Improving translation for low-resource Finno-Ugric languages with Neural Machine Translation models

Abstract

Training a good neural machine translation model requires a lot of data. The majority of languages in the world have low amounts of suitable data available for this task. One possible solution to this problem is developing a multilingual model, combining high-resource and low-resource languages and creating a shared vocabulary space, where knowledge gained from high-resource languages is applied to translating low-resource languages. Another useful technique is to produce new data for low-resource languages by creating synthetic translations of monolingual data with a baseline model. In this thesis we use both of those methods, training a multilingual baseline model on Finno- Ugric language family data and increasing the amount of data for smaller Finno-Ugric languages by translating monolingual data with the multilingual baseline model in order to improve machine translation quality for low-resource languages.