Elamur Research⌕
☰
home / papers / 2026 / Mining Naturally Romanized Seed Corpora without Romanizations

Mining Naturally Romanized Seed Corpora without Romanizations

Summary

Many languages with a non-Latin native script are also written informally in Latin letters. Corpora of such natural romanizations are rare, which makes it hard to build language identification (LID) systems that recognize this kind of text. The authors ask how to collect a first romanized seed corpus for a language that has no romanized samples at all. They use Tigrinya, written in Ge‘ez, as the test case.

They first run controlled tests on South Asian languages in Devanagari and Bengali scripts. These tests check whether a romanizer trained on one language transfers to a related language that shares its script. They also introduce a variant that routes input through a universal romanizer and generalizes better across languages.

They then train a romanizer on Amharic data from XTREME-UP, synthesize romanized Tigrinya, and add it to an LID model. That model labels Common Crawl text from MADLAD-400. Native-speaker review shows precision of 90% or more, and generic filtering raises precision further while losing some recall. The approach offers a cheap way to bootstrap resources for underserved informal registers.

Key points

  • Universal romanizers serve as baselines: uroman, uconv, and uconv′, which is uconv output with diacritics stripped and so closer to natural spellings.
  • Cross-language romanization is measured intrinsically as mCER% over 10 trials. The tests use Devanagari languages (brx, hin, kok, mai, mar, nep, san) and Bengali-script languages (asm, ben), each trained with native-script input and with uconv input.
  • Romanized Tigrinya was synthesized using Amharic romanizations from XTREME-UP, since no Tigrinya romanization lexicon exists. Mined output was judged by two L1 Tigrinya speakers and one L2 speaker.
  • The authors validated the language-agnostic paragraph filters on 540 sampled paragraphs. Estimated precision was 95.3% and recall 80.7%, with overall Cohen's Kappa of 0.81.
  • Precision of individual filters was mostly 100%. The lowest was 88%, for longest_mixed_word.

Interesting ideas

  • I1. When a language has no resources, borrow a model from a sibling language that shares its script, and treat the output only as a seed that later mining refines.
  • I2. First map input to a canonical, language-neutral form (a universal romanization), then learn the mapping from that form. This pivot transfers better across languages than learning directly from each native form.
  • I3. To cover a class absent from a classifier's training data, synthesize approximate examples. Then let the classifier harvest real examples from a large crawl, which become better training data.
  • I4. Validate a filtering pipeline by sampling separately from what each stage rejected and from what passed everything. This gives per-filter precision, not just one end-to-end number.
  • I5. Annotator guidelines include a worked example of each outcome: a positive case, an identified wrong language, an unidentified one, and non-language junk. Guidelines are refined over several rounds, with agreement tracked by kappa.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Mova could bootstrap silver data for rare English constructions from a related, better-covered construction, then keep only what the witness parsers agree on.
  • I2 (tree): For UD dialect conversion, Mova could normalize every treebank into one canonical intermediate scheme before learning conversions, so that rules transfer between treebanks.
  • I3 (learn): Synthetic sentences for missing frame or construction types could seed a detector that mines natural examples from raw text for training.
  • I4 (learn): Mova could audit each self-induced repair rule by sampling the cases it changed, a per-rule negative control that flags unreliable rules.
  • I5 (answer): Test sets for school-course and fable QA could use the same practice, with labeled negative and junk examples and multi-round agreement checks for gold answers.

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

Cite as Adrian Benton, Alexander Gutkin, Christo Kirov and Brian Roark. 2026. Mining Naturally Romanized Seed Corpora without Romanizations. Proceedings of the Fifteenth Language Resources and Evaluation Conference.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3fhfr77pugrj

@inproceedings{benton2026mining,
  title = {{Mining Naturally Romanized Seed Corpora without Romanizations}},
  author = {Adrian Benton, Alexander Gutkin, Christo Kirov and Brian Roark},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/3fhfr77pugrj},
  url = {https://doi.org/10.63317/3fhfr77pugrj}
}

