Combining Multi-Engine Translations with Moses

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 42–46, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics Combining Multi-Engine Translations with Moses Yu Chen1, Michael Jellinghaus1, Andreas Eisele1,2,Yi Zhang1,2, Sabine Hunsicker1, Silke Theison1, Christian Federmann2, Hans Uszkoreit1,2 1: Universit¨at des Saarlandes, Saarbr¨ucken, Germany 2: Deutsches Forschungszentrum f¨ur K¨unstliche Intelligenz GmbH, Saarbr¨ucken, Germany {yuchen,micha,yzhang,sabineh,sith}@coli.uni-saarland.de {eisele,cfedermann,uszkoreit}@dfki.de Abstract We present a simple method for generating translations with the Moses toolkit (Koehn et al., 2007) from existing hypotheses pro- duced by other translation engines. As the structures underlying these translation engines are not known, an evaluation- based strategy is applied to select sys- tems for combination. The experiments show promising improvements in terms of BLEU. 1 Introduction With the wealth of machine translation systems available nowadays (many of them online and for free), it makes increasing sense to investigate clever ways of combining them. Obviously, the main objective lies in finding out how to integrate the respective advantages of different approaches: Statistical machine translation (SMT) and rule- based machine translation (RBMT) systems of- ten have complementary characteristics. Previous work on building hybrid systems includes, among others, approaches using reranking, regeneration with an SMT decoder (Eisele et al., 2008; Chen et al., 2007), and confusion networks (Matusov et al., 2006; Rosti et al., 2007; He et al., 2008). The approach by (Eisele et al., 2008) aimed specifically at filling lexical gaps in an SMT sys- tem with information from a number of RBMT systems. The output of the RBMT engines was word-aligned with the input, yielding a total of seven phrase tables which where simply concate- nated to expand the phrase table constructed from the training corpus. This approach differs from the confusion network approaches mainly in that the final hypotheses do not necessarily follow any of the input translations as the skeleton. On the other hand, it emphasizes that the additional translations should be produced by RBMT systems with lexi- cons that cannot be learned from the data. The present work continues on the same track as the paper mentioned above but implements a number of important changes, most prominently a relaxation of the restrictions on the number and type of input systems. These differences are de- scribed in more detail in Section 2. Section 3 ex- plains the implementation of our system and Sec- tion 4 its application in a number of experiments. Finally, Section 5 concludes this paper with a sum- mary and some thoughts on future work. 2 Integrating Multiple Systems of Unknown Type and Quality When comparing (Eisele et al., 2008) to the present work, our proposal is more general in a way that the requirement for knowledge about the systems is minimum. The types and the identities of the participated systems are assumed unknown. Accordingly, we are not able to restrict ourselves to a certain class of systems as (Eisele et al., 2008) did. We rely on a standard phrase-based SMT framework to extract the valuable pieces from the system outputs. …

Content truncated to 500 words