Machine Translation between Uruguayan Sign Language (LSU) Glosses and Spanish

Resumen

This paper presents the construction of a parallel corpus for Uruguayan Sign Language (LSU) represented as glosses and spoken Spanish, and some experiments on machine translation for this language pair. The best MT models in our experiments are obtained by fine-tuning NLLB, and achieved 16.55 BLEU in the LSU to Spanish direction, and 17.90 BLEU in the opposite direction. Our manual analysis of the results shows that the translations are not perfect but are already showing good grasp of the ideas of the text being translated.