Elamur Research⌕
☰
home / papers / 2026 / Aligned Parallel Corpus of the Vedic Saṁhitās for Machine Translation

Aligned Parallel Corpus of the Vedic Saṁhitās for Machine Translation

Summary

Vedic Sanskrit, the oldest layer of Indo-Aryan, has long had respected scholarly translations. However, these were made for philologists, not for machine learning, so usable source–target pairs for training and testing MT systems have been lacking. Most Sanskrit MT work also targets Classical rather than Vedic language. The authors fill this gap with a parallel corpus covering three Saṁhitās: the Ṛgveda, the Śaunaka Atharvaveda, and the Taittirīya Saṁhitā. Each is paired with a century-old public-domain translation. The Sanskrit comes from digital editions such as TITUS and VedaWeb and is normalized to ISO 15919. Every source segment is linked to a single verse or, for TS prose, a single paragraph. With this data they fine-tune GPT-4.1 nano, Gemini 2.5 Flash, and Mitra, then score the outputs with sacreBLEU and COMET. Fine-tuning reliably helps when training and test data come from the same text. Performance drops sharply when a model is tested on a Saṁhitā it never saw, which reveals strong stylistic and lexical differences among the three. The released dataset gives low-resource historical MT a reproducible benchmark and argues for domain-aware evaluation.

Key points

  • Translation pairings: Geldner (1951, German) for RV, Whitney & Lanman (1905, English) for AVŚ, and Keith (1914, English) for TS. The authors present this as the first machine-readable, aligned Vedic Saṁhitā translation corpus, published on Hugging Face with standardized splits.
  • Models tested: GPT-4.1 nano, Gemini 2.5 Flash, and Mitra. All received one shared zero-shot prompt template, with no few-shot examples or retrieval, and output postprocessing was kept minimal.
  • Evaluation pairs a surface metric (case-insensitive sacreBLEU) with a meaning-based one (COMET from Unbabel), so both form and semantics are assessed.
  • For a reduced experimental setting, corpora larger than 2,000 examples were randomly sampled down to 2,000, then divided 70%/15%/15% into train, dev, and test.
  • In the qualitative examples, Gemini usually comes closest to the references; for instance, it keeps negation in AVŚ 5.30.6 and reaches COMET 0.879 on RV 5.82.4. GPT and Mitra sometimes drift toward unrelated passages.

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

Topics: Machine translation

Cite

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

@inproceedings{tsukagoshi2026aligned,
  title = {{Aligned Parallel Corpus of the Vedic Saṁhitās for Machine Translation}},
  author = {Tsukagoshi, Yuzuki and Ohmukai, Ikki},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/3faoikwftvwt},
  url = {https://doi.org/10.63317/3faoikwftvwt}
}

