Empirical lower bounds on alignment error rates in syntax-based machine translation

The empirical adequacy of synchronous context-free grammars of rank two (2-SCFGs) (Satta and Peserico, 2005), used in syntax-based machine translation systems such as Wu (1997), Zhang et al. (2006) and Chi-ang (2007), in terms of what alignments they induce, has been discussed in Wu (1997) and Wellington et al. (2006), but with a one-sided focus on so-called “inside-out alignments”. Other alignment configurations that cannot be induced by 2-SCFGs are identified in this pa-per, and their frequencies across a wide col-lection of hand-aligned parallel corpora are examined. Empirical lower bounds on two measures of alignment error rate, i.e. the one introduced in Och and Ney (2000) and one where only complete translation units are con-sidered, are derived for 2-SCFGs and related formalisms.