Syntactic reordering in statistical machine translation

Reordering has been an important topic in statistical machine translation (SMT) as long as SMT has been around. State-of-the-art SMT systems such as Pharaoh (Koehn, 2004a) still employ a simplistic model of the reordering process to do non-local reordering. This model penalizes any reordering no matter the words. The reordering is only selected if it leads to a translation that looks like a much better sentence than the alternative. Recent developments have, however, seen improvements in translation quality following from syntax-based reordering. One such development is the pre-translation approach that adjusts the source sentence to resemble target language word order prior to translation. This is done based on rules that are either manually created or automatically learned from word aligned parallel corpora. We introduce a novel approach to syntactic reordering. This approach provides better exploitation of the information in the reordering rules and eliminates problematic biases of previous approaches. Although the approach is examined within a pre-translation reordering framework, it easily extends to other frameworks. Our approach significantly outperforms a state-of-the-art phrase-based SMT system and previous approaches to pretranslation reordering, including (Li et al., 2007; Zhang et al., 2007b; Crego & Mari˜ no, 2007). This is consistent both for a very close language pair, English-Danish, and a very distant language pair, English-Arabic. We also propose automatic reordering rule learning based on a rich set of linguistic information. As opposed to most previous approaches that extract a large set of rules, our approach produces a small set of predominantly general rules. These provide a good reflection of the main reordering issues of a given language pair. We examine the influence of several parameters that may have influence on the quality of the rules learned. Finally, we provide a new approach for improving automatic word alignment. This word alignment is used in the above task of automatically learning reordering rules. Our approach learns from hand aligned data how to combine several automatic word alignments to one superior word alignment. The automatic word alignments are created from the same data that has been preprocessed with different tokenization schemes. Thus utilizing the different strengths that different tokenization schemes exhibit in word alignment. We achieve a 38% error reduction for the automatic word alignment