Model improves glossary term use by learning from hard translation mistakes
SalamandraTA at WMT 2026 Terminology Shared Task: Hard Examples Are Better Teachers
Computation and Language
Summary
Getting machine translations that use exact words from a glossary is hard because many training examples don't help the model learn to follow the glossary. The authors found that training only on examples where the model's translation disagrees with the glossary makes the model much better at using the right terms. They used this idea to build a new system that translated with over 94% accuracy on terminology and did very well in an industry-wide competition. This approach helps make sure machines stick to a defined list of terms in translations.
What this means in practice
- •For translation software developers: Improve translation models to better respect client-specific term glossaries by training only on examples that conflict with glossary terms.
- •For localization project managers: Use less training data while increasing terminology accuracy in machine translation workflows by selecting challenging training pairs.
Authors
Xixian Liao, Maite Melero
Abstract
Terminology-aware translation asks for more than a correct translation: the output must use the exact terms a glossary prescribes. The standard recipe, fine-tuning on glossary-annotated translation pairs, hides an inefficiency: for most examples the glossary prescribes exactly what the model would have produced anyway, so they teach nothing about following a glossary. We therefore keep only the examples where the model's own translation contradicts the glossary. In a controlled study at fixed data volume, this selection alone raises term accuracy from 78.7% to 89.9%. The filtered data, built by a two-way synthetic pipeline on open models, is part of the instruction-tuning mixture of our public release SalamandraTA-7b-instruct v3.0, which, used exactly as released and wrapped in a document-level inference pipeline, forms the BSC submission to the WMT26 Terminology Shared Task Track 1. At the official WMT26 evaluation, our system achieves 94.2% term success at 74.6 chrF++, with only two of the twenty-two submissions outperforming it on both metrics. On last year's benchmark, it also surpasses our GRPO-based system, despite being trained solely with ordinary supervised fine-tuning.