Abstract
We investigate different approaches to translate between similar languagesunder low resource conditions, as part of our contribution to the WMT 2020Similar Languages Translation Shared Task. We submitted Transformer-basedbilingual and multilingual systems for all language pairs, in the twodirections. We also leverage back-translation for one of the language pairs,acquiring an improvement of more than 3 BLEU points. We interpret our resultsin light of the degree of mutual intelligibility (based on Jaccard similarity)between each pair, finding a positive correlation between mutualintelligibility and model performance. Our Spanish-Catalan model has the bestperformance of all the five language pairs. Except for the case ofHindi-Marathi, our bilingual models achieve better performance than themultilingual models on all pairs.