Since reordering in phrase-based and word-based models leads to an exponential growth of the search space, reordering constraints are introduced for efficient decoding.
Reordering Constraints is the main subject of 10 publications. 9 are discussed here.
Matusov et al. (2005)
constrain reordering when the input word sequence was consistently translated monotone in the training data. Zens and Ney (2003)
; Zens et al. (2004)
; Kanthak et al. (2005)
compare different reordering constraints and their effect on translation performance, including the formal grammar ITG constraint, which may be further restricted by insisting on a match to source side syntax (Yamamoto et al., 2008)
Similarly, reordering may be restricted to syntactically cohesion (Cherry, 2008)
. Ge et al. (2008)
integrate other linguistically inspired reordering models into a phrase-based decoder. Dreyer et al. (2007)
compare reordering constraints in terms of oracle BLEU, i.e., the maximum possible BLEU score. Goh et al. (2011)
use constraints based on part-of-speech patterns mostly along clause boundaries to improve Japansese-English patent translation.
Instead of hard limits on reordering, reordering models may inform search.