Deriving Benefit from a Generalized Syntax-based Reordering

In this study we describe a syntax-based word reordering technique for n-gram-based statistical machine translation (SMT). The proposed distortion model operates with generalized unlexicalized rules and aims to order source language words so that translation is close to monotonic, simplifying the translation process. In the final step, we apply a translation units blending strategy, combining bilingual tuples extracted from the parallel corpora with monotone and reordered source parts. Experiments are reported on the BTEC corpus from tourist domain for the Arabic-English translation task, the proposed tuples blending technique significantly outperformes the monotone system.