Iterative Self-Learning for Enhanced Back-Translation in Low Resource Neural Machine Translation.

摘要: Many language pairs are low resource - the amount and/or quality of parallel data is not sufficient to train a neural machine translation (NMT) model which can reach an acceptable standard accuracy. works have explored use easier-to-get monolingual improve performance models in this category languages and even high languages. The most successful such back-translation using translations target increase training data. backward trained on available has been shown determine approach. approaches especially where model. Among self-learning iterative back-translation. These methods were perform better than This work presents self-training approach as improvement over further enhance Over several iterations, synthetic generated by used its through forward translation. Experiments that method outperforms both IWSLT'14 English German NMT. While also back-translation, though slightly, number required be reduced exactly iterations.