Elamur Research⌕
☰
home / papers / 2026 / Still Loading@DravidianLangTech 2026: Telugu Prompt-Style Recovery using Multilingual Transformers

Still Loading@DravidianLangTech 2026: Telugu Prompt-Style Recovery using Multilingual Transformers

Paper · 2026 · Samonwita Sarker, Isnat Mehrin Sami, Priyontee Mojumder et al.

Summary

This work addresses Telugu style classification in the Prompt-Style Recovery shared task at DravidianLangTech@ACL 2026. Each instance pairs an original transcript excerpt with a stylistically rewritten version. A system must determine which of nine target styles produced the rewrite. Listed alphabetically, these are Authoritative, Formal, Humorous, Informal, Inspiring, Optimistic, Persuasive, Pessimistic and Serious. The task is hard because the labelled set is small, several styles overlap, and Telugu combines agglutinative morphology with code-mixing.

The Still Loading team fine-tuned MuRIL using a preprocessing step they call "Turbo Sandwich." It applies difflib word-level matching to find inserted or replaced tokens, then gives the encoder these edits plus nearby context rather than the full text. Training used Focal Loss (γ = 3) with square-root-dampened class weights.

The submitted MuRIL system ranked 7th with Macro-F1 0.1703. After gold labels were released, IndicBERTv2 scored 0.3000 under identical settings, clearly ahead of MuRIL's 0.2072. Ablations show that diff-based input and focal loss each help. The study offers a low-resource baseline for detecting pragmatic style in Dravidian languages, although absolute scores remain modest.

Key points

  • The dataset has 3,000 training, 300 development and 301 test samples, with nearly balanced classes (321–347 training instances per style).
  • MuRIL (google/muril-base-cased) was trained with max length 256, batch size 8, learning rate 4e-5, AdamW with a cosine schedule, 15 epochs and FP16.
  • Post-submission test-set Macro-F1 scores were:
    • IndicBERTv2: 0.3000
    • mBERT: 0.2492
    • MuRIL: 0.2072
    • DistilBERT: 0.1287
  • Despite the abstract framing MuRIL as the leading model, the post-submission results favour IndicBERTv2.
  • On the dev set, the plain-concatenation baseline with cross-entropy reached Macro-F1 0.1421. Turbo Sandwich with Focal Loss reached 0.1894 (+0.0473), the largest gain among the tested configurations.
  • In error analysis, Authoritative is frequently mislabelled as Formal because both use polite imperatives and honorific forms. Serious is often confused with Pessimistic. Validation loss plateaus around epoch 6, which suggests overfitting in later epochs.

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

🎓Chittagong University of Engineering & TechnologyChittagong · Bangladesh25 works · 29 citations

 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 1Research subjects and languages the work is about.Multilingual & cross-lingual NLP · 828 works

Cite

Cite as Samonwita Sarker, Isnat Mehrin Sami, Priyontee Mojumder, Arpita Mallik and Hasan Murad. 2026. Still Loading@DravidianLangTech 2026: Telugu Prompt-Style Recovery using Multilingual Transformers. 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-1.58

@inproceedings{sarker2026still,
  title = {{Still Loading@DravidianLangTech 2026: Telugu Prompt-Style Recovery using Multilingual Transformers}},
  author = {Samonwita Sarker, Isnat Mehrin Sami, Priyontee Mojumder, Arpita Mallik and Hasan Murad},
  booktitle = {Proceedings of the Sixth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages},
  year = {2026},
  doi = {10.18653/v1/2026.dravidianlangtech-1.58},
  url = {https://doi.org/10.18653/v1/2026.dravidianlangtech-1.58}
}

References

14 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Premjith B, Bharathi Raja Chakravarthi, Prasanna Kumar Kumaresan, Saranya Rajiakodi, Sai Prashanth Karnati, Sai Rishith Reddy Mangamuru, and Chandu Janakiram. 2024. Findings of the shared task on hate and offensive language detection in Telugu codemixed text (HOLD-Telugu)@DravidianLangTech 2024. In… find
  2. Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059. Association for Computational Linguistics.
  3. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Proc…
  4. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980–2988.
  5. Pramit Bhattacharyya and Arnab Bhattacharya. 2025. BanglaByT5: Byte-level modelling for Bangla. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 5551–5560. Association for Computational Linguistics. find
  6. Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. 2024. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927.
  7. Nicholas Carlini, Florian Tramèr, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Úlfar Erlingsson, Alina Oprea, and Colin Raffel. 2021. Extracting training data from large language models. In Proceedings of the 30th USENIX Security Symposium,…
  8. Faizad Ullah, Safiullah Faizullah, Imdad Ullah Khan, Turki Alghamdi, Toqeer Ali Syed, Ahmad B. Alkho-dre, Muhammad Sohaib Ayub, and Asim Karim. 2025. Prompt-based fine-tuning with multilingual transformers for language-independent sentiment analysis. Scientific Reports, 15:20834. find
  9. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association fo… find
  10. DravidianLangTech Organizing Committee. 2024. Overview of the shared task on prompting for Telugu at DravidianLangTech 2024. In Proceedings of the Fourth Workshop on Speech, Vision, and Language Technologies for Dravidian Languages (Dravidian-LangTech 2024), pages 1–12. Association for Computational… find
  11. Shubham Vatsal, Harsh Dubey, and Aditi Singh. 2025. Multilingual prompt engineering in large language models: A survey across nlp tasks. arXiv preprint arXiv:2505.11665. — [arXiv:2505.11665](https://arxiv.org/abs/2505.11665) find
  12. R. Vemula, M. Nuthi, and M. Srivastava. 2022. TeQuAD: Telugu question answering dataset. In Proceedings of the 19th International Conference on Natural Language Processing (ICON 2022), pages 300–307. Association for Computational Linguistics. find
  13. Divyanshu Kakwani, Anoop Kunchukuttan, Satish Golla, Gokul N.C., Pushpak Bhattacharyya, Mitesh M. Khapra, and Pratyush Kumar. 2020. Indicnlpsuite: Monolingual corpora, evaluation benchmarks and pre-trained multilingual language models for indian languages. Findings of EMNLP.
  14. Simran Khanuja, Diksha Bansal, Sarvesh Mehtani, Savya Khosla, Atreyee Dey, Balaji Gopalan, Dilip Kumar Margam, Pooja Mandhale, Ashwin Mantena, Shachi Bhardwaj, Vinay Niyogi, Mirac Sharifi, Chung-Cheng Pan, and Jason Riesa. 2021. Muril: Multilingual representations for indian languages. In arXiv prep…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sarker-2026-still-loading-dravidianlangtech-2026-telugu/