References

35 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Rahul Aralikatte, Miryam de Lhoneux, Anoop Kunchukuttan, and Anders Søgaard. 2021. Itihasa: A large-scale corpus for Sanskrit to English translation. In Proceedings of the 8th Workshop on Asian Translation (WAT2021), pages 191–197, Online. Association for Computational Linguistics. find
  2. Theodor Aufrecht. 1877. Die Hymnen des Rigveda. Bonn: Adolph Marcus. find
  3. Makoto Fushimi. 1997. Taittirīya-saṃhitā. Based on Albrecht Weber’s edition (Die Taittirīya-Saṃhitā, Leipzig: Brockhaus, 1871–1872, Indische Studien 11–12). Edited by Makoto Fushimi, Ōsaka; TITUS version prepared by Jost Gippert, Frankfurt a.M., updates 1997–2012. find
  4. Karl F. Geldner. 1951. Der Rig-Veda : aus dem Sanskrit ins Deutsche übersetzt und mit einem laufenden Kommentar versehen. Number v. 33-36 in Harvard oriental series. Harvard University Press, Oxford University Press, Otto Harrassowitz, Cambridge, Mass., London, Leipzig. find
  5. Reinhold Grünendahl. 2020. GRETIL: Göttingen Register of Electronic Texts in In-https://gretil.sub. dian Languages. uni-goettingen.de/gretil.html. find
  6. Oliver Hellwig. 2010–2021. Dcs - The Digital Corpus of Sanskrit. find
  7. Oliver Hellwig, Salvatore Scarlata, Elia Ackermann, and Paul Widmer. 2020. The treebank of vedic Sanskrit. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 5137–5146, Marseille, France. European Language Resources Association. find
  8. Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations. find
  9. Gérard Huet. 2026. The sanskrit heritage engine reference manual. https://sanskrit.inria.fr/manual.html. Written on February 10th 2026 for Sanskrit Engine Version 3.76. ↗
  10. N. Biraja Isac and Himansu Das. 2025. SLIP: A Sanskrit Linguistic Intelligence Pipeline for Enhanced Neural Machine Translation of Classical Texts. IEEE Access, PP(99):1–1. License: CC BY 4.0. find
  11. Jost Gippert, Javier Martínez, Agnes Korn, and Roland Mittmann. 2016. Thesaurus Indogermanischer Text- und Sprachmaterialien. find
  12. Arthur Berriedale Keith. 1914. The Veda of the Black Yajus School, entitled Taittiriya Sanhita, translated from the original Sanskrit prose and verse. Number 18-19 in Harvard Oriental Series. Harvard University Press, Cambridge, Massachusetts. find
  13. Jeong-Soo Kim. 2022. Atharvavedasaṃhitā der Śaunakaśākhā : Eine neue Edition unter besonderer Berücksichtigung der Parallelstellen der Paippalādasaṃhitā. Universität Würzburg. find
  14. Nimrita Koul and Sunilkumar S. Manvi. 2021. A proposed model for neural machine translation of Sanskrit into English. International Journal of Information Technology, 13(1):375–381. find
  15. Daniel Kölligan, Claes Neuefeind, Uta Reinöhl, Patrick Sahle, Antje Casaretto, Anna Fischer, Börge Kiss, Natalie Korobzow, Jürgen Rolshoven, Jakob Halfmann, and Francisco Mondaca. 2024. VedaWeb. Online Research Platform for Old Indic Texts. find
  16. Ayush Maheshwari, Nikhil Singh, Amrith Krishna, and Ganesh Ramakrishnan. 2022. A benchmark and dataset for post-OCR text correction in Sanskrit. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 6258–6265, Abu Dhabi, United Arab Emirates. Association for Computational L… find
  17. Vimal Mishra and R. B. Mishra. 2009. Ann and rule based model for english to sanskrit machine translation. INFOCOMP Journal of Computer Science, 9(1):80–89. find
  18. Vandan Mujadia and Dipti Misra Sharma. 2025. Bhashaverse : Translation ecosystem for indian subcontinent languages. find
  19. Sebastian Nehrdich, Oliver Hellwig, and Kurt Keutzer. 2024. One model is all you need: ByT5-Sanskrit, a unified model for Sanskrit NLP tasks. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 13742–13751, Miami, Florida, USA. Association for Computational Linguistics. find
  20. Chatia Orlandi. 1991. Gli inni dell’Atharvaveda (Śaunaka): traslitterazione. Giardini, Pisa. Transliteration of the Atharvaveda (Śaunaka) recension. find
  21. Mrinal Pandey, Rashmikiran Pandey, and Alexey Nazarov. 2022. Machine translation of vedic sanskrit using deep learning algorithm. In 2022 4th International Conference on Advances in Computing, Communication Control and Networking (ICAC3N), pages 1477–1480. find
  22. Vladimir Petr and Petr Vavroušek. 1996. Atharvaveda saṃhitā (Śaunaka recension). Based on the editions by Chatia Orlandi (Pisa 1991) and Rudolf Roth & William D. Whitney (Berlin 1856). Entered by Vladimir Petr and Petr Vavroušek, Praha 1996; TITUS version prepared by Jost Gippert, Frankfurt a.M., up… find
  23. Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186–191, Belgium, Brussels. Association for Computational Linguistics. find our card
  24. Akanksha Shukla Pragya Shukla. 2014. English speech to sanskrit speech (esss) using rule based translation. International Journal of Computer Applications, 92(10):37–42. find
  25. Ravneet Punia, Aditya Sharma, Sarthak Pruthi, and Minni Jain. 2020. Improving neural machine translation for Sanskrit-English. In Proceedings of the 17th International Conference on Natural Language Processing (ICON), pages 234–238, Indian Institute of Technology Patna, Patna, India. NLP Association… find
  26. Sarita G. Rathod and Shanta Sondur. 2012. English to sanskrit translator and synthesizer (etsts). International Journal of Emerging Technology and Advanced Engineering, 2(12):379–382. find
  27. Jaideepsinh K. Raulji, Jatinderkumar R. Saini, Kaushika Pal, and Ketan Kotecha. 2022. A novel framework for sanskrit-gujarati symbolic machine translation system. International Journal of Advanced Computer Science and Applications, 13(4). find
  28. Ricardo Rei, José G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and André F. T. Martins. 2022. COMET-22: Unbabel-IST 2022 submission for the metrics shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages… find our card
  29. Rudolf Roth and William Dwight Whitney. 1855. Atharva Veda Sanhita. Friedrich Dümmlers Buchhandlung, Berlin. find
  30. Prakash R. Devale Sandeep R. Warhade, Suhas H. Patil. 2012. English-to-sanskrit statistical machine translation with ubiquitous application. International Journal of Computer Applications, 51(1):41–46. find
  31. Ulrich Stiehl. 1999-2025. Sanskrit-Web. find
  32. TEI Consortium. 2025. TEI P5: Guidelines for Electronic Text Encoding and Interchange. find
  33. Albrecht Weber. 1871. Die Taittirīya-Saṃhitā. F. A. Brockhaus, Leipzig. Published in Indische Studien, Vols. 11–12 (1871–1872). find
  34. William Dwight Whitney and Charles Rockwell Lanman. 1905. Atharva-veda Saṁhitā / translated with a critical and exegetical commentary, by William Dwight Whitney; revised and brought nearer to completion and edited by Charles Rockwell Lanman. Number 2 v. in Harvard oriental series ;7-8. Harvard Univ… find
  35. Dominik Wujastyk, Birgit Kellner, and Sheldon Pollock. 2015. SARIT: Search and Retrieval of Indic Texts. TEI-encoded digital library of Sanskrit and Indic texts. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tsukagoshi-2026-aligned-parallel-corpus-vedic-sa/