The statistical framework has proved to be very successful in machine translation. The main reason for this success is the existence of powerful techniques that allow to build machine translation systems automatically from available parallel corpora. Most of statistical machine translation approaches are based on single-word translation models, which do not take bilingual contextual information into account. The translation model in the phrase-based approach defines correspondences between sequences of contiguous source words (source segments) and sequences of contiguous target words (target segments) instead of only correspondences between single source words and single target words. That is, statistical phrase-based translation models make use of explicit bilingual contextual information. Different methods for the selection of adequate bilingual word sequences and for training the parameters of these models are reviewed in this paper. Improved techniques for the selection and training model parameters are also introduced. The phrase-based approach has been assessed in different tasks using different corpora and the results obtained are comparable or better than the ones obtained using other statistical and non-statistical machine translation systems.

This work has been partially supported by the Spanish project TIC2003-08681-C02, the Agencia Valenciana de Ciencia y Tecnología under contract GRUPOS03/031, the Generalitat Valenciana, and the project AMETRA (INTEK-CN03AD02).