Elamur Research⌕
☰
home / papers / 2026 / Reasoning as Supportive Context for Machine Translation: A Case Study on Hindi to Bengali Language Pair

Reasoning as Supportive Context for Machine Translation: A Case Study on Hindi to Bengali Language Pair

Summary

The paper examines whether structured reasoning helps a small LLM translate Hindi into Bengali when the reasoning is supplied as auxiliary input rather than produced as output. The two languages are related but differ in script, morphology and domain usage. The authors built a corpus of eight domains with 36,040 training pairs and 2,000 test pairs. For each pair, Llama-4-Scout-17B-16E-Instruct generated five annotation types (Key Terms, Syntactic, Semantic, Pragmatic, Paraphrase) through an English pivot, and Google Translate carried them back into Hindi and Bengali. They fine-tuned Gemma-3-1B-Instruct on all 31 non-empty component combinations and scored the outputs with BLEU, chrF and TER.

Combining many heterogeneous signals hurt quality, an effect the authors call objective diffusion. The short Semantic+Paraphrase pairing worked best. With full fine-tuning, supplying that context at test time raised BLEU from 22.12 to 23.86. The study shows that the type of supporting context and consistent use between training and inference matter more than its amount, especially for compact MT models.

Key points

  • Data: 36,040 Hindi–Bengali training pairs and a 2,000-sentence test set covering eight domains.
    • Each domain has 200 test sentences, except Education with 600.
    • The corpus spans Judiciary, Education, Healthcare, Climate, Agriculture, Governance, Tourism, and Science and Technology.
  • Annotation length varies widely: Key Terms and Syntactic annotations average over 80 tokens per instance, while Semantic and Paraphrase average about 45 and 24 tokens.
  • Ablation design: the 31 reasoning configurations were trained with LoRA. Larger mixes, including all five components together, scored below compact subsets such as R3+R5.
  • Training-inference alignment: under full fine-tuning, the baseline reached 22.12 BLEU.
    • Training with R3+R5 but testing without it gave 22.26 BLEU.
    • Providing R3+R5 at inference ("guided inference") gave 23.86 BLEU and 57.57 chrF, with TER also improving.
  • Domain-wise BLEU for guided inference vs. the baseline:
    • Largest gains: Science & Technology (24.85→30.50), Governance (17.44→20.32) and Tourism (12.87→15.32).
    • Climate dropped slightly (21.95→21.16).

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 · Machine translation

Topics: Machine translation

Cite

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

@inproceedings{singh2026reasoning,
  title = {{Reasoning as Supportive Context for Machine Translation: A Case Study on Hindi to Bengali Language Pair}},
  author = {Singh, Kshetrimayum Boynao and Singh, Saksham and Pakray, Partha and Ekbal, Asif},
  booktitle = {Proceedings of the 26th Annual Conference of the European Association for Machine Translation (Volume 2)},
  year = {2026},
  doi = {10.13140/rg.2.2.23116.37760},
  url = {https://doi.org/10.13140/rg.2.2.23116.37760}
}

