Discriminative Reordering Models for Statistical Machine Translation

Proceedings of the Workshop on Statistical Machine Translation, pages 55–63, New York City, June 2006. c⃝2006 Association for Computational Linguistics Discriminative Reordering Models for Statistical Machine Translation Richard Zens and Hermann Ney Human Language Technology and Pattern Recognition Lehrstuhl f¨ur Informatik 6 – Computer Science Department RWTH Aachen University, D-52056 Aachen, Germany {zens,ney}@cs.rwth-aachen.de Abstract We present discriminative reordering models for phrase-based statistical ma- chine translation. The models are trained using the maximum entropy principle. We use several types of features: based on words, based on word classes, based on the local context. We evaluate the overall performance of the reordering models as well as the contribution of the individual feature types on a word-aligned corpus. Additionally, we show improved transla- tion performance using these reordering models compared to a state-of-the-art baseline system. 1 Introduction In recent evaluations, phrase-based statistical ma- chine translation systems have achieved good per- formance. Still the fluency of the machine transla- tion output leaves much to desire. One reason is that most phrase-based systems use a very simple re- ordering model. Usually, the costs for phrase move- ments are linear in the distance, e.g. see (Och et al., 1999; Koehn, 2004; Zens et al., 2005). Recently, in (Tillmann and Zhang, 2005) and in (Koehn et al., 2005), a reordering model has been described that tries to predict the orientation of a phrase, i.e. it answers the question ’should the next phrase be to the left or to the right of the current phrase?’ This phrase orientation probability is con- ditioned on the current source and target phrase and relative frequencies are used to estimate the proba- bilities. We adopt the idea of predicting the orientation, but we propose to use a maximum-entropy based model. The relative-frequency based approach may suffer from the data sparseness problem, because most of the phrases occur only once in the training corpus. Our approach circumvents this problem by using a combination of phrase-level and word-level features and by using word-classes or part-of-speech information. Maximum entropy is a suitable frame- work for combining these different features with a well-defined training criterion. In (Koehn et al., 2005) several variants of the ori- entation model have been tried. It turned out that for different tasks, different models show the best per- formance. Here, we let the maximum entropy train- ing decide which features are important and which features can be neglected. We will see that addi- tional features do not hurt performance and can be safely added to the model. The remaining part is structured as follows: first we will describe the related work in Section 2 and give a brief description of the baseline system in Section 3. Then, we will present the discriminative reordering model in Section 4. Afterwards, we will evaluate the performance of this new model in Sec- tion 5. This evaluation consists of two parts: first we will evaluate the prediction capabilities of the model on a word-aligned corpus and second we will show improved translation quality compared to …

Content truncated to 500 words