Improving Neural Machine Translation with Conditional Sequence Generative Adversarial Nets

Proceedings of NAACL-HLT 2018, pages 1346–1355 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Improving Neural Machine Translation with Conditional Sequence Generative Adversarial Nets Zhen Yang1,2, Wei Chen1∗, Feng Wang1,2, Bo Xu1 1Institute of Automation, Chinese Academy of Sciences 2University of Chinese Academy of Sciences {yangzhen2014, wei.chen.media, feng.wang, xubo}@ia.ac.cn Abstract This paper proposes an approach for apply- ing GANs to NMT. We build a condition- al sequence generative adversarial net which comprises of two adversarial sub models, a generator and a discriminator. The generator aims to generate sentences which are hard to be discriminated from human-translated sen- tences ( i.e., the golden target sentences); And the discriminator makes efforts to discriminate the machine-generated sentences from human- translated ones. The two sub models play a mini-max game and achieve the win-win situation when they reach a Nash Equilibri- um. Additionally, the static sentence-level BLEU is utilized as the reinforced objective for the generator, which biases the generation towards high BLEU points. During training, both the dynamic discriminator and the stat- ic BLEU objective are employed to evaluate the generated sentences and feedback the eval- uations to guide the learning of the generator. Experimental results show that the proposed model consistently outperforms the traditional RNNSearch and the newly emerged state-of- the-art Transformer on English-German and Chinese-English translation tasks. 1 Introduction Neural machine translation (Kalchbrenner and Blunsom, 2013; Sutskever et al., 2014; Cho et al., 2014; Bahdanau et al., 2014) which directly lever- ages a single neural network to transform the source sentence into the target sentence, has drawn more and more attention in both academia and in- dustry (Shen et al., 2015; Wu et al., 2016; Johnson et al., 2016; Gehring et al., 2017; Vaswani et al., 2017). This end-to-end NMT typically consists of two sub neural networks. The encoder network reads and encodes the source sentence into the context vector representation; and the decoder net- work generates the target sentence word by word based on the context vector. To dynamically gen- erate a context vector for a target word being gen- erated, the attention mechanism which enables the model to focus on the relevant words in the source- side sentence is usually deployed. Under the encoder-decoder framework, many variants of the model structure, such as convolutional neural net- work (CNN) and recurrent neural network (RN- N) are proposed (Bahdanau et al., 2014; Gehring et al., 2017). Recently, (Gehring et al., 2017) propose the Transformer, the first sequence trans- duction model based entirely on attention, achiev- ing state-of-the-art performance on the English- German and English-French translation tasks. De- spite its success, the Transformer, similar to tradi- tional NMT models, is still optimized to maximize the likelihood estimation of the ground word (M- LE) at each time step. Such an objective poses a hidden danger to NMT models. That is, the model may generate the best candidate word for the cur- rent time step yet a bad component of the whole sentence in the long run. Minimum risk …

Content truncated to 500 words