References

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

  1. Bala Das, Sudhansu, Atharv Biradar, Tapas Kumar Mishra, and Bidyut Kr. Patra. 2023. Improving multilingual neural machine translation system for indic languages. ACM Trans. Asian Low-Resour. Lang. Inf. Process., 22(6), June. find
  2. Chen, Huadong, Shujian Huang, David Chiang, and Jiajun Chen. 2017. Improved neural machine translation with a syntax-aware encoder and decoder. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1936–1945, Vancouver, Canada, July… find our card
  3. Dabre, Raj, Bianka Buschbeck, Miriam Exel, and Hideki Tanaka. 2023. A study on the effectiveness of large language models for translation with markup. In Utiyama, Masao and Rui Wang, editors, Proceedings of Machine Translation Summit XIX, Vol. 1: Research Track, pages 148–159, Macau SAR, China, Sept… find
  4. Fan, Yuchun, Yongyu Mu, YiLin Wang, Lei Huang, Junhao Ruan, Bei Li, Tong Xiao, Shujian Huang, Xiaocheng Feng, and Jingbo Zhu. 2025. SLAM: Towards efficient multilingual reasoning via selective language alignment. In Rambow, Owen, Leo Wanner, Marianna Apidianaki, Hend Al-Khalifa, Barbara Di Eugenio,… find
  5. Gao, Hui, Jing Zhang, Peng Zhang, and Chang Yang. 2025. Consistency rating of semantic transparency: an evaluation method for metaphor competence in idiom understanding tasks. In Rambow, Owen, Leo Wanner, Marianna Apidianaki, Hend Al-Khalifa, Barbara Di Eugenio, and Steven Schockaert, editors, Proce… find
  6. Gaur, Vedant and Nikunj Saunshi. 2023. Reasoning in large language models through symbolic math word problems. find
  7. Ghazaryan, Gayane, Erik Arakelyan, Isabelle Augenstein, and Pasquale Minervini. 2025. SynDARin: Synthesising datasets for automated reasoning in low-resource languages. In Proceedings of the 31st International Conference on Computational Linguistics, pages 6459–6466, Abu Dhabi, UAE, January. Associa… find
  8. He, Jie, Tao Wang, Deyi Xiong, and Qun Liu. 2020. The box is in the pen: Evaluating commonsense reasoning in neural machine translation. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3662–3672, Online, November. Association for Computational Linguistics. find
  9. He, Zhiwei, Tian Liang, Wenxiang Jiao, Zhuosheng Zhang, Yujiu Yang, Rui Wang, Zhaopeng Tu, Shuming Shi, and Xing Wang. 2024. Exploring human-like translation strategy with large language models. Transactions of the Association for Computational Linguistics, 12:229–246. find our card
  10. Jiang, Gangwei, Yahui Liu, Zhaoyi Li, Wei Bi, Fuzheng Zhang, Linqi Song, Ying Wei, and Defu Lian. 2025. What makes a good reasoning chain? uncovering structural patterns in long chain-of-thought reasoning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pag… find
  11. Kingma, Diederik P. and Jimmy Ba. 2017. Adam: A method for stochastic optimization. find
  12. Lee, Minjae, Youngbin Noh, and Seung Jin Lee. 2025. A testset for context-aware LLM translation in Korean-to-English discourse level translation. In Proceedings of the 31st International Conference on Computational Linguistics, pages 1632–1646, Abu Dhabi, UAE, January. Association for Computational… find
  13. Liu, Xuebo, Yutong Wang, Derek F. Wong, Runzhe Zhan, Liangxuan Yu, and Min Zhang. 2023. Revisiting commonsense reasoning in machine translation: Training, evaluation and challenge. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pag… find
  14. Nguyen, Lam and Yang Xu. 2025. Reasoning for translation: Comparative analysis of chain-of-thought and tree-of-thought prompting for LLM translation. In Zhao, Jin, Mingyang Wang, and Zhu Liu, editors, Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 4:… find
  15. Papineni, Kishore, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318. find our card
  16. Singh, Kshetrimayum Boynao, Ningthoujam Avichandra Singh, Loitongbam Sanayai Meetei, Ningthoujam Justwant Singh, Thoudam Doren Singh, and Sivaji Bandyopadhyay. 2023. A comparative study of transformer and transfer learning MT models for English-Manipuri. In Proceedings of the 20th International Conf… find
  17. Singh, Kshetrimayum Boynao, Asif Ekbal, and Partha Pakray. 2025a. Evaluating IndicTrans2 and ByT5 for English–Santali machine translation using the ol chiki script. In Proceedings of the 1st Workshop on Multimodal Models for Low-Resource Contexts and Social Impact (MMLoSo 2025), pages 95–100, Mumbai… find
  18. Singh, Ningthoujam Avichandra, Boynao Kshetrimayum, Ningthoujam Justwant Singh, Thoudam Doren Singh, and Shilpa Sharma. 2025b. Quantifying the impact of data scale and quality on neural machine translation for low-resource language pair. In 2025 OITS International Conference on Information Technolog… find
  19. Singh, Kshetrimayum Boynao, Soham Bhattacharjee, Baban Gain, Asif Ekbal, and Partha Pakray. 2026a. A comparative study of fine-tuned mbart and indictrans2 for english-hindi legal machine translation. TechRxiv, 2026(0102). find
  20. Singh, Kshetrimayum Boynao, Soham Bhattacharjee, Baban Gain, Asif Ekbal, and Partha Pakray. 2026b. A comparative study of fine-tuned mbart and indictrans2 for english-hindi legal machine translation. TechRxiv, 2026(0102). find
  21. Verma, Sshubam, Mohammed Safi Ur Rahman Khan, Vishwajeet Kumar, Rudra Murthy, and Jaydeep Sen. 2025. MILU: A multi-task Indic language understanding benchmark. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Languag… find
  22. Wang, Jiaan, Fandong Meng, Yunlong Liang, and Jie Zhou. 2025. DRT: Deep reasoning translation via long chain-of-thought. In Findings of the Association for Computational Linguistics: ACL 2025, pages 6770–6782, Vienna, Austria, July. Association for Computational Linguistics. find
  23. Yan, Shen, Leonard Dahlmann, Pavel Petrushkov, Sanjika Hewavitharana, and Shahram Khadivi. 2018. Word-based domain adaptation for neural machine translation. In Proceedings of the 15th International Conference on Spoken Language Translation, pages 31–38, Brussels, October 29-30. International Confer… find
  24. Yao, Shunyu, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. find
  25. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809–11822. find
  26. Zhu, Wenhao, Hongyi Liu, Qingxiu Dong, Jingjing Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. 2024. Multilingual machine translation with large language models: Empirical results and analysis. In Duh, Kevin, Helena Gomez, and Steven Bethard, editors, Findings of the Association for Comp… find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/singh-2026-reasoning-as-supportive-context-machine/