![[2601.09982] Context Volume Drives Performance: Tackling Domain Shift in Extremely Low-Resource Translation via RAG](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2601.09982] Context Volume Drives Performance: Tackling Domain Shift in Extremely Low-Resource Translation via RAG
Summary
This article presents a hybrid framework for improving neural machine translation performance in low-resource languages, specifically addressing domain shift challenges using Retrieval-Augmented Generation (RAG).
Why It Matters
The research highlights the difficulties faced by neural machine translation models when applied to low-resource languages, emphasizing the need for innovative solutions like RAG to enhance translation accuracy. This is crucial for preserving linguistic diversity and improving communication in underrepresented languages.
Key Takeaways
- Domain shift significantly impacts NMT performance in low-resource languages.
- A hybrid framework combining NMT and LLM via RAG can recover translation quality.
- The number of retrieved examples is more critical than the retrieval algorithm itself.
- LLMs serve as a safety net, effectively repairing translation failures.
- This approach can be applied to other low-resource languages facing similar challenges.
Computer Science > Computation and Language arXiv:2601.09982 (cs) [Submitted on 15 Jan 2026 (v1), last revised 16 Feb 2026 (this version, v2)] Title:Context Volume Drives Performance: Tackling Domain Shift in Extremely Low-Resource Translation via RAG Authors:David Samuel Setiawan, Raphaël Merx, Jey Han Lau View a PDF of the paper titled Context Volume Drives Performance: Tackling Domain Shift in Extremely Low-Resource Translation via RAG, by David Samuel Setiawan and 2 other authors View PDF HTML (experimental) Abstract:Neural Machine Translation (NMT) models for low-resource languages suffer significant performance degradation under domain shift. We quantify this challenge using Dhao, an indigenous language of Eastern Indonesia with no digital footprint beyond the New Testament (NT). When applied to the unseen Old Testament (OT), a standard NMT model fine-tuned on the NT drops from an in-domain score of 36.17 chrF++ to 27.11 chrF++. To recover this loss, we introduce a hybrid framework where a fine-tuned NMT model generates an initial draft, which is then refined by a Large Language Model (LLM) using Retrieval-Augmented Generation (RAG). The final system achieves 35.21 chrF++ (+8.10 recovery), effectively matching the original in-domain quality. Our analysis reveals that this performance is driven primarily by the number of retrieved examples rather than the choice of retrieval algorithm. Qualitative analysis confirms the LLM acts as a robust "safety net," repairing seve...