References

47 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Niat Gebremichael Afeworki. 2018. Eritrean nationalism and the digital diaspora: Expanding diasporic networks via Twitter. Master’s thesis, African Studies, University of California (UCLA), Los Angeles, CA. find
  2. Cyril Allauzen, Michael Riley, Johan Schalkwyk, Wojciech Skut, and Mehryar Mohri. 2007. Open-Fst: A general and efficient weighted finite-state transducer library. In Proceedings of 12th International Conference on Implementation and Application of Automata (CIAA), pages 11–23, Prague, Czech Republic. Springer. find
  3. Adrian Benton, Alexander Gutkin, Christo Kirov, and Brian Roark. 2025. Improving informally romanized language identification. In Proceedings of EMNLP, forthcoming. find
  4. Victoria Bernal. 2005. Eritrea on-line: Diaspora, cyberspace, and the public sphere. American Ethnologist, 32(4):660–675. find
  5. Maximilian Bisani and Hermann Ney. 2008. Joint-sequence models for grapheme-to-phoneme conversion. Speech Communication, 50(5):434–451. find
  6. Verena Blaschke, Hinrich Schuetze, and Barbara Plank. 2023. A survey of corpora for Germanic low-resource languages and dialects. In Proceedings of the 24th Nordic Conference on Computational Linguistics (NoDaLiDa), pages 392– 414, Tórshavn, Faroe Islands. University of Tartu Library. our card
  7. Carmen Brandt. 2020. From a symbol of colonial conquest to the scripta franca: The Roman script for South Asian languages. In Carmen Brandt and Hans Harder, editors, Wege durchs Labyrinth: Festschrift zu Ehren von Rahul Peter Das, pages 1–36. CrossAsia-eBooks, Heidelberg, Germany. find
  8. Ralf Brown. 2014. Non-linear mapping for improved identification of 1300+ languages. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 627–632, Doha, Qatar. Association for Computational Linguistics. find
  9. Laurie Burchell, Alexandra Birch, Nikolay Bogoychev, and Kenneth Heafield. 2023. An open dataset and model for language identification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 865–879, Toronto, Canada. Association for Computational Linguistics. our card
  10. Isaac Caswell, Theresa Breiner, Daan van Esch, and Ankur Bapna. 2020. Language ID in the wild: Unexpected challenges on the path to a thousand-language web text corpus. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6588–6608, Barcelona, Spain (Online). International Committee on Computational Linguistics. our card
  11. Stephen Chrisomalis. 2010. Numerical notation: A comparative history, chapter Alphabetic Systems. Cambridge University Press, Cambridge, UK. find
  12. Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37–46. ↗
  13. Bettina Conrad. 2010. “We are the Prisoners of our Dreams“: Long-distance Nationalism and the Eritrean Diaspora in Germany. Ph.D. thesis, Department of Social Sciences, University of Hamburg, Hamburg, Germany. find
  14. Arthur P. Dempster, Nan M. Laird, and Donald B. Rubin. 1977. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society: series B (methodological), 39(1):1–22. find
  15. Rasul Dent, Pedro Ortiz Suarez, Thibault Clérice, and Benoît Sagot. 2025. Krèyolid from language identification towards language mining. arXiv preprint arXiv:2503.06547. ↗
  16. Paul D. Fallon. 2006. Blin orthography: a history and an assessment. In Selected proceedings of the 36th Annual Conference on African Linguistics (ACAL), pages 93–98, Statesboro, GA. find
  17. Paul D. Fallon. 2008. Language development in Eritrea: The case of Blin. In Barbara Soukup, Lyn Fogle, Jia Jackie Lou, Kendall A. King, and Natalie Schilling-Estes, editors, Endangered and Minority Languages and Language Varieties: Sustaining Linguistic Diversity, chapter 10, pages 145–158. Georgetown University Press, Washington DC. find
  18. Yitna Firdyiwek and Daniel Yaqob. 1997. The system for Ethiopic representation in https://www.researchgate.net/ASCII. publication/2682324_The_System_for_ Ethiopic_Representation_in_ASCII . ↗
  19. Fitsum Gaim, Wonsuk Yang, and Jong C. Park. 2022. GeezSwitch: Language identification in typologically related low-resourced East African languages. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 6578–6584, Marseille, France. European Language Resources Association. our card
