Abstract

We present the first neural machine translation system for translationbetween the endangered Erzya language and Russian and the dataset collected byus to train and evaluate it. The BLEU scores are 17 and 19 for translation toErzya and Russian respectively, and more than half of the translations arerated as acceptable by native speakers. We also adapt our model to translatebetween Erzya and 10 other languages, but without additional parallel data, thequality on these directions remains low. We release the translation modelsalong with the collected text corpus, a new language identification model, anda multilingual sentence encoder adapted for the Erzya language. These resourceswill be available at https://github.com/slone-nlp/myv-nmt.

 

Quick Read (beta)