Multiple reorderings in phrase-based machine translation

Abstract

This paper presents a method to integrate multiple reordering strategies in phrase-based statistical machine translation. Recently there has been much research effort in reordering problems in machine translation. State-of-the-art decoders incorporate sophisticated local reordering strategies, but there is little research on a unified approach to incorporate various kinds of reordering methods. We present a phrase-based decoder which easily allows multiple reordering schemes. We show how to use this framework to perform distance-based reordering and HIERO-style (Chiang 2005) hierarchical reordering. We also present two novel syntax-based reordering methods, one built on part-of-speech tags and the other based on parse trees. We will give experimental results using these relatively easy to implement methods on standard tests.