Retrieval-Augmented Terminology Translation for English-Russian: A Multi-Domain Study
Abstract
AbstractNeural machine translation reaches near-human quality on general-domain text but systematically fails on specialized terminology, where polysemous and low-frequency terms carry domain-specific meanings absent from general training data. We investigate two complementary interventions for English-Russian specialized translation across ten professional domains: (i) a four-stage retrieval-augmented pipeline of LLM-based term extraction, domain classification, lemma-based knowledge-base retrieval, and glossary-augmented translation; and (ii) a preference-based fine-tuning stage combining supervised fine-tuning (SFT) with Contrastive Preference Optimization (CPO) layered on top of retrieval. We release an open multi-domain English-Russian terminology knowledge base of 50,000 terms with 120,161 domain-tagged translations, a 200-item curated test set selected for terminology difficulty, and a 1,300-item Wikipedia-derived test set. Across four open base models, retrieval-augmented prompting alone delivers the bulk of the terminology gain and already exceeds the two commercial reference systems; the additional preference-tuning stage contributes only a small further gain at a measurable fluency cost. The pattern transfers to the WMT25 Terminology Translation Task, the most recent iteration of the shared task that included English-Russian as a language pair.