This paper presents and compares various single-word based alignment models for statistical machine translation. The authors discuss the five IBM alignment models, the Hidden Markov alignment model, smoothing techniques, and various modifications. They present different methods to combine alignments and use the quality of the resulting Viterbi alignment compared to a manually produced reference alignment as the evaluation criterion. They show that models with first-order dependence and a fertility model lead to significantly better results than the simple IBM-1 or IBM-2 models. The paper discusses the IBM translation models IBM-1 to IBM-5 and the Hidden-Markov alignment model. The different alignment models provide different decompositions of the probability. The Viterbi alignment is the alignment that maximizes the probability for a specific model. The authors propose using the quality of the Viterbi alignment compared to a manually produced alignment as an evaluation criterion. The paper describes the structure of the six models, including the IBM-1 to IBM-5 models and the Hidden-Markov alignment model. The models differ in the alignment model (zero-order or first-order), the existence of an explicit fertility model, and whether the model is deficient or not. The authors also discuss the training of the models using the EM-algorithm and the use of smoothing techniques to avoid over-fitting and handle rare words. The paper also discusses the problem of deficiency in IBM-3 and IBM-4 models, where the EM-algorithm can increase the likelihood by aligning more words to the empty word. The authors modify these models to avoid this issue. The paper presents an evaluation methodology for alignment models, including an annotation scheme and a corresponding evaluation criterion. The authors also discuss generalized alignments, where a source word can be aligned with multiple target words. They show that combining Viterbi alignments from both translation directions can improve precision and recall. The paper presents experiments on the VERBMOBIL and HANSARDS tasks, showing that models with first-order dependence and a fertility model lead to better results. The authors also discuss the effect of smoothing, the use of a larger training corpus, and the impact of using different alignment models. They conclude that sophisticated alignment models with first-order dependence and a fertility model lead to significantly better results than simple models.This paper presents and compares various single-word based alignment models for statistical machine translation. The authors discuss the five IBM alignment models, the Hidden Markov alignment model, smoothing techniques, and various modifications. They present different methods to combine alignments and use the quality of the resulting Viterbi alignment compared to a manually produced reference alignment as the evaluation criterion. They show that models with first-order dependence and a fertility model lead to significantly better results than the simple IBM-1 or IBM-2 models. The paper discusses the IBM translation models IBM-1 to IBM-5 and the Hidden-Markov alignment model. The different alignment models provide different decompositions of the probability. The Viterbi alignment is the alignment that maximizes the probability for a specific model. The authors propose using the quality of the Viterbi alignment compared to a manually produced alignment as an evaluation criterion. The paper describes the structure of the six models, including the IBM-1 to IBM-5 models and the Hidden-Markov alignment model. The models differ in the alignment model (zero-order or first-order), the existence of an explicit fertility model, and whether the model is deficient or not. The authors also discuss the training of the models using the EM-algorithm and the use of smoothing techniques to avoid over-fitting and handle rare words. The paper also discusses the problem of deficiency in IBM-3 and IBM-4 models, where the EM-algorithm can increase the likelihood by aligning more words to the empty word. The authors modify these models to avoid this issue. The paper presents an evaluation methodology for alignment models, including an annotation scheme and a corresponding evaluation criterion. The authors also discuss generalized alignments, where a source word can be aligned with multiple target words. They show that combining Viterbi alignments from both translation directions can improve precision and recall. The paper presents experiments on the VERBMOBIL and HANSARDS tasks, showing that models with first-order dependence and a fertility model lead to better results. The authors also discuss the effect of smoothing, the use of a larger training corpus, and the impact of using different alignment models. They conclude that sophisticated alignment models with first-order dependence and a fertility model lead to significantly better results than simple models.