Non-projective Parsing for Pre-ordering Statistical Machine Translation

Abstract

This dissertation thesis treats of the word reordering issue in machine translation Word reordering is an important issue in machine translation because the word orders in the source and target languages are generally different. The degree of the difference between linguistically divergent languages, such as Korean and English, becomes larger than that between similar ones. Although previous methods utilized syntactic trees for word reordering between divergent languages, their hierarchically imposed restrictions often prevents correct word reordering. Although the Inversion Transduction Grammar(ITG) constraints have been widely accepted, it often prevent to improve translation quality between linguistically divergent language pairs because of the restriction.

With three parallel corpora, we manually categorized the source of non-ITG word reordering. As a consequence, They appear in approximately 4\% to 10\% of all sentences. We also propose a novel reordering method via non-projective parsing. Instead of a syntactic tree from a traditional parser, a \textit{reordering tree} is proposed, which is produced by a reordering parser that is trained with a parallel corpus. It aims to model reordering phenomena more suitably than a syntactic tree by relaxing projective constraints and explicitly encoding reordering orientation in the tree. A reordering parser inspired by a non-projective parser produces the reordering tree. Under a pre-ordering SMT framework, we conducted experiments for three language pairs in both directions. Our proposed method showed promising improvement in both parsing accuracy and translation quality.