It Takes Two: A Dual Stage Approach for Terminology-Aware Translation

TL;DR

DuTerm tackles domain-specific terminology translation by coupling a terminology-aware NMT trained on large-scale synthetic, tagged data with a prompt-based LLM for post-editing. The approach integrates a terminology extraction/generation pipeline, tag standardization, quality filtering, and multilingual model adaptation to preserve term integrity across languages. Evaluation under the WMT 2025 Terminology Shared Task shows a trade-off: strict term enforcement boosts lexical accuracy but may reduce fluency, while context-aware LLM post-editing can achieve higher overall translation quality, especially in morphologically rich languages. This hybrid workflow demonstrates the potential of combining training-time terminology handling with inference-time post-editing to improve terminology-driven MT in real-world settings.

Abstract

This paper introduces DuTerm, a novel two-stage architecture for terminology-constrained machine translation. Our system combines a terminology-aware NMT model, adapted via fine-tuning on large-scale synthetic data, with a prompt-based LLM for post-editing. The LLM stage refines NMT output and enforces terminology adherence. We evaluate DuTerm on English-to German, English-to-Spanish, and English-to-Russian with the WMT 2025 Terminology Shared Task corpus. We demonstrate that flexible, context-driven terminology handling by the LLM consistently yields higher quality translations than strict constraint enforcement. Our results highlight a critical trade-off, revealing that an LLM's work best for high-quality translation as context-driven mutators rather than generators.