiment classification using transformer models. In ICON 2021: 18th International Conference on Natural Language Processing. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/reddy-2026-abusive-content-detection-telugu-english/