The RWTH machine translation system for IWSLT 2007

The RWTH Machine Translation System for IWSLT 2007 Arne Mauser, David Vilar, Gregor Leusch, Yuqi Zhang and Hermann Ney Human Language Technology and Pattern Recognition Lehrstuhl f¨ur Informatik 6, Computer Science Department RWTH Aachen University, D-52056 Aachen, Germany {mauser,vilar,leusch,yzhang,ney}@cs.rwth-aachen.de Abstract The RWTH system for the IWSLT 2007 evaluation is a combination of several statistical machine translation sys- tems. The combination includes Phrase-Based models, a n-gram translation model and a hierarchical phrase model. We describe the individual systems and the method that was used for combining the system outputs. Compared to our 2006 system, we newly introduce a hierarchical phrase-based translation model and show improvements in system combi- nation for Machine Translation. RWTH participated in the Italian-to-English and Chinese-to-English translation direc- tions. 1. Introduction The RWTH system for the evaluation campaign of the International Workshop on Spoken Language Translation (IWSLT) 2007 is a combination of different statistical ma- chine translation systems: several phrase-based, an n-gram based and a hierarchical MT system. In this paper, we de- scribe the individual systems and the algorithm used to com- bine them. In the evaluation, the system participated in the Italian- to-English and Chinese-to-English translation directions. It was ranked first for Italian-to-English correct transcription and second for automatic recognition output. For Chinese- to-English, the primary system was ranked fourth, the best secondary system was ranked second. The following section describes the statistical approach to machine translation. Section 3 forms the main part of this paper and gives a detailed description of the models and submodels used in the combination. The method for system combination is described in Section 4. The task, system setup and results are discussed in Sections 5, 6 and 7 separately for the two translation directions. Finally, Section 9 presents the conclusions drawn from the 2007 IWSLT evaluation. 2. Statistical Machine Translation In statistical machine translation , we are given a source lan- guage sentence f J 1 = f1 . . . fj . . . fJ, which is to be trans- lated into a target language sentence eI 1 = e1 . . . ei . . . eI. Among all possible target language sentences, we will choose the sentence with the highest translation probability: ˆe 1ˆI = argmax I,eI 1 � Pr(eI 1|f J 1 ) � (1) We model this probability directly using a log-linear model: Pr(eI 1|f J 1 ) = exp ��M m=1 λmhm(eI 1, f J 1 ) � � e′I′ 1 exp ��M m=1 λmhm(e′I′ 1 , f J 1 ) � (2) This equation can be considered as a generalization of the source-channel approach presented in [1]. The hm(·) rep- resent feature functions which can be bilingual, and thus represent the correspondence between source and target lan- guage, or monolingual, which represent additional features like grammaticality of the output. Typically, the features are statistical models or simple heuristics. This approach has the advantage that additional models h(·) can be easily integrated into the overall system. The model scaling factors λM 1 are trained …

Content truncated to 500 words