Back-Translation Still Matters: Leveraging Monolingual Data for Low-Resource Machine Translation with LLMs
arXiv cs.CL · 14 Oct 2026, 03:50 AEST
Back-translation proves to be a significant technique for improving machine translation in low-resource languages. It outperforms continued pretraining and shows substantial gains in translation quality.
Key facts
- Pre-trained Large Language Model (LLMs) are effective on many tasks in zero-shot and few-shot settings.
- We argue that, rather than continued pretraining, a more effective use of this data is the pre-LLM era technique of back-translation (BT).
- With 1K parallel pairs on Odia and Swahili, BT beats continued pretraining into the low-resource language by 7.7 to 16.4 XCOMET-XXL points in domain.
- Across five low-resource languages, with Hindi as a high-resource control, and two 3B backbones, BT improves translation into the low-resource language by 16 to 173% XCOMET-XXL over parallel-only fine-tuning.
Read aloud
Back-translation proves to be a significant technique for improving machine translation in low-resource languages. It outperforms continued pretraining and shows substantial gains in translation quality. Pre-trained Large Language Model L L M s are effective on many tasks in zero-shot and few-shot settings. We argue that, rather than continued pretraining, a more effective use of this data is the pre-L L M era technique of back-translation BT. With 1K parallel pairs on Odia and Swahili, BT beats continued pretraining into the low-resource language by seven point seven to sixteen point four XCOMET-XXL points in domain. Across five low-resource languages, with Hindi as a high-resource control, and two 3B backbones, BT improves translation into the low-resource language by sixteen to one hundred seventy three percent XCOMET-XXL over parallel-only fine-tuning.
Sources in this cluster
- Back-Translation Still Matters: Leveraging Monolingual Data for Low-Resource Machine Translation with LLMs
arXiv cs.CL · 14 Oct 2026, 03:50 AEST