lexically constrained neural machine translation (NMT), which controls the
generation of NMT models with pre-specified constraints, is important in many
practical scenarios. Due to the representation gap between discrete constraints
and continuous vectors in NMT models, most existing w