Dynamic Oracle for Neural Machine Translation in Decoding Phase

Dynamic Oracle for Neural Machine Translation in Decoding Phase Zi-Yi Dou, Hao Zhou, Shu-Jian Huang, Xin-Yu Dai, Jia-Jun Chen National Key Laboratory for Novel Software Technology, Nanjing University Nanjing, 210023, China 141242042@smail.nju.edu.cn, {zhouh, huangsj, daixy, chenjj}@nlp.nju.edu.cn Abstract The past several years have witnessed the rapid progress of end-to-end Neural Machine Translation (NMT). However, there exists discrepancy between training and inference in NMT when decoding, which may lead to serious problems since the model might be in a part of the state space it has never seen during training. To address the issue, Scheduled Sampling has been proposed. However, there are certain limitations in Scheduled Sampling and we propose two dynamic oracle-based methods to improve it. We manage to mitigate the discrepancy by changing the training process towards a less guided scheme and meanwhile aggregating the oracle’s demonstrations. Experimental results show that the proposed approaches improve translation quality over standard NMT system. Keywords: machine translation, dynamic oracle, language model 1. Introduction Neural networks have been widely used contemporarily and have achieved great performance on a variety of fields like sentiment analysis (Santos and Gattit, 2014) and visual ob- ject recognition (Ciregan et al., 2012). For sequential prob- lems, recurrent neural networks can be applied to process sequences. To address issues like long term dependen- cies in the data (Bengio et al., 1994), the Long Short-Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997) or Gated Recurrent Unit (GRU) can be used to tackle the prob- lem (Cho et al., 2014). A straightforward application of the LSTM and GRU architecture have already shown impres- sive performance in several difficult tasks, including ma- chine translation (Sutskever et al., 2014), and image cap- tioning (Vinyals et al., 2015). Generally, a basic sequence-to-sequence model consists of two recurrent neural networks: an encoder that processes the input and a decoder that generates the output (Cho et al., 2014). In many applications of sequence-to-sequence models, at inference time, the output of the decoder at time t is fed back and becomes the input of decoder at time t+1. However, during training, it is more common to provide the correct input to the decoder at every time-step even if the decoder made a mistake before, which leads to a discrep- ancy between how the model is used at training and infer- ence. As has been pointed out by Bengio et. al. (2015), although this discrepancy can be mitigated by the use of a beam search heuristic maintaining several generated tar- get sequences, for continuous state space models like re- current neural networks, there is no dynamic programming approach, so the effective number of sequences considered remains small. The main problem is that mistakes made earlier in the sequence generation process are fed as input to the model and can be quickly amplified because the model might be in a part of the state space it has never seen at training time (Bengio et al., 2015). There are several existing methods to bridge the gap be- tween training and inference. Bengio …

Content truncated to 500 words