This paper presents and compares various statistical and heuristic models for computing word alignments in bilingual sentence-aligned corpora. The authors evaluate five alignment models from Brown et al. (1993), the hidden Markov alignment model, smoothing techniques, and refinements. These are compared with two heuristic models based on the Dice coefficient. The authors propose a method for combining word alignments to symmetrize directed statistical alignment models. The evaluation criterion is the quality of the resulting Viterbi alignment compared to a manually produced reference alignment. The models are evaluated on the German-English Verbmobil task and the French-English Hansards task. The authors analyze various design decisions of their statistical alignment system and evaluate these on training corpora of various sizes. An important result is that refined alignment models with first-order dependence and a fertility model yield significantly better results than simple heuristic models. The paper also presents an efficient training algorithm for the alignment models. The authors describe different statistical alignment models, heuristic models, and their comparison. They also describe the training of the alignment models and present a new training schedule that yields significantly better results. They describe heuristic methods for improving alignment quality by performing a symmetrization of word alignments. They describe an evaluation methodology for word alignment methods dealing with ambiguities associated with the word alignment annotation based on generalized precision and recall measures. The paper presents a systematic comparison of the various statistical alignment models with regard to alignment quality and translation quality. The authors show that the refined alignment models of Brown et al. (1993) are suitable for small corpora if parametrized appropriately. The paper also describes fertility-based alignment models and their training. The authors describe the computation of the Viterbi alignment for each alignment model. They describe the EM algorithm for training the model parameters. They discuss the problem of deficiency in the alignment models and propose solutions. They describe smoothing techniques to avoid overfitting and enable the models to cope better with rare words. They describe the use of a bilingual dictionary as an additional knowledge source in training. They describe methods for symmetrizing directed statistical alignment models by combining alignments in both translation directions. They describe an evaluation methodology for word alignment methods based on generalized precision and recall measures. The paper presents results of experiments involving the Verbmobil and Hansards tasks. The Verbmobil task is a speech translation task in the domain of appointment scheduling, travel planning, and hotel reservation. The Hansards task consists of debates in the Canadian parliament. The paper shows that the refined alignment models with first-order dependence and a fertility model yield significantly better results than simple heuristic models. The authors also show that the use of a bilingual dictionary can improve the alignment quality. The paper concludes that the refined alignment models with first-order dependence and a fertility model are more effective than simple heuristic models.This paper presents and compares various statistical and heuristic models for computing word alignments in bilingual sentence-aligned corpora. The authors evaluate five alignment models from Brown et al. (1993), the hidden Markov alignment model, smoothing techniques, and refinements. These are compared with two heuristic models based on the Dice coefficient. The authors propose a method for combining word alignments to symmetrize directed statistical alignment models. The evaluation criterion is the quality of the resulting Viterbi alignment compared to a manually produced reference alignment. The models are evaluated on the German-English Verbmobil task and the French-English Hansards task. The authors analyze various design decisions of their statistical alignment system and evaluate these on training corpora of various sizes. An important result is that refined alignment models with first-order dependence and a fertility model yield significantly better results than simple heuristic models. The paper also presents an efficient training algorithm for the alignment models. The authors describe different statistical alignment models, heuristic models, and their comparison. They also describe the training of the alignment models and present a new training schedule that yields significantly better results. They describe heuristic methods for improving alignment quality by performing a symmetrization of word alignments. They describe an evaluation methodology for word alignment methods dealing with ambiguities associated with the word alignment annotation based on generalized precision and recall measures. The paper presents a systematic comparison of the various statistical alignment models with regard to alignment quality and translation quality. The authors show that the refined alignment models of Brown et al. (1993) are suitable for small corpora if parametrized appropriately. The paper also describes fertility-based alignment models and their training. The authors describe the computation of the Viterbi alignment for each alignment model. They describe the EM algorithm for training the model parameters. They discuss the problem of deficiency in the alignment models and propose solutions. They describe smoothing techniques to avoid overfitting and enable the models to cope better with rare words. They describe the use of a bilingual dictionary as an additional knowledge source in training. They describe methods for symmetrizing directed statistical alignment models by combining alignments in both translation directions. They describe an evaluation methodology for word alignment methods based on generalized precision and recall measures. The paper presents results of experiments involving the Verbmobil and Hansards tasks. The Verbmobil task is a speech translation task in the domain of appointment scheduling, travel planning, and hotel reservation. The Hansards task consists of debates in the Canadian parliament. The paper shows that the refined alignment models with first-order dependence and a fertility model yield significantly better results than simple heuristic models. The authors also show that the use of a bilingual dictionary can improve the alignment quality. The paper concludes that the refined alignment models with first-order dependence and a fertility model are more effective than simple heuristic models.