Neural Machine Translation with Fully Information Transformation

Neural machine translation (NMT) has become the de facto standard in the machine translation community due to its competitive performance and simple model structure. During the Ph.D., I am engaged to develop NMT models with fully information transformation, i.e. sufficient information transformation (adequate translation), and efficient information transformation (fast translation). In this thesis, I first leverage the explicit linguistic knowledge, e.g. syntactic dependencies and cross-lingual reordering information, and implicit knowledge, e.g. better model initialization derived from self-supervised pretraining, to facilitate the sufficient information transformation. Then, I focus on explaining the reason why the non-autoregressive translation performs worse than its autoregressive counterpart: the low-frequency word translation problem, and address it from several perspectives, e.g. learning objective-level, data-level