Large Language Models im Umfeld der maschinellen Übersetzung. Eine Analyse des Potenzials von ChatGPT-4o zum Pre-Editing in der Wirtschaftsübersetzung

Abstract The translation industry is changing, and the capabilities of current artificial intelligence (AI) technologies now extend beyond machine translation (MT). In this paper, neural machine translation (NMT) and the use of AI will be assessed from the perspective of quality assurance. A quantitative error analysis was performed in order to determine the extent to which the usage of AI chatbots for automated pre-editing can improve the quality of NMT: Several English-language extracts of a Deutsche Bank Earnings Report were pre-edited and then translated into German multiple times, using ChatGPT-4o for pre-editing and DeepL for translation. The translation quality was evaluated both manually (MQM) and automatically (TER and COMET), the reference translation being the published German version of the annual report. Pre-editing was performed with one naïve prompt and two best practice prompts, the latter being created with the OpenAI Chat Playground and manually improved. The results show a potential in using AI for automated pre-editing, especially when it comes to style and linguistic conventions.