Fine-tuning NMT Models and LLMs for Specialised EN-ES Translation Using Aligned Corpora, Glossaries, and Synthetic Data: MULTITAN at WMT25 Terminology Shared Task
Résumé
This paper presents a hybrid evaluation of terminology-aware English-to-Spanish machine translation systems developed for the WMT25 Terminology Shared Task, specifically targeting the Information Technology (IT) domain. Our objective was to improve terminology accuracy and overall translation quality and highlight the potential of specialised terminology-aware translation models for technical domains. We used different enhancement strategies for both neural machine translation (NMT) systems and large language models (LLMs). These strategies include fine-tuning with synthetic data, the use of in-domain parallel corpora, and hard constraint methods such as placeholder substitution and in-context glossary integration. The results demonstrate distinct lexical and stylistic profiles in the outputs of fine-tuned NMT systems and LLMs, as well as the complementary advantages of different terminology injection methods. Systems behave differently with and without a glossary, as demonstrated by experimental results. The NMT systems seem to be rather limited in adapting to special lexicons and resizing embeddings, which is the opposite of LLMs, which prefer structured instructions. Although our translation systems achieved their highest scores on the NoTerm, Consistency metrics, exceeding 81%, demonstrating their ability to produce stable and coherent translations of recurring terms and phrases in unconstrained settings, the precision of the terminology and overall quality of the translation could have been improved by additional training.
| Licence |
|---|