Review-Based Curriculum Learning for Neural Machine Translation

For Neural Machine Translation (NMT) tasks with limited domain resources, curriculum learning provides a way to simulate the human learning process from simple to difficult to adapt the general NMT model to a specific domain. However, previous curriculum learning methods suffer from catastrophic forgetting and learning inefficiency. In this paper, we introduce a review-based curriculum learning method, targetedly selecting curriculum according to long time interval or unskilled mastery. Furthermore, we add general domain data to curriculum learning, using the mixed fine-tuning method, to improve generalization and robustness of translation. Extensive experimental results and analysis show that our method outperforms other curriculum learning baselines across three specific domains.