Fuzzy Syntactic Reordering for Phrase-based Statistical Machine Translation

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 227–236, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Fuzzy Syntactic Reordering for Phrase-based Statistical Machine Translation Jacob Andreas and Nizar Habash and Owen Rambow Center for Computational Learning Systems Columbia University jda2129@columbia.edu {habash,rambow}@ccls.columbia.edu Abstract The quality of Arabic-English statistical ma- chine translation often suffers as a result of standard phrase-based SMT systems’ inabil- ity to perform long-range re-orderings, specif- ically those needed to translate VSO-ordered Arabic sentences. This problem is further ex- acerbated by the low performance of Arabic parsers on subject and subject span detection. In this paper, we present two parse “fuzzi- fication” techniques which allow the transla- tion system to select among a range of pos- sible S–V re-orderings. With this approach, we demonstrate a 0.3-point improvement in BLEU score (69% of the maximum possible using gold parses), and a corresponding im- provement in the percentage of syntactically well-formed subjects under a manual evalua- tion. 1 Introduction The question of how to effectively use phrase-based statistical machine translation (PSMT) to translate between language pairs which require long-range re- ordering has attracted a great deal of interest in re- cent years. The inability to capture long-range re- ordering behaviors is a weakness inherent in PSMT systems, which typically have only two mechanisms to control the reordering between source and tar- get language: (1) distortion penalties, which penal- ize or forbid long-distance re-orderings in order to reduce the search space explored by the decoder, and (2) lexicalized reordering models, which cap- ture the preferences of individual phrases to orient themselves monotonically, reversed with their pre- ceding phrases or discontinuously. Because both of these mechanisms work at the phrase level, they have proven very effective at capturing short-range reordering behaviors, but unable to describe long range movements; in fact, the distortion penalty ef- fectively causes the translation system to not pre- fer long-range re-orderings, even when they are as- signed significantly higher probability by the lan- guage model. The problem is particularly acute in translating from Arabic to English: Arabic sentences frequently exhibit a VSO ordering (both VSO and SVO are permitted in Arabic), while English permits only an SVO order. Past research has shown that verb anticipation and subject-span detection is a ma- jor source of error when translating from Arabic to English (Green et al., 2009; Bisazza and Fed- erico, 2010). Unable to perform long-range reorder- ing, PSMT frequently produces English sentences in which verbs precede their subjects (sometimes with “hallucinated” pronouns in front of them) or do not appear at all. Intuitively, better handling of these re- orderings has the potential to improve both accuracy and fluency of translation. In this paper, we present two parse fuzzification techniques which allow the translation system to se- lect among a range of possible S–V re-orderings. With this approach, we demonstrate a 0.3-point im- provement in BLEU score (69% of the maximum possible using gold parses), and a corresponding im- provement in the percentage of syntactically …

Content truncated to 500 words