Improving Domain-Specific Translation from English into Ukrainian with Retrieval-Augmented Generation

2026-06-28 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A study explores retrieval-augmented generation (RAG) to enhance English-Ukrainian machine translation within a private legal/military corpus. The approach feeds semantically similar translation units, found via vector embeddings, as in-context examples to large language models. Researchers evaluated three open-weight Gemma 3 models (4B, 12B, 27B) against Gemini 3 Flash, testing five augmentation conditions with k values of 0, 3, 5, 10, and 25 on a 2,581-pair index and a 258-pair test set. Findings indicate that context augmentation significantly improves both ChrF++ and COMET scores across all models. Notably, the 4B model's COMET score increased by 0.076 at k=3. However, smaller models like the 4B Gemma 3 exhibited context saturation, peaking at k=10 and degrading with further context, losing 9.72 ChrF++ points and 0.007 COMET between k=10 and k=25, while larger models continued to benefit from increased context.

Key takeaway

For NLP Engineers developing domain-specific translation systems, you should integrate retrieval-augmented generation (RAG) to significantly boost quality, especially for lower-resourced languages like Ukrainian. When implementing RAG, carefully tune the number of retrieved examples (k) for your chosen LLM. Be aware that smaller models, such as Gemma 3 4B, can experience context saturation, degrading performance if k is too high. Larger models, like Gemma 3 27B, generally benefit from more context.

Key insights

Retrieval-augmented generation significantly improves domain-specific English-Ukrainian translation, but smaller LLMs can suffer from context saturation.

Principles

• RAG enhances domain-specific translation quality.
• Context saturation affects smaller LLMs.
• Larger LLMs benefit from more context.

Method

Semantically similar translation units, retrieved via vector embeddings, are provided as in-context examples to a large language model for domain-specific translation.

In practice

• Use RAG for specialized translation tasks.
• Optimize k value for model size.
• Consider larger models for extensive context.

Topics

• Retrieval-Augmented Generation
• Machine Translation
• English-Ukrainian Translation
• Large Language Models
• Gemma 3
• Context Saturation
• Legal/Military Domain

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Retrieval-Augmented Generation as a Bridge Between Classical NLP Pipelines and Modern LLM Systems — Retrieval-augmented generation (RAG) addresses the limitation of static knowledge in large language models (LLMs) by connecting them to external, up-to-date knowledge sources.
• Combining Retrieval-Augmented Text Generation with LLMs for Reading Content Recommendations — RAG significantly enhances LLM-generated personalized reading content, improving relevance and groundedness.
• Linguistic Symptoms: Augmented Generation by Retrieval and Reasoning in LLMs under Portuguese-English Variation in Medical Contexts — Linguistic variation significantly impacts smaller LLMs in medical contexts, with larger models showing greater cross-language robustness.
• A Comparative Study of Language Models for Khmer Retrieval-Augmented Question Answering — RAG efficacy for low-resource languages like Khmer is bottlenecked by retriever performance, with generator strengths varying across RAGAS metrics.
• Büyük Dil Modellerinin Hafızasını Güçlendiren Yaklaşım:Retrieval-Augmented Generation (RAG) — RAG enhances LLMs by providing external, real-time context, reducing hallucinations and improving factual accuracy.
• AI-assisted cultural heritage dissemination: Comparing NMT and glossary-augmented LLM translation in rock art documents — Glossary-augmented LLMs significantly improve terminology accuracy in specialized translation without sacrificing overall quality.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.