Elamur Research⌕
☰
home / papers / 2026 / Nepali Lemmatization with Multilingual Transformers: Intrinsic and Extrinsic Evaluation in a Low-Resource Setting

Nepali Lemmatization with Multilingual Transformers: Intrinsic and Extrinsic Evaluation in a Low-Resource Setting

Summary

Nepali morphology is highly inflected, and earlier Nepali lemmatizers were mostly rule-based or TRIE-based. Those systems break down on unseen or misspelled words. The authors recast lemmatization as text-to-text generation and fine-tune three multilingual encoder–decoder models: mBART-large-50, mT5-base and mT5-small. They do not use POS tags or morphological feature supervision.

The training data has 8,000 word–lemma pairs. Of these, 5,000 are gold pairs from a public Nepali lemmatization repository. The other 3,000 were derived from inflection and suffix rules and then checked by a native speaker. Target lemmas follow the dictionary infinitive convention (e.g., garnu) rather than bare stems.

Models were scored on accuracy, CER, character-level BLEU and morphological coverage. They were also tested extrinsically on two downstream tasks: Hindi–Nepali cross-lingual alignment and information retrieval.

mT5-base gave the best overall scores, with mBART-large-50 nearly level. Lemmatizing with mBART produced large downstream gains. The study shows that pretrained multilingual seq2seq models can learn morphological normalization for a low-resource language from a few thousand examples.

Key points

  • Data: 5,000 gold pairs from the dpakpdl NepaliLemmatizer project plus 3,000 rule-derived, human-verified pairs. The 80/20 train/test split has no overlap. A separate 700-word error-analysis set (100 words for each of 7 POS) was drawn from the Pragya Nepali Brihat Shabdakosh, with word frequencies taken from FineWeb2.
  • Setup: Inputs were given as "lemmatize: <word>" prompts, normalized to NFC and tokenized with SentencePiece. The maximum sequence length was 32 tokens, and fine-tuning used Hugging Face Transformers.
  • mT5-base: best overall, with 96.1% accuracy and 1.1% CER at a learning rate of 5×10⁻⁴. It was somewhat weaker on complex morphological variation.
  • mBART-large-50: 96.0% accuracy and the strongest morphological coverage (0.970).
  • Downstream gains with mBART: Hindi–Nepali alignment rose from 12.86% to 41.61%. Retrieval MAP@1 with a binary index rose from 0.71 to 0.90.

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: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/2o6euz7qakr5

@inproceedings{regmi2026nepali,
  title = {{Nepali Lemmatization with Multilingual Transformers: Intrinsic and Extrinsic Evaluation in a Low-Resource Setting}},
  author = {Regmi, Sunil and Dawadi, Sundeep and Bal, Bal Krishna},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/2o6euz7qakr5},
  url = {https://doi.org/10.63317/2o6euz7qakr5}
}

