Redefining Terminology Work: The Role of Glossaries and TMs in Customized AI-driven Machine Translation

Abstract This study investigates whether traditional terminology work in the customization of machine translation (MT) systems can be effectively replaced by translation memories (TMs) alone. Given the growing reliance on AI-driven translation tools, we evaluated three MT configurations using English–slovak technical documentation: a baseline (non-customized system), a system customized with both TMs and a glossary, and a system customized with TMs only. since the text corpora for the given area were sufficient, we used the LLM model to generate additional training data. Results show that TM-only customization can achieve terminology translation accuracy nearly equivalent to setups that include glossaries—particularly when supported by high-quality, domain-specific bilingual data. Nonetheless, glossary-based customization further improves consistency, and terminology errors persist across all systems. This suggests that although automation of translation processes can reduce dependence on traditional terminology building, terminology databases remain essential for ensuring the quality (QA) of the output text. The study offers practical guidance for translators, terminologists, and developers of translation tools by emphasizing the importance of collaboration between automated and human-driven translation processes. It also underscores both the promise and limitations of LLM-generated data for domain adaptation in low-resource language settings.