This repository contains the implementation, experiments, and resources for the research paper:
Reasoning as Supportive Context for Machine Translation: A Case Study on Hindi to Bengali Language Pair
This work investigates how structured reasoning can improve machine translation by acting as supportive contextual guidance during translation generation.
Overall framework of reasoning-guided Hindi → Bengali machine translation.
The framework introduces five reasoning categories:
| ID | Reasoning Type |
|---|---|
| R1 | Key Terms |
| R2 | Syntactic |
| R3 | Semantic |
| R4 | Pragmatic |
| R5 | Paraphrase |
The study performs exhaustive experiments across all:
31 reasoning combinations
- Increasing reasoning signals does not always improve translation
- Translation quality depends on reasoning composition
- Excessive reasoning may introduce noisy contextual signals
- The best-performing combination is:
Semantic (R3) + Paraphrase (R5)
- Guided inference significantly improves translation quality
Reasoning is not generated as the final output.
Instead, reasoning acts as:
Supportive context to guide machine translation
- 36,040 parallel sentence pairs
- 2,000 sentence pairs
- Agriculture
- Tourism
- Governance
- Climate
- Healthcare
- Science & Technology
- Judiciary
- Education
- Instruction-tuned base model
- LoRA fine-tuning
- Structured reasoning prompts
- Full reasoning ablation study
- Zero-Reasoning Inference
- Guided Inference using optimal reasoning subset
The table below shows the BLEU scores obtained across different domains for multiple reasoning configurations and fine-tuning setups.
Performance comparison of LoRA and full fine-tuning models across different domains.
Key observations from the experiments:
- Among LoRA-based models, R3 + R5 achieved the best overall performance
- Using all five reasoning signals together reduced translation quality
- Full fine-tuning significantly outperformed LoRA configurations
- Guided inference consistently improved results over normal inference
- The best overall system was:
Full SFT: R3 + R5 (Guided Inference) with an overall BLEU score of 23.86
If you use this work, please cite:
@article{EAMT26-IITP,
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},
publisher={European Association for Machine Translation (EAMT)}
}This work contributes toward reasoning-enhanced machine translation for low-resource Indian languages and multilingual NLP research.