Franz Josef Och

A systematic comparison of various statistical alignment models

We present and compare various methods for computing word alignments using statistical or heuristic models. We consider the five alignment models presented in Brown, Della Pietra, Della Pietra, and Mercer (1993), the hidden Markov alignment model, smoothing techniques, and refinements. These statistical models are compared with two heuristic models based on the Dice coefficient. We present different methods for combining word alignments to perform a symmetrization of directed statistical alignment models. As evaluation criterion, we use the quality of the resulting Viterbi alignment compared to a manually produced reference alignment. We evaluate the models on the German-English Verbmobil task and the French-English Hansards task. We perform a detailed analysis of various design decisions of our statistical alignment system and evaluate these on training corpora of various sizes. An important result is that refined alignment models with a first-order dependence and a fertility model yield significantly better results than simple heuristic models. In the Appendix, we present an efficient training algorithm for the alignment models presented.

Statistical phrase-based translation

We propose a new phrase-based translation model and decoding algorithm that enables us to evaluate and compare several, previously proposed phrase-based translation models. Within our framework, we carry out a large number of experiments to understand better and explain why phrase-based models out-perform word-based models. Our empirical results, which hold for all examined language pairs, suggest that the highest levels of performance can be obtained through relatively simple means: heuristic learning of phrase translations from word-based alignments and lexical weighting of phrase translations. Surprisingly, learning phrases longer than three words and learning phrases from high-accuracy word-level alignment models does not have a strong impact on performance. Learning only syntactically motivated phrases degrades the performance of our systems.

Minimum Error Rate Training in Statistical Machine Translation

Often, the training procedure for statistical machine translation models is based on maximum likelihood or related criteria. A general problem of this approach is that there is only a loose relation to the final translation quality on unseen text. In this paper, we analyze various training criteria which directly optimize translation quality. These training criteria make use of recently proposed automatic evaluation metrics. We describe a new algorithm for efficient training an unsmoothed error count. We show that significantly better results can often be obtained if the final evaluation criterion is taken directly into account as part of the training procedure.

Discriminative Training and Maximum Entropy Models for Statistical Machine Translation

We present a framework for statistical machine translation of natural languages based on direct maximum entropy models, which contains the widely used source-channel approach as a special case. All knowledge sources are treated as feature functions, which depend on the source language sentence, the target language sentence and possible hidden variables. This approach allows a baseline machine translation system to be extended easily by adding new feature functions. We show that a baseline statistical machine translation system is significantly improved using this approach.

Improved statistical alignment models

In this paper, we present and compare various single-word based alignment models for statistical machine translation. We discuss the five IBM alignment models, the Hidden-Markov alignment model, smoothing techniques and various modifications. We present different methods to combine alignments. As evaluation criterion we use the quality of the resulting Viterbi alignment compared to a manually produced reference alignment. We show that models with a first-order dependence and a fertility model lead to significantly better results than the simple models IBM-1 or IBM-2, which are not able to go beyond zero-order dependencies.

The Alignment Template Approach to Statistical Machine Translation

A phrase-based statistical machine translation approach the alignment template approach is described. This translation approach allows for general many-to-many relations between words. Thereby, the context of words is taken into account in the translation model, and local changes in word order from source to target language can be learned explicitly. The model is described using a log-linear modeling approach, which is a generalization of the often used source-channel approach. Thereby, the model is easier to extend than classical statistical machine translation systems. We describe in detail the process for learning phrasal translations, the feature functions used, and the search algorithm. The evaluation of this approach is performed on three different tasks. For the German-English speech Verbmobil task, we analyze the effect of various system components. On the French-English Canadian Hansards task, the alignment template system obtains significantly better results than a single-word-based translation model. In the Chinese-English 2002 National Institute of Standards and Technology (NIST) machine translation evaluation it yields statistically significantly better NIST scores than all competing research and commercial translation systems.

Lattice-based Minimum Error Rate Training for Statistical Machine Translation

Minimum Error Rate Training (MERT) is an effective means to estimate the feature function weights of a linear model such that an automated evaluation criterion for measuring system performance can directly be optimized in training. To accomplish this, the training procedure determines for each feature function its exact error surface on a given set of candidate translations. The feature function weights are then adjusted by traversing the error surface combined over all sentences and picking those values for which the resulting error count reaches a minimum. Typically, candidates in MERT are represented as N-best lists which contain the N most probable translation hypotheses produced by a decoder. In this paper, we present a novel algorithm that allows for efficiently constructing and representing the exact error surface of all translations that are encoded in a phrase lattice. Compared to N-best MERT, the number of candidate translations thus taken into account increases by several orders of magnitudes. The proposed method is used to train the feature function weights of a phrase-based statistical machine translation system. Experiments conducted on the NIST 2008 translation tasks show significant runtime improvements and moderate BLEU score gains over N-best MERT.

Phrase-Based Statistical Machine Translation

This paper is based on the work carried out in the framework of the VERBMOBIL project, which is a limited-domain speech translation task (German-English). In the final evaluation, the statistical approach was found to perform best among five competing approaches.In this paper, we will further investigate the used statistical translation models. A shortcoming of the single-word based model is that it does not take contextual information into account for the translation decisions. We will present a translation model that is based on bilingual phrases to explicitly model the local context. We will show that this model performs better than the single-word based model. We will compare monotone and non-monotone search for this model and we will investigate the benefit of using the sum criterion instead of the maximum approximation.

Automatic Evaluation of Machine Translation Quality Using Longest Common Subsequence and Skip-Bigram Statistics

In this paper we describe two new objective automatic evaluation methods for machine translation. The first method is based on longest common subsequence between a candidate translation and a set of reference translations. Longest common subsequence takes into account sentence level structure similarity naturally and identifies longest co-occurring in-sequence n-grams automatically. The second method relaxes strict n-gram matching to skip-bigram matching. Skip-bigram is any pair of words in their sentence order. Skip-bigram cooccurrence statistics measure the overlap of skip-bigrams between a candidate translation and a set of reference translations. The empirical results show that both methods correlate with human judgments very well in both adequacy and fluency.

A comparison of alignment models for statistical machine translation

In this paper, we present and compare various alignment models for statistical machine translation. We propose to measure the quality of an alignment model using the quality of the Viterbi alignment compared to a manually-produced alignment and describe a refined annotation scheme to produce suitable reference alignments. We also compare the impact of different alignment models on the translation quality of a statistical machine translation system.