Integrating Vectorized Lexical Constraints for Neural Machine Translation