A systematic evaluation of MBOT in statistical machine translation

A Systematic Evaluation of MBOT in Statistical Machine Translation Nina Seemann seemanna@ims.uni-stuttgart.de Fabienne Braune braunefe@ims.uni-stuttgart.de Andreas Maletti Institute for Natural Language Processing, University of Stuttgart,maletti@ims.uni-stuttgart.de Pfaffenwaldring 5b, 70569 Stuttgart, Germany Abstract Shallow local multi-bottom up tree transducers (MBOTs) have been successfully used as trans- lation models in several settings because of their ability to model discontinuities. In this con- tribution, several additional settings are explored and evaluated. The first rule extractions for tree-to-tree MBOT with non-minimal rules and for string-to-string MBOT are developed. All existing MBOT systems are systematically evaluated and compared to corresponding base- line systems in three large-scale translation tasks: English-to-German, English-to-Chinese, and English-to-Arabic. Particular emphasis is placed on the use of discontinuous rules. The devel- oped rule extractions and analysis tools will be made publicly available. 1 Introduction The area of statistical machine translation (SMT) (Koehn, 2009) concerns itself mostly with the development of automatic methods of deriving probabilistic translation models from a parallel corpus. Such a corpus contains corresponding sentences in two languages. The rule extrac- tion mechanism automatically extracts rules for the translation model from such a corpus and assigns them a probability. Several different translation approaches are currently discussed in SMT. Among the best-performing systems one often finds phrase-based (Koehn et al., 2003) and syntax-based translation systems. Phrase-based systems use no linguistic information and can be obtained directly from the corpus. The same observation is true for hierarchical phrase- based systems (Chiang, 2005) as those are syntactic in a formal sense only. Syntax-based systems require some form of syntactic annotation, which is often represented as a (parse) tree. Correspondingly, we obtain several models, namely tree-to-tree, string-to-tree, and tree- to-string systems, which use trees on the source or the target language side. A multitude of for- malisms has been proposed as syntax-based translation models. The most prominent might be the synchronous tree substitution grammars (STSGs) of Eisner (2003) and the non-contiguous synchronous tree sequence substituition grammars (STSSGs) of Sun et al. (2009). Recently, Maletti (2011) proposed local multi bottom-up tree transducers (MBOT) as a translation model for syntax-based SMT. An MBOT is an extension of an STSG that allows sequences of tree fragments on the target side of its rules. In this manner, it can model discontinuities. It can also be understood as a restricted form of a STSSG, in which the rules consist of sequences of source and target tree fragments. Recently, MBOTs have been implemented as a translation model inside the Moses frame- work (Koehn et al., 2007) by Braune et al. (2013). Initially, only the rule extraction for minimal Proceedings of MT Summit XV, vol.1: MT Researchers' Track Miami, Oct 30 - Nov 3, 2015 | p. 200 tree-to-tree rules of Maletti (2011) was available. In combination with synchronous context-free grammar (SCFG) rules, already this system led to significant improvements over a correspond- ing system using SCFG rules only. Later, Seemann et al. (2015) proposed a rule extraction procedure for string-to-tree MBOTs that is able to extract non-minimal rules. Since the number …

Content truncated to 500 words