Unsupervised Pretraining for Neural Machine Translation Using Elastic Weight Consolidation
This work addresses efficiency and data requirements in machine translation, but it is incremental as it builds on existing regularization methods with limited effectiveness when tasks are unrelated.
The paper tackles the problem of unsupervised pretraining for neural machine translation by using Elastic Weight Consolidation (EWC) to prevent forgetting during fine-tuning, achieving BLEU scores similar to prior work while converging 2-3 times faster and eliminating the need for unlabeled data during fine-tuning.
This work presents our ongoing research of unsupervised pretraining in neural machine translation (NMT). In our method, we initialize the weights of the encoder and decoder with two language models that are trained with monolingual data and then fine-tune the model on parallel data using Elastic Weight Consolidation (EWC) to avoid forgetting of the original language modeling tasks. We compare the regularization by EWC with the previous work that focuses on regularization by language modeling objectives. The positive result is that using EWC with the decoder achieves BLEU scores similar to the previous work. However, the model converges 2-3 times faster and does not require the original unlabeled training data during the fine-tuning stage. In contrast, the regularization using EWC is less effective if the original and new tasks are not closely related. We show that initializing the bidirectional NMT encoder with a left-to-right language model and forcing the model to remember the original left-to-right language modeling task limits the learning capacity of the encoder for the whole bidirectional context.