Long-distance reordering during search for hierarchical phrase-based SMT

Long-distance reordering during search for hierarchical phrase-based SMT Fabienne Braune Anita Gojun Institute for NLP Universit¨at Stuttgart Pfaffenwaldring 5b D-70569 Stuttgart, Germany braunefe,gojunaa,fraser@ims.uni-stuttgart.de Alexander Fraser Abstract Long-distance reordering of syntactically divergent language pairs is a critical prob- lem. SMT has had limited success in han- dling these reorderings during inference, and thus deterministic preprocessing based on reordering parse trees is used. We con- sider German-to-English translation using Hiero. We show how to effectively model long-distance reorderings during search. Our work is novel in that we look at re- ordering distances of up to 50 words, and conduct a detailed manual analysis based on a new gold standard. 1 Introduction Word reordering is a well-known issue in SMT. One successful approach has been to use rule- based preprocessing to reorder parse trees. We would like to perform reordering during inference. Phrase-based hierarchical models (Chiang, 2007) have helped, but reordering over long distances is still a difficult open problem. Consider the fol- lowing German sentence and English output taken from the hierarchical component of the Moses toolkit (Hoang et al., 2009). These sentences il- lustrate the successful reordering of the partici- ple geeinigt (agreed) from the end of the German clause, to be next to the English auxiliary have. (1) deutschland (germany) , frankreich (france) , israel (israel) und (and) die (the) usa- (us) haben (have) sich (themselves) im (in) mai (may) 2006 darauf (on) geeinigt (agreed) , es (it) zu (to) tun (do). (2) germany , france , israel and the us have agreed in may 2006 , to do it . This reordering involves a word movement over 5 tokens and is therefore not a long-distance reorder- c⃝ 2012 European Association for Machine Translation. ing. However, there can be many more words be- tween the German auxiliary and participle, so the movement required can become arbitrarily large. Restrictions on reordering distance are typically used with hierarchical systems like Hiero because previous experiments have shown some evidence that long-distance reorderings are not effective (Chiang, 2007). We are not aware of careful explo- rations focusing exclusively on long-distance re- orderings in search, prior to our work. We present the first step towards solving the problem of long-distance reorderings during search. We first analyze the rule geometry re- quired for long-distance German-to-English move- ment and modify extraction of Hiero’s SCFG rules to focus on these rules. We then introduce a new idea, span-width-specific rules in the grammar. By span, we denote the number of tokens that are allowed to be covered by a non-terminal symbol (usually “X”) in the source language side of an SCFG rule. We define long-distance reordering as occurring over spans containing 11 to 50 source words, and define a new set of rules which ap- ply over spans of 11 to 50 words, which we call long spans. We combine these rules (applied on 11 to 50 word spans) with the standard Hiero X rules (applied on 1 to 10 word spans). We further restrict our rules by applying …

Content truncated to 500 words