Method of Input Masking for Training Translation Models
摘要
Abstract
Despite significant advances in the field of machine translation, automatic systems of machine translation still have some systematic errors. A novel approach to training translation models is proposed in this work; it is based on masking input and output sequences. The proposed loss function is a generalization not only for the classical translation task but also for translation postediting and the task of masked language modeling. Training by means of the proposed method is studied for the quality in the task of translation from English to Russian both separately and combined with other methods for improving the quality of machine translation.