The state-of-the-art MT approach is statistical MT, which learns statistical models from parallel bilingual corpus to model the translation process. There are mainly two very different statistical MT approaches, symbolic MT (e.g. phrase-based MT, hierarchical phrase-based MT and tree-to-string MT) and neural MT.
Symbolic MT extracts translation rules from parallel corpus to translate a new source sentence. A big challenge in symbolic MT is that the extracted translation rules are context-free and ambiguous. For a particular input sentence, symbolic MT needs to select the appropriate translation rules based on the sentence context. We proposed three novel neural models (a neural word reordering model, a binarized neural joint model and a neural rule selection model) to guide the translation process of symbolic MT, which outperformed previous related work significantly.
Neural MT (NMT) uses only a single large neural network to model the entire translation process. NMT takes the source sentence as input of the neural network and outputs the target sentence, which does not need any translation rules. Compared to symbolic MT, NMT generally generates more fluent translations, but still has a few disadvantages, such as NMT is more likely to generate completely unrelated translation, under and over translation. We proposed a method to incorporate symbolic MT decoding probabilities into NMT. We showed that the proposed method can improve NMT translation quality significantly.