Online Distilling from Checkpoints for Neural Machine Translation
Current predominant neural machine translation (NMT) models often have a deep structure with large amounts of parameters, making these models hard to train and easily suffering from over-fitting. A common practice is to utilize a validation set to evaluate the training process and select the best checkpoint. Average and ensemble techniques on checkpoints can lead to further performance improvement. However, as these methods do not affect the training process, the system performance is restricted to the checkpoints generated in original training procedure. In contrast, we propose an online knowledge distillation method. Our method on-the-fly generates a teacher model from checkpoints, guiding the training process to obtain better performance. Experiments on several datasets and language pairs show steady improvement over a strong self-attention-based baseline system. We also provide analysis on data-limited setting against over-fitting. Furthermore, our method leads to an improvement in a machine reading experiment as well.
PDF Abstract
