The paper introduces the alignment template approach to statistical machine translation, which allows for general many-to-many relations between words and explicitly learns local changes in word order. The model is described using a log-linear modeling approach, which generalizes the source-channel approach. The process for learning phrasal translations, the feature functions used, and the search algorithm are detailed. The approach is evaluated on three tasks: the German-English VERBMOBIL task, the French-English HANSARDS task, and the Chinese-English NIST 2002 machine translation evaluation. The alignment template system outperforms single-word-based translation models and yields statistically significantly better results than competing systems. The paper also discusses the training and implementation details of the model, including the use of alignment templates, feature functions, and a dynamic-programming-based beam search algorithm.The paper introduces the alignment template approach to statistical machine translation, which allows for general many-to-many relations between words and explicitly learns local changes in word order. The model is described using a log-linear modeling approach, which generalizes the source-channel approach. The process for learning phrasal translations, the feature functions used, and the search algorithm are detailed. The approach is evaluated on three tasks: the German-English VERBMOBIL task, the French-English HANSARDS task, and the Chinese-English NIST 2002 machine translation evaluation. The alignment template system outperforms single-word-based translation models and yields statistically significantly better results than competing systems. The paper also discusses the training and implementation details of the model, including the use of alignment templates, feature functions, and a dynamic-programming-based beam search algorithm.