Modelling specialized language through AI: challenges in training systems for technical domain translation [Articol]

Abstract

Modelling specialized language with artificial intelligence is a significant challenge for specialized translation, especially in fields where controlled terminology is essential, such as medicine, engineering, law, or the exact sciences. The performance of neural machine translation systems depends directly on the quality of the corpora used for training. However, obtaining clean, standardized, and representative corpora is difficult because access to technical documents is often limited, the structure of specialized texts varies, and terminology may be inconsistent across existing sources. As a result, unedited data can lead to semantic errors, incorrect generalizations, and conceptual confusion—issues that are particularly serious in fields where accuracy is critical. Another risk is the tendency of AI models to fill information gaps using statistical assumptions, which can lead to the use of inappropriate terms or to distorted conceptual relations. In this context, it becomes necessary to develop effective strategies for selecting, annotating, and cleaning corpora to ensure the terminological consistency required for training the models. This article proposes a hybrid AI–translator approach, in which technology and human expertise work together: AI increases the speed and volume of processing, while the translator validates terminology, checks conceptual coherence, and corrects semantic errors. By analysing current challenges and presenting practical solutions, this contribution aims to show how artificial intelligence can be integrated responsibly into specialized translation, without compromising the precision and accuracy required in technical fields.