DFKI Hybrid Machine Translation System for WMT 2011 - On the Integration of SMT and RBMT

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 485–489, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics DFKI Hybrid Machine Translation System for WMT 2011 - On the Integration of SMT and RBMT Jia Xu and Hans Uszkoreit and Casey Kennington and David Vilar and Xiaojun Zhang DFKI GmbH, Language Technology Lab Stuhlsatzenhausweg 3 D-66123 Saarbr¨ucken Germany {Jia.Xu,uszkoreit,David.Vilar}@dfki.de, {bakuzen,xiaojun.zhang.iiken}@gmail.com Abstract We present the DFKI hybrid translation sys- tem at the WMT workshop 2011. Three SMT and two RBMT systems are combined at the level of the final translation output. The trans- lation results show that our hybrid system sig- nificantly outperformed individual systems by exploring strengths of both rule-based and sta- tistical translations. 1 Introduction Machine translation (MT), in particular the statisti- cal approach to it, has undergone incremental im- provements in recent years. While rule-based ma- chine translation (RBMT) maintains competitive- ness in human evaluations. Combining the advan- tages of both approaches have been investigated by many researchers such as (Eisele et al., 2008). Nonetheless, significant improvements over statis- tical approaches still remain to be shown. In this paper, we present the DFKI hybrid system in the WMT workshop 2011. Our system is different from the system of the last year (Federmann et al., 2010), which is based on the shallow phrase substitution. In this work, two rule-based translation systems are applied. In addition, three statistical machine trans- lation systems are built, including a phrase-based, a hierarchical phrase-based and a syntax-based sys- tem. Instead of combining with rules or post-editing, we perform system combination on the final transla- tion hypotheses. We applied the CMU open toolkit (Heafield and Lavie, 2010) among numerous com- bination methods such as (Matusov, 2009), (Sim et al., 2007) and (He et al., 2008). The final transla- tion output outperforms each individual output sig- nificantly. 2 Individual translation systems 2.1 Phrase-based system We use the IBM model 1 and 4 (Brown et al., 1993) and Hidden-Markov model (HMM) (Vogel et al., 1996) to train the word alignment using the mgiza toolkit1. We applied the EMS in Moses (Koehn et al., 2007) to build up the phrase-based translation system. Features in the log-linear model include translation models in two directions, a language model, a distortion model and a sentence length penalty. A dynamic programming beam search al- gorithm is used to generate the translation hypoth- esis with maximum probability. We applied a 5- gram mixture language model with each sub-model trained on one fifth of the monolingual corpus with Kneser-Ney smoothing using SRILM toolkit (Stol- cke, 2002). We did not perform any tuning, because it hurts the evaluation performance in our experi- ments. 2.2 Syntax-based system To capture the syntactic structure, we also built a tree-based system using the same configuration of EMS in Moses (Koehn et al., 2007). Tree-based models operate on so-called grammar rules, which include variables in the mapping rules. To increase the diversity of models in combination, the lan- guage model in each individual translation system is …

Content truncated to 500 words