Automating translation checks of financial documents using large language models

Abstract

We introduce a tool for automated translation checking of financial reports in German-English. It uses a heuristic matching algorithm followed by a transformer encoder based error detection model on sentence pair level. For generating the training data, we leverage state-of-the-art large language models such as GPT-4o, thereby alleviating the need for expert annotations. The results suggest that smaller models fine-tuned specifically for this task significantly outperform large multi-purpose generative models like GPT-4 for this particular problem, and that a combination of informed and deep learning approaches works best in this case. The tool is being made publicly available as a demonstrator.