Syntactic Reordering Integrated with Phrase-Based SMT

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 209–216 Manchester, August 2008 Syntactic Reordering Integrated with Phrase-based SMT Jakob Elming Computational Linguistics Copenhagen Business School jel.isv@cbs.dk Abstract We present a novel approach to word reordering which successfully integrates syntactic structural knowledge with phrase-based SMT. This is done by con- structing a lattice of alternatives based on automatically learned probabilistic syntactic rules. In decoding, the alter- natives are scored based on the output word order, not the order of the input. Unlike previous approaches, this makes it possible to successfully integrate syntactic reordering with phrase-based SMT. On an English-Danish task, we achieve an absolute improvement in translation qual- ity of 1.1 % BLEU. Manual evaluation supports the claim that the present ap- proach is significantly superior to previous approaches. 1 Introduction The emergence of phrase-based statistical machine translation (PSMT) (Koehn et al., 2003) has been one of the major developments in statistical ap- proaches to translation. Allowing translation of word sequences (phrases) instead of single words provides SMT with a robustness in word selection and local word reordering. PSMT has two means of reordering the words. Either a phrase pair has been learned where the tar- get word order differs from the source (phrase in- ternal reordering), or distance penalized orderings of target phrases are attempted in decoding (phrase c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. external reordering). The first solution is strong, the second is weak. The second solution is necessary for reorderings within a previously unseen sequence or over dis- tances greater than the maximal phrase length. In this case, the system in essence relies on the tar- get side language model to get the correct word order. The choice is made without knowing what the source is. Basically, it is a bias against phrase external reordering. It seems clear that reordering often depends on higher level linguistic information, which is ab- sent from PSMT. In recent work, there has been some progress towards integrating syntactic infor- mation with the statistical approach to reorder- ing. In works such as (Xia and McCord, 2004; Collins et al., 2005; Wang et al., 2007; Habash, 2007), reordering decisions are done “determinis- tically”, thus placing these decisions outside the actual PSMT system by learning to translate from a reordered source language. (Crego and Mari˜no, 2007; Zhang et al., 2007; Li et al., 2007) are more in the spirit of PSMT, in that multiple reorderings are presented to the PSMT system as (possibly weighted) options. Still, there remains a basic conflict between the syntactic reordering rules and the PSMT system: one that is most likely due to the discrepancy be- tween the translation units (phrases) and units of the linguistic rules, as (Zhang et al., 2007) point out. In this paper, we proceed in the spirit of the non- deterministic approaches by providing the decoder with multiple source reorderings. But instead of scoring the input word order, we score the …

Content truncated to 500 words