End-to-End Training of Both Translation Models in the Back-Translation Framework

Semi-supervised learning algorithms in neural machine translation (NMT) have significantly improved translation quality compared to the supervised learning algorithms by using additional monolingual corpora. Among them, back-translation is a theoretically well-structured and cutting-edge method. Given two pre-trained NMT models between source and target languages, one translates a monolingual sentence as a latent sentence, and the other reconstructs the monolingual input sentence given the latent sentence. Therefore, previous works tried to apply the variational auto-encoder's (VAE) training framework to the back-translation framework. However, the discrete property of the latent sentence made it impossible to use backpropagation in the framework. This paper proposes a categorical reparameterization trick that generates a differentiable sentence, with which we practically implement the VAE's training framework for the back-translation and train it by end-to-end backpropagation. In addition, we propose several regularization techniques that are especially advantageous to this framework. In our experiments, we demonstrate that our method makes backpropagation available through the latent sentences and improves the BLEU scores on the datasets of the WMT18 translation task.