Evaluating the Use of Generative LLMs for Intralingual Diachronic Translation of Middle-Polish Texts into Contemporary Polish
Abstract: This paper presents efforts towards creating a tool for translating texts from Middle Polish into modern Polish. Archaic texts sourced from the CBDU digital library were translated into modern language using ChatGPT and the resulting parallel corpus was used to train a neural text-to-text model. We assessed the results using automatic metrics and performed human evaluation of translations of the best-performing model and ChatGPT. Even though the performance of the trained models was far from perfect, the quality of translations produced with ChatGPT was good in most cases. Although caution should be exercised, we believe that LLMs have a high potential for text-to-text annotation applications.
External IDs:dblp:conf/icadl/KlamraKO23
Loading