Interactive Machine Translation with Large Language Models for Low-resource Languages

Large language models (LLM) have been applied to machine translation with notable success. However, while the quality of automatic translations is good enough for some applications, those translations still do not fit the requirements for critical tasks. Thus, interactive machine translation (IMT) emerges as a post-edition variant between automatic and traditional (human) translation. It consists of an iterative and collaborative framework in which the machine provides translation hypotheses that are partially validated and corrected by a human expert. The ultimate goal of this procedure is to reduce the human effort required to generate the final high-quality translations. Studies that have integrated LLMs into IMT tools have obtained significant improvements for languages with large resources. Nevertheless, the low-resource setting is one of the hardest challenges in machine translation (MT). We contribute to the research in this field by testing LLMs in low-resource languages as a part of an IMT experiment. Thus, we chose six representative LLMs{\textemdash}mBART, M2M, Flan-T5, NLLB, EuroLLM and Gemma{\textemdash}to perform an IMT experiment between Galician{\textendash}English, Swahili{\textendash}English and Catalan{\textendash}English pairs. Our results reflect the complexity of low-resource MT, but also that LLMs can effectively reduce the required human effort.