The Impact of Source–Side Syntactic Reordering on Hierarchical Phrase-based SMT

The Impact of Source–Side Syntactic Reordering on Hierarchical Phrase-based SMT Jinhua Du, Andy Way CNGL, School of Computing, Dublin City University, Dublin, Ireland jdu,away@computing.dcu.ie Abstract Syntactic reordering has been demon- strated to be helpful and effective for han- dling different word orders between source and target languages in SMT. However, in terms of hierarchial PB-SMT (HPB), does the syntactic reordering still has a signif- icant impact on its performance? This paper introduces a reordering approach which explores the � (DE) grammati- cal structure in Chinese. We employ the Stanford DE classifier to recognise the DE structures in both training and test sentences of Chinese, and then per- form word reordering to make the Chi- nese sentences better match the word or- der of English. The annotated and re- ordered training data and test data are ap- plied to a re-implemented HPB system and the impact of the DE construction is ex- amined. The experiments are conducted on the NIST 2008 evaluation data and ex- perimental results show that the BLEU and METEOR scores are significantly im- proved by 1.83/8.91 and 1.17/2.73 abso- lute/relative points respectively. 1 Introduction Syntactic structure-based reordering has been shown to be significantly helpful for handling word order issues in phrase-based machine translation (PB-SMT) (Xia and McCord, 2004; Collins et al., 2005; Li et al., 2007; Wang et al., 2007; Elming, 2008; Chang et al., 2009). Generally, PB-SMT has an independent reordering model because the phrases themselves in PB-SMT do not have an ability to perform word reordering. However, as regards the hierarchical PB-SMT (HPB) (Chiang, c⃝ 2010 European Association for Machine Translation. 2005), it possesses an inherent word reordering ca- pability because each rule is a hierarchical struc- ture which contains the sub-phrases. The order of a sub-phrase pair in a source–target hierarchial phrase pair is decided by the positions of non- terminals when generating the hierarchical phrase based on the word alignment links between the source and target language sentence. Generally, there is no specific reordering model inside the HPB system. When the hierarchical phrase is ex- panded during the decoding process, the word or- der is carried out automatically according to the positions of the corresponding non-terminal sym- bols and the synchronous context-free grammars (SCFG) between the source and target phrases. The phrases in HPB are generally variable- based phrases, i.e., short phrases are substituted in a long phrase with variables or non-terminals based on SCFG rules. The method of constructing variable-based phrases facilitates learning of the different orderings between the source and target language to some extent. Therefore, in some sense, it can be regarded that the hierarchical phrases con- tain a hidden reordering model. However, this hid- den model cannot handle some flexible syntactic structures well such as the � (DE) in Chinese. In this paper, we utilise the Stanford DE-annotated approach (Chang et al., 2009) in both training and test sentences of Chinese, and then carry out word reordering on the Chinese side so as to adjust the Chinese sentences to …

Content truncated to 500 words