Since around the year 2000, the dominant paradigm in research on machine translation has changed from an “expert system” approach to one based on machine learning, called “statistical machine translation” (SMT). The purpose of the talk is to give an overview of SMT and to encourage researchers in machine learning to consider entering the field. After a description of how the SMT systems are evaluated experimentally, the five IBM translation models will be summarized. Then, the key concepts underlying today’s phrase-based SMT systems will be described. Finally, the talk will discuss five interesting open problems in SMT: system adaptation, feature selection, generalizing beyond phrases, discriminative training of SMT systems, and the application of kernel methods to SMT.