Abstract
Large language models such as ChatGPT are transforming (specialised) translation and terminology work. Yet, their performance with respect to terminological variation and regional language diversity remains limited. In the German-speaking context, standard varieties such as Austrian German, Swiss German and South Tyrolean German display significant differences to the dominant variety (German German). These differences are not marginal but embedded in social, educational and legal systems (Wissik 2010). When LLMs overlook such differences, they risk homogenising language use and decreasing the usage of non-dominant language varieties (Fleisig et al. 2024; Joshi et al. 2020), raising questions about power relations between language technology providers and language communities (Blodgett et al. 2020) as well as domain adaptation in LLMs (Mai et al. 2025).
The UniTermGPT project addresses this issue by collecting and annotating a corpus of university-related texts from Austria, Germany, Switzerland and South Tyrol, reflecting different varieties of the German language. From this corpus, terminology is extracted and aligned with existing terminological resources. In the next stage, ChatGPT is tested by means of different prompts so that expert annotators can assess how ChatGPT manages and translates university texts in different German varieties. By combining corpus linguistics, computational methods and expert annotation, UniTermGPT not only documents the challenges but also develops recommendations for practice and policy.
The overarching research question addressed by this project is: How can ChatGPT solve, if prompted accordingly, the issue of terminology injection in the translation of specialised language taking German language varieties into account? UniTermGPT thus aims to investigate how ChatGPT handles university-related terminology across selected German language varieties in Europe and how the output changes depending on the prompt. It also seeks to explore strategies (especially prompt engineering and terminology injection) to improve variety-specific translations and provide practical recommendations for their use.
In the ongoing project, the current phase involves compiling and processing both the corpus and the relevant terminological resources. Building on previous research in terminology work with LLMs (Reineke 2023) and in LLM-based translation (He 2024; He et al. 2024; Moslem et al. 2023), a pilot study testing prompting frameworks was already conducted in the field of university admission in the Austrian higher education system (Heinisch 2025), which showed that prompts specifying the “Austrian variety of the German language”, “Austrian German” or “Austrian university system” resulted in LLM outputs that contained terminology from the German university system (instead of the Austrian university system) as well as hallucinated terminology (Heinisch 2025).
In light of recent developments, promising solutions include retrieval-augmented generation (Lewis et al. 2020), which (applied to terminology tasks) entails supplying documents or glossaries containing the relevant terminology, and terminology-augmented generation (Di Nunzio 2025; Fleischmann and Lang 2025), which enables the integration of terminological resources such as terminological databases directly into the LLM workflow.
For practical application, UniTermGPT also formulates recommendations for using ChatGPT in specialised translation and prepares a policy brief on terminological variation in the context of LLM use.