Generative adversarial training for neural machine translation

Abstract Neural machine translation (NMT) is typically optimized to generate sentences which cover n-grams with ground target as much as possible. However, it is widely acknowledged that n-gram precisions, the manually designed approximate loss function, may mislead the model to generate suboptimal translations. To solve this problem, we train the NMT model to generate human-like translations directly by using the generative adversarial net, which has achieved great success in computer vision. In this paper, we build a conditional sequence generative adversarial net (CSGAN-NMT) which comprises of two adversarial sub models, a generative model (generator) which translates the source sentence into the target sentence as the traditional NMT models do and a discriminative model (discriminator) which discriminates the machine-translated target sentence from the human-translated one. The two sub models play a minimax game and achieve a win-win situation when reaching a Nash Equilibrium. As a variant of the single generator-discriminator model, the multi-CSGAN-NMT which contains multiple discriminators and generators, is also proposed. In the multi-CSGAN-NMT model, each generator is viewed as an agent which can interact with others and even transfer messages. Experiments show that the proposed CSGAN-NMT model obtains substantial improvements than the strong baseline and the improvement of the multi-CSGAN-NMT model is more remarkable.