Data-driven machine translation has shifted from traditional statistical machine translation to neural machine translation (NMT) powered by deep learning. Although translations generated by NMT systems are now much better than those by SMT in some languages, it is still difficult for NMT to handle discourse-level machine translation, translation of sentences with complex linguistic patterns and so on. Prior and external knowledge is widely considered helpful on these issues. However, incorporating external symbolic knowledge into NMT is non-trivial. In this talk, I will discuss the connections among data, knowledge and NMT models, as well as our recent efforts in exploring external knowledge in NMT.