Discontinuous Statistical Machine Translation with Target-Side Dependency Syntax

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 239–247, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. Discontinuous Statistical Machine Translation with Target-Side Dependency Syntax Nina Seemann and Andreas Maletti Institute for Natural Language Processing, Universit¨at Stuttgart Pfaffenwaldring 5b, 70569 Stuttgart, Germany {seemanna,maletti}@ims.uni-stuttgart.de Abstract For several languages only potentially non-projective dependency parses are readily available. Projectivizing the parses and utilizing them in syntax-based trans- lation systems often yields particularly bad translation results indicating that those translation models cannot properly utilize such information. We demonstrate that our system based on multi bottom-up tree transducers, which can natively handle discontinuities, can avoid the large transla- tion quality deterioration, achieve the best performance of all classical syntax-based translation systems, and close the gap to phrase-based and hierarchical systems that do not utilize syntax. 1 Introduction Syntax-based machine translation, in which the transfer is achieved from and/or to the level of syntax, has become widely used in the statisti- cal machine translation community (Bojar et al., 2014). Different grammar formalisms have been proposed and evaluated as translation models driv- ing the translation systems. We use a variant of the local multi bottom-up tree transducer as pro- posed by Maletti (2011). More precisely, we use a string-to-tree variant of it, which offers two imme- diate advantages: (i) The source side of the rules is a simple string containing terminal symbols and the unique non-terminal X. Consequently, we do not need to match an input sentence parse, which allows additional flexibility. It has been demonstrated that this flexibility in the input of- ten yields improved translation quality (Chiang, 2010). (ii) The target language side offers discon- tinuities because rules can contain a sequence of target tree fragments instead of a single tree frag- ment. These fragments are applied synchronously, which allows the model to synchronously develop discontinuous parts in the output (e.g., to realize agreement). Overall, this translation model al- ready proved to be useful when translating from English into German, Chinese, and Arabic as demonstrated by Seemann et al. (2015). The goal of the current contribution is to adjust the approach and the system to Eastern European languages, for which we expect discontinuities to occur. The ex- isting system (Seemann et al., 2015) cannot read- ily be applied since it requires constituent-like parses for the target side in our string-to-tree set- ting. However, for the target languages discussed here (Polish and Russian), only dependency parses are readily available. Those parses relate the lexi- cal items of the sentence via edges that are labeled with the syntactic function between the head and its dependent. Overall, these structures also form trees, but they are often non-projective for our tar- get languages. Such non-projective dependency trees do not admit a constituent-like tree represen- tation, so we first need to convert them into pro- jective dependency trees, which can be converted easily into a constituent-like tree representation. The conversion into projective dependency trees is known to preserve discontinuities, so we expect that out model is …

Content truncated to 500 words