Elamur Research⌕
☰
home / papers / 2026 / A Joint Detection Framework for Latvian Loanwords and Calques Using Monolingual Data

A Joint Detection Framework for Latvian Loanwords and Calques Using Monolingual Data

Summary

Most automatic borrowing detection relies on bilingual resources and handles only direct loanwords that resemble their source in sound or spelling. Calques, which copy a foreign structure or meaning using native material, have been largely ignored. The authors tackle both loanwords and calques in Latvian using only monolingual data. With help from Latvian linguists and native speakers, they manually labeled over 3,000 words and phrases sampled from a Latvian Wikipedia dump as loanword, calque, or local word. For practical reasons, the calque class was widened to include hybrid loan+native compounds. The method fine-tunes mBERT in two binary stages, following the language-contact view that borrowing ranges from surface to deep. Loanwords, which look visibly foreign, are separated first. Calques are then distinguished from native vocabulary. Pseudo-labeling supplies extra semi-supervised training data. On an external test set, the semi-supervised joint pipeline reaches macro-F1 0.854, outperforming a direct three-class classifier and zero-shot GPT-4o. This is the first benchmark for Latvian borrowing detection and a reusable approach for other resource-poor languages.

Key points

  • Data: The source is the Latvian Wikipedia dump of 02-Feb-2025, cleaned of markup and non-Latvian content. The training data were split into stratified 25/50/75/100% subsets. The annotated dataset is on Zenodo (doi:10.5281/zenodo.17380901), and the code is on GitHub.
  • External test set: It contains 471 loanwords, 306 calques and 465 local words.
  • Main results (macro-F1, full data):
    • Joint binary semi-supervised: 0.854 (accuracy 0.857).
    • Joint binary supervised: 0.828.
    • Direct 3-way: 0.753 supervised and 0.792 semi-supervised.
    • GPT-4o zero-shot: 0.820.
  • Pseudo-to-real label ratios: Ratios of 1:1, 2:1 and 3:1 were tested. With 75% of the data, macro-F1 was 0.821 supervised, 0.829 at 2:1 and 0.832 at 3:1. The gain was small at 50% (0.799 → 0.800).
  • Label scheme: The scheme is explicitly a computational operationalization, not a revision of linguistic typology. Long-established historical borrowings are counted as local words.

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 10.63317/2f6e35y4dgkn

@inproceedings{zhang2026joint,
  title = {{A Joint Detection Framework for Latvian Loanwords and Calques Using Monolingual Data}},
  author = {Zhang, Yelingyun and Kapenieks, Atis and Platonova, Marina},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/2f6e35y4dgkn},
  url = {https://doi.org/10.63317/2f6e35y4dgkn}
}

References

20 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Felermino Dario Mario Ali, Henrique Lopes Cardoso, and Rui Sousa-Silva. 2024. Detecting Loanwords in Emakhuwa: An Extremely Low-Resource Bantu Language Exhibiting Significant Borrowing from Portuguese. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language R… find
  2. Mateusz Buda, Atsuto Maki, and Maciej A. Mazurowski. 2018. A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106:249–259. find
  3. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,… find
  4. Johann-Mattis List. 2019. Automated methods for the investigation of language contact, with a focus on lexical borrowing. Language and Linguistics Compass, 13(10):e12355. E12355 LNCO-0781.R2. find
  5. Yaron Matras. 2009. Language Contact. Cambridge Textbooks in Linguistics. Cambridge University Press, Cambridge. find
  6. Tamara Matthews and David Lillis. 2022. Experimenting with ensembles of pre-trained language models for classification of custom legal datasets. In Proceedings of the 5th International Conference on Natural Language and Speech Processing (ICNLSP 2022), pages 68–78, Trento, Italy. Association for Com… find
  7. Chenggang Mi. 2023. Loanword identification based on web resources: A case study on wikipedia. Computer Speech & Language, 81:101517. find
  8. Chenggang Mi, Lei Xie, and Yanning Zhang. 2020. Loanword Identification in Low-Resource Languages with Minimal Supervision. ACM Trans. Asian Low-Resour. Lang. Inf. Process., 19(3):43:1–43:22. find
  9. Abhijnan Nath, Sina Mahdipour Saravani, Ibrahim Khebour, Sheikh Mannan, Zihui Li, and Nikhil Krishnaswamy. 2022. A Generalized Method for Automated Multilingual Loanword Detection. In Proceedings of the 29th International Conference on Computational Linguistics, pages 4996–5013, Gyeongju, Republic o… find
  10. Peteris Paikens, Lauma Pretkalnin , a, and Laura Rituma. 2024. A Computational Model of Latvian Morphology. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 221–232, Torino, Italia. ELRA and ICCL. find
  11. Penelope Sewell. 2001. The occurrence of calque in translation scripts. Meta, 46(3):607–615. find
  12. Yulia Spektor. 2021. Detection and Morphological Analysis of Novel Russian Loanwords. Thesis, City University of New York. find
  13. Andrejs Veisbergs. 2017. Translation language: The major force in shaping modern latvian. Vertimo studijos, 2:54. find
  14. Andrejs Veisbergs. 2018. Word-Formation Pattern Borrowing in Latvian. Baltic Journal of English Language, Literature and Culture, 8:129–146. find
  15. Uriel Weinreich. 2010. Languages in Contact: Findings and Problems. De Gruyter Mouton. find
  16. Liqin Zhang, Ray Fabri, John Nerbonne, and John Nerbonne. 2021. Detecting loan words computationally. In Enoch O. Aboh and Cécile B. Vigouroux, editors, Variation Rolls the Dice, pages 269–288. John Benjamins Publishing Company. 13. Language Resource References Latvijas Universitātes Literatūras,… find
  17. Latvijas Zināt¸nu akadēmijas Terminolo¸gijas komisija. 2025. Latvijas nacionālais terminolo¸gijas portāls. https://termini.gov.lv/. Accessed: 2025-10-17. ↗
  18. Tilde. 2025. Letonika.lv. https://www.letonika.lv/. Accessed: 2025-10-17. ↗
  19. Wikimedia Foundation. 2025. Latvian wikipedia dump (02-feb-2025). https://dumps.wikimedia.org/lvwiki/20250202/. Data dump generated 02-Feb-2025. The live server index for this date has since been rotated. ↗
  20. Yelingyun Zhang, Atis Kapenieks, and Marina Platonova. 2025. Annotated latvian borrowing dataset. https://doi.org/10.5281/zenodo.17380901. Dataset created for this study. — [doi:10.5281/zenodo.17380901](https://doi.org/10.5281/zenodo.17380901) ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2026-joint-detection-framework-latvian-loanwords/