  20. Michael Gasser. 2011. HornMorpho: a system for morphological processing of Amharic, Oromo, and Tigrinya. In Proceedings of Conference on Human Language Technology for Development, pages 94–99, Alexandria, Egypt. https://github.com/hltdi/HornMorpho . ↗
  21. Kanika Gupta, Monojit Choudhury, and Kalika Bali. 2012. Mining Hindi-English transliteration pairs from online Hindi lyrics. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 2459–2465, Istanbul, Turkey. European Language Resources Association (ELRA). find
  22. Ulf Hermjakob, Jonathan May, and Kevin Knight. 2018. Out-of-the-box universal Romanization tool uroman. In Proceedings of ACL 2018, System Demonstrations, pages 13–18, Melbourne, Australia. Association for Computational Linguistics. find
  23. Bhakti Jadhav, Himanshu Dutta, Shruti Kanitkar, Malhar Kulkarni, and Pushpak Bhattacharyya. 2025. An introduction to computational identification and classification of Upamā alaṇkāra. In Computational Sanskrit and Digital Humanities - World Sanskrit Conference 2025, pages 1–14, Kathmandu, Nepal. Association for Computational Lingustics. find
  24. Pratik Joshi, Christain Barnes, Sebastin Santy, Simran Khanuja, Sanket Shah, Anirudh Srinivasan, Satwik Bhattamishra, Sunayana Sitaram, Monojit Choudhury, and Kalika Bali. 2019. Unsung challenges of building and deploying language technologies for low resource language communities. In Proceedings of the 16th International Conference on Natural Language Processing, pages 211–219, International Institute of Information Technology, Hyderabad, India. NLP Association of India. find
  25. Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Hérve Jégou, and Tomas FastText.zip: Compressing Mikolov. 2016. text classification models. arXiv preprint arXiv:1612.03651. ↗ our card
  26. Amir Hossein Kargaran, Ayyoob Imani, François Yvon, and Hinrich Schuetze. 2023. GlotLID: Language identification for low-resource languages. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6155–6218, Singapore. Association for Computational Linguistics. our card
  27. Amir Hossein Kargaran, François Yvon, and Hinrich Schütze. 2024. GlotCC: An open broad-coverage CommonCrawl corpus and pipeline for minority languages. Advances in Neural Information Processing Systems, 37:16983–17005. find
  28. Christo Kirov, Cibu Johny, Anna Katanova, Alexander Gutkin, and Brian Roark. 2024. Context-aware transliteration of Romanized South Asian languages. Computational Linguistics, 50(2):475–534. find
  29. Sneha Kudugunta, Isaac Caswell, Biao Zhang, Xavier Garcia, Derrick Xin, Aditya Kusupati, Romi Stella, Ankur Bapna, and Orhan Firat. 2023. MADLAD-400: A multilingual and document-level large audited dataset. Advances in Neural Information Processing Systems (NeurIPS), 36:67284–67296. our card
  30. Lucy Linder, Michael Jungo, Jean Hennebert, Claudiu Cristian Musat, and Andreas Fischer. 2020. Automatic creation of text corpora for low-resource languages from the Internet: The case of Swiss German. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 2706–2711, Marseille, France. European Language Resources Association. find
  31. Marco Lui and Timothy Baldwin. 2012. langid.py: An off-the-shelf language identification tool. In Proceedings of the ACL 2012 System Demonstrations, pages 25–30, Jeju Island, Korea. Association for Computational Linguistics. our card
  32. Yash Madhani, Mitesh M. Khapra, and Anoop Bhasa-Abhijnaanam: Kunchukuttan. 2023a. Native-script and romanized language identification for 22 Indic languages. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 816–826, Toronto, Canada. Association for Computational Linguistics. find
  33. Yash Madhani, Sushane Parthan, Priyanka Bedekar, Gokul Nc, Ruchi Khapra, Anoop Kunchukuttan, Pratyush Kumar, and Mitesh Khapra. 2023b. Aksharantar: Open Indic-language transliteration datasets and models for the next billion users. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 40–57, Singapore. Association for Computational Linguistics. find
