📄 TaMTAS: Terminology-Aware Machine Translation for Accessible Science. Large Corpus compilation, terminology extraction and data augmentation

Summary

Most scientific work is published in English. This limits access for non-English readers and makes it harder for researchers to share findings in their own languages. This short project paper introduces TaMTAS, a 36-month consortium effort (Dec 2025–Dec 2028) coordinated by UOC with BSC, University of Surrey, DCU and University of Tartu. It aims to build an open-source translation ecosystem for Life Sciences texts. Conventional NMT often loses terminological consistency across long documents, so TaMTAS frames translation as multi-step reasoning with Large Reasoning Models (LRMs), using chain-of-thought prompting, glossary constraints and self-correction. The paper outlines seven work packages and details WP2. WP2 will build Life Sciences corpora for five languages and upgrade TBXTools to output richly annotated TBX termbases. It will also create synthetic training data by swapping terms inside aligned segments. The project is still in its early phase, so no experimental results are reported. Its significance lies in the planned openly licensed, FAIR-compliant resources, especially for lower-resourced Estonian and Irish.

Key points

• Languages and domain: English, Spanish, Catalan, Estonian and Irish in the Life Sciences domain. Planned sources include open-access journals, public health guidelines and regulatory documents, aligned at sentence and paragraph level.
• Work package structure: WP1/WP2 (UOC) cover management and data. WP3 (BSC) covers terminology-aware LRM translation. WP4 (UoS) covers QE and APE. WP5 (UT) covers simplification and explanatory enrichment. WP6/WP7 (DCU) cover evaluation and dissemination.
• TBXTools (Oliver and Vàzquez, 2015): the tool will be extended to produce TBX termbases with morphological and syntactic metadata. It will also handle complex multi-word and highly specialised terms better.
• Data augmentation: rare or OOV terms are addressed by substituting contextually similar terms within existing parallel segments. This avoids manual data collection.
• Release and funding: all outputs will be released under open-source licences in FAIR repositories. The project is funded by CHIST-ERA / MICIU/AEI (PCI2025-167063-2).