Elamur Research⌕
☰
home / papers / 2026 / Representation-Aware Prompting for Zero-Shot Marathi Text Classification: IPA, Romanization, Repetition

Representation-Aware Prompting for Zero-Shot Marathi Text Classification: IPA, Romanization, Repetition

Paper · 2026 · Van-Hien Tran, Huy Hien Vu, Hideki Tanaka et al.

Summary

Instruction-tuned LLMs tend to struggle with zero-shot classification of low-resource languages written in non-Latin scripts, partly because poor tokenizer coverage produces weak input representations. Focusing on Marathi, the authors test whether the prompt alone, with no fine-tuning, can compensate. Each prompt keeps the original Devanagari sentence and adds one extra view of it: an IPA transcription (Epitran), a romanized version (Indic-Transliteration), or simply a second copy of the sentence. The repetition option is a fallback for cases where no conversion tools exist. They also test a prompt that combines all three.

Two frozen models, LLaMA-3.3-70B-Instruct and Qwen2.5-72B-Instruct, were evaluated with greedy decoding on L3CubeMahaSent and L3Cube-MahaHate. Every single augmentation improved accuracy over the script-only prompt, by as much as 2.6 points. The best option differed by model: phonetic input worked best for LLaMA, and repetition worked best for Qwen. Combining all three never matched the best single cue, which suggests that short, targeted additions work better than piling on context. This is a cheap, training-free technique that could apply to other underserved scripts.

Key points

  • Benchmarks:
    • L3CubeMahaSent: 3-class sentiment, 1,500 balanced test items.
    • L3Cube-MahaHate: binary HOF/NOT_HOF tweets, 3,750 balanced test items.
    • Accuracy is the metric.
  • LLaMA-3.3-70B-Instruct gains most from +IPA:
    • Sentiment rises from 70.47 to 73.07.
    • Hate detection rises from 92.51 to 94.69.
    • Romanization also helps, by +2.13 and +1.54 respectively.
  • Qwen2.5-72B-Instruct gains most from +Repetition:
    • Sentiment rises from 71.80 to 73.40.
    • Hate detection rises from 89.89 to 90.99.
  • +All (all three cues concatenated) underperforms the best single cue for both models on both tasks:
    • For Qwen on sentiment, it falls slightly below the baseline (71.60).
    • The authors attribute this to redundancy, longer prompts, and conversion noise.
  • Stated limitations:
    • Only Marathi and two very large models were tested.
    • Prompts that drop the original script were not explored.

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

🔬National Institute of Information and Communications TechnologyTokyo · Japan23 works · 2,090 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.Language models & pretraining · 3,320 works

Cite

Cite as Van-Hien Tran, Huy Hien Vu, Hideki Tanaka and Masao Utiyama. 2026. Representation-Aware Prompting for Zero-Shot Marathi Text Classification: IPA, Romanization, Repetition. Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.loreslm-1.37

@inproceedings{tran2026representation,
  title = {{Representation-Aware Prompting for Zero-Shot Marathi Text Classification: IPA, Romanization, Repetition}},
  author = {Van-Hien Tran, Huy Hien Vu, Hideki Tanaka and Masao Utiyama},
  booktitle = {Proceedings of the Second Workshop on Language Models for Low-Resource Languages (LoResLM 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.loreslm-1.37},
  url = {https://doi.org/10.18653/v1/2026.loreslm-1.37}
}

