Using Self-Training to Improve Back-Translation in Low Resource Neural Machine Translation

摘要: Improving neural machine translation (NMT) models using the back-translations of monolingual target data (synthetic parallel data) is currently state-of-the-art approach for training improved systems. The quality backward system - which trained on available and used back-translation has been shown in many studies to affect performance final NMT model. In low resource conditions, usually not enough train a model that can produce qualitative synthetic needed standard This work proposes self-training strategy where output improve itself through forward technique. technique was baseline IWSLT'14 English-German IWSLT'15 English-Vietnamese by 11.06 1.5 BLEUs respectively. generated out-performed another 2.7 BLEU.