Mutual Information and Diverse Decoding Improve Neural Machine Translation

Sequence-to-sequence neural translation models learn semantic and syntactic relations between sentence pairs by optimizing the likelihood of the target given the source, i.e., p(y|x), an objective that ignores other potentially useful sources of information. We introduce an alternative objective function for neural MT that maximizes the mutual information between the source and target sentences, modeling the bi-directional dependency of sources and targets. We implement the model with a simple re-ranking method, and also introduce a decoding algorithm that increases diversity in the Nbest list produced by the first pass. Applied to the WMT German/English and French/English tasks, both mechanisms offer a consistent performance boost on both standard LSTM and attention-based neural MT architectures. The result is the best published performance for a single (non-ensemble) neural MT system, as well as the potential application of our diverse decoding algorithm to other NLP re-ranking tasks.