During the last few years, the statistical approach has found widespread use in machine translation of both written and spoken language. In many comparative evaluations, the statistical approach was found to be competitive or superior to the existing conventional approaches. Like other natural language processing tasks, machine translation requires four major components: an error measure for the decision rule that is used to generate the target sentence from the source sentence; a set of probability models that replace the true but unknown probability distributions in the decision rule, a training criterion that is used to learn the unknown model parameters from training data; an efficient implementation of the decision rule, which is referred to as generation or, like in speech recognition, as search or decoding.
We will consider each of these four components in more detail and review the attempts that have been made to improve the state of the art. In addition, we will address the problem of recognition-translation integration which is specific of spoken language translation.