Statistical Machine Translation with Word- and Sentence-Aligned Parallel Corpora

Statistical Machine Translation with Word- and Sentence-Aligned Parallel Corpora Chris Callison-Burch David Talbot School on InformaticsMiles Osborne University of Edinburgh 2 Buccleuch Place Edinburgh, EH8 9LW callison-burch@ed.ac.uk Abstract The parameters of statistical translation models are typically estimated from sentence-aligned parallel corpora. We show that significant improvements in the alignment and translation quality of such mod- els can be achieved by additionally including word- aligned data during training. Incorporating word- level alignments into the parameter estimation of the IBM models reduces alignment error rate and increases the Bleu score when compared to training the same models only on sentence-aligned data. On the Verbmobil data set, we attain a 38% reduction in the alignment error rate and a higher Bleu score with half as many training examples. We discuss how varying the ratio of word-aligned to sentence- aligned data affects the expected performance gain. 1 Introduction Machine translation systems based on probabilistic translation models (Brown et al., 1993) are gener- ally trained using sentence-aligned parallel corpora. For many language pairs these exist in abundant quantities. However for new domains or uncommon language pairs extensive parallel corpora are often hard to come by. Two factors could increase the performance of statistical machine translation for new language pairs and domains: a reduction in the cost of cre- ating new training data, and the development of more efficient methods for exploiting existing train- ing data. Approaches such as harvesting parallel corpora from the web (Resnik and Smith, 2003) address the creation of data. We take the second, complementary approach. We address the prob- lem of efficiently exploiting existing parallel cor- pora by adding explicit word-level alignments be- tween a number of the sentence pairs in the train- ing corpus. We modify the standard parameter esti- mation procedure for IBM Models and HMM vari- ants so that they can exploit these additional word- level alignments. Our approach uses both word- and sentence-level alignments for training material. In this paper we: 1. Describe how the parameter estimation frame- work of Brown et al. (1993) can be adapted to incorporate word-level alignments; 2. Report significant improvements in alignment error rate and translation quality when training on data with word-level alignments; 3. Demonstrate that the inclusion of word-level alignments is more effective than using a bilin- gual dictionary; 4. Show the importance of amplifying the contri- bution of word-aligned data during parameter estimation. This paper shows that word-level alignments im- prove the parameter estimates for translation mod- els, which in turn results in improved statistical translation for languages that do not have large sentence-aligned parallel corpora. 2 Parameter Estimation Using Sentence-Aligned Corpora The task of statistical machine translation is to choose the source sentence, e, that is the most prob- able translation of a given sentence, f, in a for- eign language. Rather than choosing e∗ that di- rectly maximizes p(e|f), Brown et al. (1993) apply Bayes’ rule and select the source sentence: e∗ = arg max e p(e)p(f|e). (1) In this equation p(e) is a language model probabil- …

Content truncated to 500 words