The first attempts at Machine Translation (MT) date from the fifties. They were based on information theoretic principles and crude statistical techniques. Disappointment about the results, as worded in the ALPAC report (1966), and the growing tendency toward a formal and computational approach in linguistics led to a more knowledge-based methodology with the integration of morphological, syntactic and semantic modules. This was typical of the MT systems that were developed in the seventies and the eighties. Their growing complexity and high cost, both financially and computationally, led to a grinding halt in the early nineties and prepared the minds for a new paradigm shift, away from linguistics and toward statistics again. At this point by the end of its second decade, statistical MT has matured into a field with a basic consensus on matters of methodology, objectives and the measurement of progress. The time is hence ripe for a book-length synthesis, preferably by an author who has contributed to the field, who has the ability to explain complex issues in an accessible way, and who has enough background and experience to offer a broad survey. With this book, Philipp Koehn shows himself worthy of the challenge: Statistical Machine Translation provides an excellent synthesis of a vast amount of literature (the bibliography section takes up 45 double-column pages) and presents it in a well-structured and articulate way. Moreover, the book has been class-tested and contains a set of exercises at the end of each chapter, as well as numerous references to open source tools and resources which enable the diligent reader to build MT systems for any language pair.