Dual-Path Phrase-Based Statistical Machine Translation

Dual-Path Phrase-Based Statistical Machine Translation Susan Howlett and Mark Dras Centre for Language Technology Macquarie University Sydney, Australia {susan.howlett,mark.dras}@mq.edu.au Abstract Preceding a phrase-based statistical ma- chine translation (PSMT) system by a syntactically-informed reordering prepro- cessing step has been shown to improve overall translation performance compared to a baseline PSMT system. However, the improvement is not seen for every sen- tence. We use a lattice input to a PSMT system in order to translate simultaneously across both original and reordered ver- sions of a sentence, and include a number of confidence features to support the sys- tem in choosing on a sentence-by-sentence basis whether to use the reordering pro- cess. In German-to-English translation, our best system achieves a BLEU score of 21.39, an improvement of 0.62. 1 Introduction Machine translation (MT) is the automatic transla- tion of text from one human language to another. Statistical MT accomplishes this through a proba- bilistic model of the translation process. In phrase-based statistical MT (PSMT), trans- lation proceeds by dividing a sentence into se- quences of adjacent words called phrases, then translating each phrase and reordering the phrases according to a distortion model. The distortion model may be lexicalised but does not typically incorporate information about the syntactic struc- ture of the sentence. As such, although PSMT has been very successful, it suffers from the lack of a principled mechanism for handling long-distance reordering phenomena due to word order differ- ences between languages. One method for addressing this difficulty is the reordering-as-preprocessing approach, exem- plified by Collins et al. (2005) and Xia and Mc- Cord (2004), where PSMT is coupled with a pre- processing step that reorders input sentences to more closely parallel the target language word order. Although this leads to improved perfor- mance overall, Collins et al. (2005) show that the reordering-as-preprocessing system does not con- sistently provide better translations than the PSMT baseline on a sentence-by-sentence basis. One possible reason could be errors in the parse or the consequent reordering. Chiang et al. (2009) used features indicating problematic use of syntax to improve performance within hierarchical and syntax-based translation. In this work, we want to see whether syntax-related features can help choose between original and reordered sentence translations in PSMT. We use as our starting point the PSMT system Moses (Koehn et al., 2007). In order to use fea- tures within the system’s log-linear model to as- sess the reliability of syntax, it is necessary to in- put both variants simultaneously. To do this, we adapt in a novel way the lattice input of Moses; we refer to this new system as dual-path PSMT (§3). We then augment the model with a number of confidence features to enable it to evaluate which of the two paths is more likely to yield the best translation (§3.2). We reimplement the Collins et al. (2005) reordering preprocessing step and con- duct some preliminary experiments in German-to- English translation (§4). Our results (§5) do not replicate the finding of Collins et al. (2005) that the …

Content truncated to 500 words