Using Moses to Integrate Multiple Rule-Based Machine Translation Engines into a Hybrid System

Proceedings of the Third Workshop on Statistical Machine Translation, pages 179–182, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Using Moses to Integrate Multiple Rule-Based Machine Translation Engines into a Hybrid System Andreas Eisele1,2, Christian Federmann2, Herv´e Saint-Amand1, Michael Jellinghaus1, Teresa Herrmann1, Yu Chen1 1: Saarland University, Saarbr¨ucken, Germany 2: DFKI GmbH, Saarbr¨ucken, Germany Abstract Based on an architecture that allows to com- bine statistical machine translation (SMT) with rule-based machine translation (RBMT) in a multi-engine setup, we present new results that show that this type of system combination can actually increase the lexical coverage of the resulting hybrid system, at least as far as this can be measured via BLEU score. 1 Introduction (Chen et al., 2007) describes an architecture that allows to combine statistical machine translation (SMT) with one or multiple rule-based machine translation (RBMT) systems in a multi-engine setup. It uses a variant of standard SMT technology to align translations from one or more RBMT systems with the source text and incorporated phrases extracted from these alignments into the phrase table of the SMT system. Using this approach it is possible to employ a vanilla installation of the open-source de- coder Moses1 (Koehn et al., 2007) to find good com- binations of phrases from SMT training data with the phrases derived from RBMT. A similar method was presented in (Rosti et al., 2007). This setup provides an elegant solution to the fairly complex task of integrating multiple MT re- sults that may differ in word order using only stan- dard software modules, in particular GIZA++ (Och and Ney, 2003) for the identification of building blocks and Moses for the recombination, but the authors were not able to observe improvements in 1see http://www.statmt.org/moses/ terms of BLEU score. A closer investigation re- vealed that the experiments had suffered from a cou- ple of technical difficulties, such as mismatches in character encodings generated by different MT en- gines and similar problems. This motivated us to re-do these experiments in a somewhat more sys- tematic way for this year’s shared translation task, paying the required attention to all the technical de- tails and also to try it out on more language pairs. 2 System Architecture For conducting the translations, we use a multi- engine MT approach based on a ”vanilla” Moses SMT system with a modified phrase table as a cen- tral element. This modification is performed by aug- menting the standard phrase table with entries ob- tained from translating the data with several rule- based MT systems. The resulting phrase table thus combines statistically gathered phrase pairs with phrase pairs generated by linguistic rules. Basing its decision about the final translation on the obtained ”combined” phrase table, the SMT de- coder searches for the best translation by recombin- ing the building blocks that have been contributed by the different RBMT systems and the original SMT system trained on Europarl data. A sketch of the overall architecture is given in Fig. 1, where the lighter parts represent the mod- …

Content truncated to 500 words