The Reordering Problem in Statistical Machine Translation

Reordering of words is one of the most visible changes when translating a sentence from one language to another. The reordering problem has always been a central concern for machine translation. In this report, we look at the reordering problem in the context of Statistical Machine Translation. We first study and classify the reordering divergences between languages. We emphasize the contribution that linguistic understanding of these divergence patterns plays in developing efficient and effective solutions to the reordering problem. We study how the reordering is modeled in the major SMT models and analyze these models with respect to coverage of divergence patterns, model complexity and use of linguistic resources. Machine translation is distinguished from other machine learning problems with respect to the complexity of inference, known as decoding in machine translation terminology. We describe methods used to efficiently perform decoding while attempting to get the best possible reordering. We look at extensions beyond the noisy channel model, in the form of source reordering and discriminative re-ranking, which attempt to further improve reordering. In this report, our endeavour has been to do a holistic study of all aspects of the reordering problem, across all components of an SMT system and identify potential areas of research.