Abstract: Neural Machine Translation (NMT) has been through great revolutions in recent years. Accompanied with improvements in translation quality are works that attempted to understand the working mechanism of various aspects of the NMT framework. In our paper, we survey those efforts on unveiling the \textit{black box} of the standard NMT framework. To begin with, we briefly introduce the three critical components of the holistic NNT framework; nextly, we deliver a clear \textit{component-centric} categorization and clean summary of these specific works \textit{guided} by \textit{frequently-asked} questions (FAQs) that aim at making up \textit{lack} of understanding; finally, we discuss several limitations, future directions and inspirations. We believe this paper could facilitate the community to weave a holistic and clear picture of our current understandings of the standard NMT framework and shed light on its future improvements and developments. Please check this website https://nmtology.github.io/ for a visual guidance of the FAQs.
0 Replies
Loading