  34. Ronny Meyer and Bedilu Wakjira. 2023. Scripts and writing in Ethiopia. In Ronny Meyer, Bedilu Wakjira, and Zelealem Leyew, editors, The Oxford Handbook of Ethiopian Languages, Oxford Handbooks, pages 86–100. Oxford University Press, Oxford, UK. find
  35. Alice Millour and Karën Fort. 2020. Text corpora and the challenge of newly written languages. In Proceedings of the 1st Joint Workshop on Spoken Language Technologies for Under-resourced languages (SLTU) and Collaboration and Computing for Under-Resourced Languages (CCURL), pages 111–120, Marseille, France. European Language Resources association. find
  36. Anne Murphy. 2018. Writing Punjabi across borders. South Asian history and culture, 9(1):68– 91. find
  37. Guilherme Penedo, Hynek Kydlíček, Vinko Sabolčec, Bettina Messmer, Negar Foroutan, Amir Hossein Kargaran, Colin Raffel, Martin Jaggi, Leandro Von Werra, and Thomas Wolf. 2025. FineWeb2: One pipeline to scale them all — adapting pre-training data processing to every language. https: arXiv preprint arXiv:2506.20920. //github.com/huggingface/fineweb-2 . ↗
  38. Alan Ramponi. 2024. Language varieties of Italy: Technology challenges and opportunities. Transactions of the Association for Computational Linguistics, 12:19–38. find
  39. Brian Roark, Lawrence Wolf-Sonkin, Christo Kirov, Sabrina J. Mielke, Cibu Johny, Isin Demirsahin, and Keith Hall. 2020. Processing South Asian languages written in the Latin script: the Dakshina dataset. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 2413–2423, Marseille, France. European Language Resources Association. find
  40. Sebastian Ruder, Jonathan H. Clark, Alexander Gutkin, Mihir Kale, Min Ma, Massimo Nicosia, Shruti Rijhwani, Parker Riley, Jean-Michel A Sarr, Xinyi Wang, John Wieting, Nitish Gupta, Anna Katanova, Christo Kirov, Dana L. Dickinson, Brian Roark, Bidisha Samanta, Connie Tao, David I. Adelani, Vera Axelrod, Isaac Caswell, Colin Cherry, Dan Garrette, Reeve Ingle, Melvin Johnson, Dmitry Panteleev, and Partha Talukdar. 2023. XTREME-UP: A user-centric scarce-data benchmark for under-represented languages. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1856–1884, Singapore. Association for Computational Linguistics. find
  41. Satyendra Kumar Sarmah. 2014. Script movement among the Bodo of Assam. Proceedings of the Indian History Congress, 75:1335–1340. find
  42. Hedwig G. Sekeres, Wilbert Heeringa, Wietse de Vries, Oscar Yde Zwagers, Martijn Wieling, and Goffe Th. Jensma. 2024. Developing infrastructure for low-resource language corpus building. In Proceedings of the 3rd Annual Meeting of the Special Interest Group on Under-resourced Languages @ LREC-COLING 2024, pages 72– 78, Torino, Italia. ELRA and ICCL. find
  43. Yingli Shen, Wen Lai, Shuo Wang, Xueren Zhang, Kangyang Luo, Alexander Fraser, and Maosong Sun. 2025. DCAD-2000: A multilingual dataset across 2000+ languages with data cleaning as anomaly detection. arXiv preprint arXiv:2502.11546. https://huggingface.co/datasets/openbmb/DCAD-2000 . ↗
  44. Gary F. Simons, Abbey L. L. Thomas, and Chad K. K. White. 2022. Assessing digital language support on a global scale. In Proceedings of the 29th International Conference on Computational Linguistics, pages 4299–4305, Gyeongju, Republic of Korea. International Committee on Computational Linguistics. find
  45. Atnafu Lambebo Tonja, Israel Abebe Azime, Tadesse Destaw Belay, Mesay Gemeda Yigezu, Moges Ahmed Ah Mehamed, Abinew Ali Ayele, Ebrahim Chekol Jibril, Michael Melese Woldeyohannis, Olga Kolesnikova, Philipp Slusallek, Dietrich Klakow, and Seid Muhie Yimam. 2024. EthioLLM: Multilingual large language models for Ethiopian languages with task evaluation. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 6341–6352, Torino, Italia. ELRA and ICCL. find
  46. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483–498, Online. Association for Computational Linguistics. ↗ our card
  47. Daniel Yaqob. 1997. Transliteration on the Internet: the case of Ethiopic. In Proceedings of the International Symposium on Multilingual Information Processing, Tsukuba, Japan. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/benton-2026-mining-naturally-romanized-seed-corpora/