References

26 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Tatiana Anikina, Jan Cegin, Jakub Simko, and Simon Ostermann. 2025. A rigorous evaluation of LLM data generation strategies for low-resource languages. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 8282–8303, Suzhou, China. Association for Computati…
  2. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Vikrant Dewangan, Bharath Raj S, Garvit Suri, and Raghav Sonavane. 2025. When every token counts: Optimal segmentation for low-resource language models. In Proceedings of the 31st International Conferen… find
  3. Aleksandra Edwards and José Camacho-Collados. 2024. Language models for text classification: Is in-context learning enough? In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, LREC/COLING 2024, 20-25 May, 2024, Torino, Italy, pag… find
  4. Taisei Enomoto, Hwichan Kim, Zhousi Chen, and Mamoru Komachi. 2025. A fair comparison without translationese: English vs. target-language instructions for multilingual LLMs. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics:… find
  5. Barah Fazili, Ashish Agrawal, and Preethi Jyothi. 2024. Boosting zero-shot crosslingual performance using llm-based augmentations with effective data selection. In Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeting, August 11-16, 2024, pages 13… find
  6. International Phonetic Association. 1999. Handbook of the International Phonetic Association: A Guide to the Use of the International Phonetic Alphabet. Cambridge University Press, Cambridge, UK. find
  7. Jaavid J, Raj Dabre, Aswanth M, Jay Gala, Thanmay Jayakumar, Ratish Puduppully, and Anoop Kunchukuttan. 2024. RomanSetu: Efficiently unlocking multilingual capabilities of large language models via Romanization. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguisti… find
  8. Haeji Jung, Changdae Oh, Jooeon Kang, Jimin Sohn, Kyungwoo Song, Jinkyu Kim, and David R Mortensen. 2024. Mitigating the linguistic gap with phonemic representations for robust cross-lingual transfer. In Proceedings of the Fourth Workshop on Multilingual Representation Learning (MRL 2024), pages 200… find
  9. Vani Kanjirangat, Tanja Samardzic, Ljiljana Dolamic, and Fabio Rinaldi. 2025. Tokenization and representation biases in multilingual models on dialectal NLP tasks. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 23992–24010, Suzhou, China. Association… find
  10. Atharva Kulkarni, Meet Mandhane, Manali Likhitkar, Gayatri Kshirsagar, and Raviraj Joshi. 2021. L3CubeMahaSent: A Marathi tweet-based sentiment analysis dataset. In Proceedings of the Eleventh Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 213–220, O… find
  11. Aristides Milios, Siva Reddy, and Dzmitry Bahdanau. 2023. In-context learning for text classification with many labels. In Proceedings of the 1st GenBench Workshop on (Benchmarking) Generalisation in NLP, pages 173–184, Singapore. Association for Computational Linguistics. find
  12. David R. Mortensen, Siddharth Dalmia, and Patrick Littell. 2018. Epitran: Precision G2P for many languages. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).
  13. Hoang Nguyen, Chenwei Zhang, Tao Zhang, Eugene Rohrbaugh, and Philip Yu. 2023. Enhancing cross-lingual transfer via phonemic transcription integration. In Findings of the Association for Computational Linguistics: ACL 2023, pages 9163–9175, Toronto, Canada. Association for Computational Linguistics. find
  14. Ercong Nie, Sheng Liang, Helmut Schmid, and Hinrich Schütze. 2023. Cross-lingual retrieval augmented prompt for low-resource languages. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8320–8340, Toronto, Canada. Association for Computational Linguistics. find
  15. Hrushikesh Patil, Abhishek Velankar, and Raviraj Joshi. 2022. L3Cube-MahaHate: A tweet-based Marathi hate speech detection dataset and BERT models. In Proceedings of the Third Workshop on Threat, Aggression and Cyberbullying (TRAC 2022), pages 1–9, Gyeongju, Republic of Korea. Association for Comput… find
  16. Priyaranjan Pattnayak, Hitesh Laxmichand Patel, and Amit Agarwal. 2025. Tokenization matters: Improving zero-shot NER for indic languages. CoRR, abs/2504.16977. — [arXiv:2504.16977](https://arxiv.org/abs/2504.16977) find
  17. Parth Patwa, Simone Filice, Zhiyu Chen, Giuseppe Castellucci, Oleg Rokhlenko, and Shervin Malmasi. 2024. Enhancing low-resource LLMs classification with PEFT and synthetic data. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation… find
  18. Aleksandar Petrov, Emanuele La Malfa, Philip H. S. Torr, and Adel Bibi. 2023. Language model tokenizers introduce unfairness between languages. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA…
  19. Aabha Pingle, Aditya Vyawahare, Isha Joshi, Rahul Tangsali, and Raviraj Joshi. 2023. L3Cube-MahaSent-MD: A multi-domain Marathi sentiment analysis dataset and transformer models. In Proceedings of the 37th Pacific Asia Conference on Language, Information and Computation, pages 274–281, Hong Kong, Ch… find
  20. Sukannya Purkayastha, Sebastian Ruder, Jonas Pfeiffer, Iryna Gurevych, and Ivan Vulić. 2023. Romanization-based large-scale adaptation of multilingual language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7996–8005, Singapore. Association for Computational… find
  21. Ammon Shurtz, Christian Richardson, and Stephen D. Richardson. 2025. When scripts diverge: Strengthening low-resource neural machine translation through phonetic cross-lingual transfer. In Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025), pages 336–346, Suzhuo, Chin… find
  22. Jimin Sohn, Haeji Jung, Alex Cheng, Jooeon Kang, Yilin Du, and David R Mortensen. 2024. Zero-shot cross-lingual NER using phonemic representations for low-resource languages. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 13595–13602, Miami, Florida,… find
  23. Jacob Mitchell Springer, Suhas Kotha, Daniel Fried, Graham Neubig, and Aditi Raghunathan. 2025. Repetition improves language model embeddings. In The Thirteenth International Conference on Learning Representations. find
  24. Hailay Kidu Teklehaymanot and Wolfgang Nejdl. 2025. Tokenization disparities as infrastructure bias: How subword systems create inequities in LLM access and efficiency. CoRR, abs/2510.12389. — [arXiv:2510.12389](https://arxiv.org/abs/2510.12389) find
  25. Abhishek Velankar, Hrushikesh Patil, and Raviraj Joshi. 2022. Mono vs multilingual BERT for hate speech detection and text classification: A case study in marathi. In Artificial Neural Networks in Pattern Recognition - 10th IAPR TC3 Workshop, ANNPR 2022, Dubai, United Arab Emirates, November 24-26,… find
  26. Yuqi Wang, Wei Wang, Qi Chen, Kaizhu Huang, Anh Nguyen, and Suparna De. 2023. Prompt-based zero-shot text classification with conceptual knowledge. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop), pages 30–38, Toronto,… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tran-2026-representation-aware-prompting-zero-shot/