Elamur Research⌕
☰
home / papers / 2026 / Robust Language Identification for Romansh Varieties

Robust Language Identification for Romansh Varieties

Paper · 2026 · Charlotte Model, Sina Ahmadi and Jannis Vamvas

Summary

Romansh, spoken by roughly 40,000 people in the Swiss canton of Grisons, exists as five regional idioms (Sursilvan, Sutsilvan, Surmiran, Puter, Vallader) plus the supra-regional written standard Rumantsch Grischun (RG). Existing language identification tools either skip Romansh or treat it as one language. Current dictionary and spell-checking services also make users pick the idiom manually. The authors build what they describe as the first documented LID system that tells all six varieties apart. They combine five sources: the Pledari Grond dictionary, La Quotidiana news, RTR broadcast transcripts, RTR Telesguard newsroom notes and Mediomatix textbooks. These are cleaned, deduplicated and optionally masked for named entities using a SwissBERT-based NER model. They then compare linear scikit-learn classifiers on TF–IDF character and word n-grams. A tuned linear SVM performs best, averaging about 97% accuracy on in-domain test sets, though results drop clearly on out-of-domain data. The released classifier can route texts to idiom-specific tools such as spell checkers or machine translation. It also offers a template for other low-resource regional varieties.

Key points

  • Data and splits: The corpus has 487,172 samples, about 12.4M tokens. Training uses PG, LQ, RTR and TB. Evaluation uses four test sets: three in-domain (LQ, RTR, TB) and one out-of-domain (TG newsroom notes).
  • Model comparison: The best SVM uses ℓ1 penalty, character 1–4-grams and word unigrams, reaching 97.1 macro F1 on the dev set. Logistic Regression reaches 96.8. Before tuning, the baselines scored only about 69–77 macro F1.
  • Scores across test sets: Macro F1 reaches up to 98 on balanced in-domain data but falls to 69.1 on out-of-domain test-d.
  • Named-entity masking: Masking has almost no effect on scores (≤0.1 F1 on most sets, 0.7 at most). This suggests the model relies on linguistic cues rather than memorised proper nouns.
  • Discriminative features: Character n-grams carry the most signal. Two of the top features are artefacts that may not transfer to ordinary text:
    • underdotted stress markers in Puter/Vallader dictionary entries;
    • whitespace for RG, caused by class imbalance (≈171K RG samples versus 44K–86K for each other idiom).
  • Availability: The code is at github.com/ZurichNLP/romansh-lid.

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

🎓University of ZurichZurich · Switzerland80 works · 12,260 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 1Research subjects and languages the work is about.Idioms & multiword expressions · 222 works

Cite

Cite as Charlotte Model, Sina Ahmadi and Jannis Vamvas. 2026. Robust Language Identification for Romansh Varieties. Proceedings of the 11th Edition of the Swiss Text Analytics Conference.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{model2026robust,
  title = {{Robust Language Identification for Romansh Varieties}},
  author = {Charlotte Model, Sina Ahmadi and Jannis Vamvas},
  booktitle = {Proceedings of the 11th Edition of the Swiss Text Analytics Conference},
  year = {2026},
  url = {https://aclanthology.org/2026.swisstext-1.9/}
}

References

14 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Noëmi Aepli, Antonios Anastasopoulos, Adrian-Gabriel Chifu, William Domingues, Fahim Faisal, Mihaela Gaman, Radu Tudor Ionescu, and Yves Scherrer. 2022. Findings of the VarDial evaluation campaign 2022. In Proceedings of the Ninth Workshop on NLP for Similar Languages, Varieties and Dialects, pages… find
  2. Stephen R. Anderson. 2016. Romansh (rumantsch). In Adam Ledgeway and Martin Maiden, editors, The Oxford Guide to the Romance Languages, pages 169– 184. Oxford University Press, Oxford. find
  3. Gabriel Bernier-colborne, Cyril Goutte, and Serge Leger. 2023. Dialect and variant identification as a multi-label classification task: A proposal based on near-duplicate analysis. In Tenth Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial 2023), pages 142–151, Dubrovnik, Croati… find
  4. Negar Foroutan, Jakhongir Saydaliev, Grace Kim, and Antoine Bosselut. 2026. ConLID: Supervised contrastive learning for low-resource language identification. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages… find
  5. Matthias Grünert. 2024. Rätoromanisch. In Elvira Glaser, Johannes Kabatek, and Barbara Sonnenhauser, editors, Sprachenräume der Schweiz. Band 1: Sprachen, pages 156–184. Narr Francke Attempto, Tübingen. find
  6. Alejandro Hernández-Cano, Alexander Hägele, Allen Hao Huang, Angelika Romanou, Antoni-Joan Solergibert, Barna Pasztor, Bettina Messmer, Dhia Garbaya, Eduard Frank Ďurech, Ido Hakimi, Juan García Giraldo, Mete Ismayilzada, Negar Foroutan, Skander Moalla, Tiancheng Chen, Vinko Sabolčec, Yixuan Xu, M… — Apertus: Democratizing Open and Compliant LLMs for Global Language Environments
  7. Laurie Burchell, Alexandra Birch, Robert P. Thompson, and Kenneth Heafield. 2024. Code-switched language identification is harder than you think. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2024 - Volume 1: Long Papers, St. Jul… find
  8. Zachary Hopton, Jannis Vamvas, Andrin Büchler, Anna Rutkiewicz, Rico Cathomas, and Rico Sennrich. 2026. The mediomatix corpus: Parallel data for Romansh language varieties via comparable schoolbooks. In Findings of the Association for Computational Linguistics: EACL 2026, pages 290–306, Rabat, Moroc… find
  9. Bharathi Raja Chakravarthi, Mihaela Gaman, Radu Tudor Ionescu, Heidi Jauhiainen, Tommi Jauhiainen, Krister Lindén, Nikola Ljubešić, Niko Partanen, Ruba Priyadharshini, Christoph Purschke, Eswari Rajagopal, Yves Scherrer, and Marcos Zampieri. 2021. Findings of the VarDial evaluation campaign 2021. I… find
  10. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:282… find
  11. Elizabeth Salesky, Badr M. Abdullah, Sabrina Mielke, Elena Klyachko, Oleg Serikov, Edoardo Maria Ponti, Ritesh Kumar, Ryan Cotterell, and Ekaterina Vylomova. 2021. find
  12. SIGTYP 2021 shared task: Robust spoken language identification. In Proceedings of the Third Workshop on Computational Typology and Multilingual NLP, pages 122–129, Online. Association for Computational Linguistics. find
  13. Jannis Vamvas, Johannes Graën, and Rico Sennrich. 2023. SwissBERT: The multilingual language model for Switzerland. In Proceedings of the 8th edition of the Swiss Text Analytics Conference, pages 54–69, Neuchatel, Switzerland. Association for Computational Linguistics. find
  14. Jannis Vamvas, Ignacio Pérez Prat, Not Soliva, Sandra Baltermia-Guetg, Andrina Beeli, Simona Beeli, Madlaina Capeder, Laura Decurtins, Gian Peder Gregori, Flavia Hobi, Gabriela Holderegger, Arina Lazzarini, Viviana Lazzarini, Walter Rosselli, Bettina Vital, Anna Rutkiewicz, and Rico Sennrich. 2025.… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/model-2026-robust-language-identification-romansh-varieties/