回译依然重要：利用单语数据改进 LLM 低资源机器翻译
Back-Translation Still Matters: Leveraging Monolingual Data for Low-Resource Machine Translation with LLMs
AI 导读
一项 arXiv 研究提出，对低资源语言机器翻译，用 LLM 做回译（BT）比继续预训练更有效：当前微调后的模型翻译单语句子，再以反向方向训练，并与平行集更新交错进行。在 Odia 和 Swahili 上仅用 1K 平行句对，BT 在域内比继续预训练高 7.7 到 16.4 XCOMET-XXL 分；跨 5 种低资源语言和两个 3B 骨干模型，比仅平行集微调提升 16% 到 173%。
正文 · 原文
Abstract:Pre-trained Large Language Model (LLMs) are effective on many tasks in zero-shot and few-shot settings. For machine translation of low-resource languages, the obvious options are few-shot prompting and fine-tuning on a small parallel set. We ask: what if moderate amounts of low-resource monolingual data are also available? We argue that, rather than continued pretraining, a more effective use of this data is the pre-LLM era technique of back-translation (BT). The current fine-tuned LLM translates monolingual sentences in each language, and the resulting pairs train it in the reverse direction. Updates on the parallel set are interleaved for stable learning. 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. With beam-decoded intermediate translations, it also beats a model with 10 time more parallel data. In-context examples retrieved from the parallel data further improve BT. We also analyze the source, quality and amount of monolingual text. 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. Blind human evaluation on Odia and Sanskrit agrees with the automatic metrics.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.18818 [cs.CL] |
| (or arXiv:2610.18818v3 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.18818 arXiv-issued DOI via DataCite |
Submission history
From: Jitin Singla [view email]
[v1]
Tue, 13 Oct 2026 17:50:18 UTC (706 KB)
[v2]
Wed, 14 Oct 2026 07:55:24 UTC (706 KB)
[v3]
Mon, 26 Oct 2026 09:27:18 UTC (706 KB)
来源：arXiv cs.AI · arxiv.org