MATREX: The DCU MT System for WMT 2009

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 95–99, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics MATREX: The DCU MT System for WMT 2009 Jinhua Du, Yifan He, Sergio Penkale, Andy Way Centre for Next Generation Localisation Dublin City University Dublin 9, Ireland {jdu,yhe,spenkale,away}@computing.dcu.ie Abstract In this paper, we describe the machine translation system in the evaluation cam- paign of the Fourth Workshop on Statisti- cal Machine Translation at EACL 2009. We describe the modular design of our multi-engine MT system with particular focus on the components used in this par- ticipation. We participated in the translation task for the following translation directions: French–English and English–French, in which we employed our multi-engine ar- chitecture to translate. We also partic- ipated in the system combination task which was carried out by the MBR de- coder and Confusion Network decoder. We report results on the provided devel- opment and test sets. 1 Introduction In this paper, we present a multi-engine MT system developed at DCU, MATREX (Machine Translation using Examples). This system exploits EBMT, SMT and system combination techniques to build a cascaded translation framework. We participated in both the French–English and English-French News tasks. In these two tasks, we employ three individual MT system which are 1) Baseline: phrase-based system (PB); 2) EBMT: Monolingually chunking both source and target sides of the dataset using a marker-based chun- ker (Gough and Way, 2004). 3) HPB: a typical hierarchical phrase-based system (Chiang, 2005). Meanwhile, we also use a word-level combina- tion framework (Rosti et al., 2007) to combine the multiple translation hypotheses and employ a new rescoring model to generate the final result. For the system combination task, we first use the minimum Bayes-risk (MBR) (Kumar and Byrne, 2004) decoder to select the best hypothe- sis as the alignment reference for the Confusion Network (CN) (Mangu et al., 2000). We then build the CN using the TER metric (Snover et al., 2006), and finally search and generate the translation. The remainder of this paper is organised as fol- lows: Section 2 details the various components of our system, in particular the multi-engine strate- gies used for the shared task. In Section 3, we outline the complete system setup for the shared task and provide results on the development and test sets. Section 4 is our conclusion. 2 The MATREX System 2.1 System Architecture The MATREX system is a combination-based multi-engine architecture, which exploits aspects of both the EBMT and SMT paradigms. This architecture includes three individual sys- tems which are phrase-based, example-based and hierarchical phrase-based. The combination structure is the MBR decoder and CN decoder, which is based on the word-level combination strategy. In the final stage, we use a new rescoring mod- ule to process the N-best list generated by the combination module. See Figure 1 as a detailed illustration. 2.2 Example-Based Machine Translation EBMT obtains resources using the Marker Hy- pothesis (Green, 1979), a psycholinguistic con- straint which posits that all languages …

Content truncated to 500 words