References

32 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Darwin Acharya, Sundeep Dawadi, Shivram Saud, and Sunil Regmi. 2025. Paramananda@nlu of devanagari script languages 2025: Detection of language, hate speech and targets using fasttext and bert. In Proceedings of the First Workshop on Challenges in Natural Language Understanding of Devanagari Script… find
  2. Emre Acikgoz, Tilek Chubakov, Müge Kural, Gözde İşgüder, and Deniz Yuret. 2022. Transformers on multilingual clause-level morphology. pages 100–105. find
  3. Mikel Artetxe, Gorka Labaka, and Eneko Agirre. 2018. A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 789–798. find our card
  4. B. K. Bal and P. Shrestha. 2004. The nepali spellchecker using hunspell framework. In PAN Localization Working Papers 2004–2007. PAN Localization Group. find
  5. Bal Krishna Bal, Basanta Karki, and Prajol Shrsetha. 2004. Nepali spellchecker 1.1 and the thesaurus, research and development. find
  6. Bal Krishna Bal and Prajol Shrestha. 2007. Architectural and system design of the nepali grammar checker. find
  7. Sunil M. Bhat and Rajendra Rai. 2012. Building morphological analyzer for nepali. Journal of Modern Languages, 22(1):45–58. find
  8. Sudha Bhingardive and Pushpak Bhattacharyya. 2016. Word sense disambiguation using indowordnet. In The WordNet in Indian Languages, pages 243–260. Springer. find
  9. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association fo… find
  10. R. Dave and P. Balani. 2015. Survey paper of different lemmatization approaches. In Proceedings of the International Conference on Advances in Technology and Engineering (ICAT-EST 2015), volume 8, pages 366–370, India. IJRAT. find
  11. Dadhi Ram Ghimire, Sanjeev Panday, and Aman Shakya. 2024. Information extraction from a large knowledge graph in the nepali language. National College of Computer Studies Research Journal, 3(1):33–49. find
  12. Pravesh Koirala and Nobal B Niraula. 2021. Npvec1: Word embeddings for nepali-construction and evaluation. In Proceedings of the 6th Workshop on Representation Learning for NLP (RepL4NLP-2021), pages 174–184. find
  13. Pravesh Koirala and Aman Shakya. 2020. A nepali rule-based stemmer and its performance on different nlp applications. arXiv preprint arXiv:2002.09901. — [arXiv:2002.09901](https://arxiv.org/abs/2002.09901) find
  14. Taku Kudo and John Richardson. 2018. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of EMNLP: System Demonstrations, pages 66–71. find our card
  15. Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020a. Multilingual denoising pre-training for neural machine translation. find
  16. Yinhan Liu, Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2020b. Multilingual denoising pre-training for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 162–176, Online. Association for Computational Linguistics. find
  17. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. find
  18. Tomas Mikolov, Edouard Grave, Piotr Bojanowski, Christian Puhrsch, and Armand Joulin. 2018. Advances in pre-training distributed word representations. In Proceedings of the International Conference on Language Resources and Evaluation (LREC 2018). find our card
  19. Minh Van Nguyen, Viet Dac Lai, Amir Pouran Ben Veyseh, and Thien Huu Nguyen. 2021. Trankit: A light-weight transformer-based toolkit for multilingual natural language processing. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2021… find our card
  20. R. Paul, S. Sah, and S. Joshi. 2014. An affix removal stemmer for nepali language. Proceedings of the National Conference on Information Technology. find
  21. Balaram Prasain. 2011a. A Computational Analysis of Nepali Morphology: A Model for Natural Language Processing. Ph.D. thesis, Central Department of Linguistics, Tribhuvan University, Kathmandu, Nepal. find
  22. Puskar Prasain. 2011b. Nepali lemmatizer: A rule based approach. In Proceedings of the 9th International Conference on Natural Language Processing (ICON), Kharagpur, India. NLP Association of India. find
  23. Chiranjibi Sitaula. 2013. A hybrid algorithm for stemming of nepali text. Intelligent Information Management, 5(4). find
  24. B. Subedi, S. Regmi, B. K. Bal, and P. Acharya. 2024. Exploring the potential of large language models (llms) for low-resource languages: A study on named-entity recognition (ner) and part-of-speech (pos) tagging for nepali language. In Proceedings of the 2024 Joint International Conference on Compu… find
  25. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Ankur Barua, and Colin Raffel. 2021. mt5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 483–4… find
  26. S. Yadav and S. Shakya. 2020. A neural morphological analyzer for nepali language. In Proceedings of LREC. 12. find
  27. Language Resource References Disisbig. 2025. Nepali News Dataset. Kaggle Datasets. PID https://www.kaggle.com/datasets/disisbig/nepali-news-dataset. Accessed 1 October 2025. dpakpdl. 2025. Gold Data for Manually Annotated Corpus in the NepaliLem-matizer Project. GitHub. PID https://github.com/dpakpd… ↗
  28. CFILT (Computational Foundations for Indian Language Technology). 2025. RoundTripOCR-Nepali. Hugging Face Datasets. PID https://huggingface.co/datasets/cfilt/RoundTripOCR-nepali. Post-OCR error correction dataset for Nepali language using RoundTripOCR technique, accessed 1 October 2025. ↗
  29. Nepal Academy. 2022. Pragya Nepali Brihat Shab-dakosh. Nepal Pragya Pratisthan, 10th (Revised). find
  30. Nivre, Joakim and de Marneffe, Marie-Catherine and Ginter, Filip and Hajić, Jan and Manning, Christopher D. and Pyysalo, Sampo and Schuster, Sebastian and Tyers, Francis and Zeman, Daniel. 2020. Universal Dependencies v2: An Evergrowing Multilingual Treebank Collection. European Language Resources A… find
  31. Guilherme Penedo and Hynek Kydlíček and Vinko Sabolčec and Bettina Messmer and Negar Foroutan and Amir Hossein Kargaran and Colin Raffel and Martin Jaggi and Leandro Von Werra and Thomas Wolf. 2025. FineWeb2: One Pipeline to Scale Them All – Adapting Pre-Training Data Processing to Every Language. find
  32. C. V. Shivaramakrishna. 1977. Trilingual Dictionary: Hindi–English–Nepali. Jai Gyan (via Digital Library of India). PID https://archive.org/details/dli.language.1977. Digitised version, accessed 1 October 2025. ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/regmi-2026-nepali-lemmatization-multilingual-transformers-intrinsic/