Word reordering is one of the most difficult aspects of statistical machinetranslation (SMT), and an important factor of its quality and efficiency.Despite the vast amount of research published to date, the interest of thecommunity in this problem has not decreased, and no single method appears to bestrongly dominant across language pairs. Instead, the choice of the optimalapproach for a new translation task still seems to be mostly driven byempirical trials. To orientate the reader in this vast and complex researcharea, we present a comprehensive survey of word reordering viewed as astatistical modeling challenge and as a natural language phenomenon. The surveydescribes in detail how word reordering is modeled within differentstring-based and tree-based SMT frameworks and as a stand-alone task, includingsystematic overviews of the literature in advanced reordering modeling. We thenquestion why some approaches are more successful than others in differentlanguage pairs. We argue that, besides measuring the amount of reordering, itis important to understand which kinds of reordering occur in a given languagepair. To this end, we conduct a qualitative analysis of word reorderingphenomena in a diverse sample of language pairs, based on a large collection oflinguistic knowledge. Empirical results in the SMT literature are shown tosupport the hypothesis that a few linguistic facts can be very useful toanticipate the reordering characteristics of a language pair and to select theSMT framework that best suits them.
展开▼
