ABSTRACT
This chapter introduces the novel discourse-processing methods in statistical machine translation (SMT) and neural machine translation (NMT) architectures. It describes our first attempt at investigating the potential for implicitly incorporating discourse information into machine translation (MT) and addresses the research question about MT problems at discourse level, regarding the influence of global context on SMT and NMT performance. Focusing on two typical discourse MT scenarios as pro-drop language and large-context NMT, the chapter presents some novel approaches to translation quality improvement targeting these problems. Experiment results show that it is crucial to identify the dropped pronouns to improve translation performance and demonstrate that the novel model significantly outperforms a strong attention-based large-context NMT baseline system.
