Bidirectional Generative Adversarial Networks for Neural Machine Translation

Proceedings of the 22nd Conference on Computational Natural Language Learning (CoNLL 2018), pages 190–199 Brussels, Belgium, October 31 - November 1, 2018. c⃝2018 Association for Computational Linguistics 190 Bidirectional Generative Adversarial Networks for Neural Machine Translation Zhirui Zhang†∗, Shujie Liu§, Mu Li¶, Ming Zhou§, Enhong Chen† †University of Science and Technology of China, Hefei, China §Microsoft Research Asia †zrustc11@gmail.com †cheneh@ustc.edu.cn §{shujliu,mingzhou}@microsoft.com ¶limugx@outlook.com Abstract Generative Adversarial Network (GAN) has been proposed to tackle the exposure bias problem of Neural Machine Translation (NMT). However, the discriminator typically results in the instability of the GAN train- ing due to the inadequate training problem: the search space is so huge that sampled translations are not sufficient for discrimina- tor training. To address this issue and stabi- lize the GAN training, in this paper, we pro- pose a novel Bidirectional Generative Adver- sarial Network for Neural Machine Transla- tion (BGAN-NMT), which aims to introduce a generator model to act as the discriminator, whereby the discriminator naturally considers the entire translation space so that the inade- quate training problem can be alleviated. To satisfy this property, generator and discrimina- tor are both designed to model the joint prob- ability of sentence pairs, with the difference that, the generator decomposes the joint prob- ability with a source language model and a source-to-target translation model, while the discriminator is formulated as a target lan- guage model and a target-to-source transla- tion model. To further leverage the symme- try of them, an auxiliary GAN is introduced and adopts generator and discriminator mod- els of original one as its own discriminator and generator respectively. Two GANs are alter- nately trained to update the parameters. Exper- iment results on German-English and Chinese- English translation tasks demonstrate that our method not only stabilizes GAN training but also achieves significant improvements over baseline systems. 1 Introduction The past several years have witnessed the rapid de- velopment of Neural Machine Translation (NMT) ∗This work was done when the first author was the intern at Microsoft Research Asia. (Cho et al., 2014; Sutskever et al., 2014; Bahdanau et al., 2014), from catching up with Statistical Ma- chine Translation (SMT) (Koehn et al., 2003; Chi- ang, 2007) to outperforming it by significant mar- gins on many languages (Sennrich et al., 2016; Gehring et al., 2017; Vaswani et al., 2017; Has- san et al., 2018). The most common approach to training NMT is to maximize the conditional log-probability of the correct translation given the source sentence. However, as argued in Bengio et al. (2015), the Maximum Likelihood Estimation (MLE) principle suffers from so-called exposure bias in the inference stage: the model predicts next token conditional on its previously predicted ones that may be never observed in the training data. To address this problem, much recent work attempts to reduce the inconsistency between training and inference, such as adopting sequence-level objec- tives and directly maximizing BLEU scores (Ben- gio et al., 2015; Ranzato et al., 2015; Shen et al., 2016; Wiseman and Rush, 2016). Generative Adversarial Network (GAN) …

Content truncated to 500 words