The RWTH System Combination System for WMT 2009

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 51–55, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics The RWTH System Combination System for WMT 2009 Gregor Leusch, Evgeny Matusov, and Hermann Ney RWTH Aachen University Aachen, Germany Abstract RWTH participated in the System Combi- nation task of the Fourth Workshop on Sta- tistical Machine Translation (WMT 2009). Hypotheses from 9 German→English MT systems were combined into a consen- sus translation. This consensus transla- tion scored 2.1% better in BLEU and 2.3% better in TER (abs.) than the best sin- gle system. In addition, cross-lingual output from 10 French, German, and Spanish→English systems was combined into a consensus translation, which gave an improvement of 2.0% in BLEU/3.5% in TER (abs.) over the best single system. 1 Introduction The RWTH approach to MT system combination is a refined version of the ROVER approach in ASR (Fiscus, 1997), with additional steps to cope with reordering between different hypotheses, and to use true casing information from the input hy- potheses. The basic concept of the approach has been described by Matusov et al. (2006). Several improvements have been added later (Matusov et al., 2008). This approach includes an enhanced alignment and reordering framework. In con- trast to existing approaches (Jayaraman and Lavie, 2005; Rosti et al., 2007), the context of the whole corpus rather than a single sentence is considered in this iterative, unsupervised procedure, yielding a more reliable alignment. Majority voting on the generated lattice is performed using the prior prob- abilities for each system as well as other statistical models such as a special n-gram language model. 2 System Combination Algorithm In this section we present the details of our system combination method. Figure 1 gives an overview of the system combination architecture described in this section. After preprocessing the MT hy- potheses, pairwise alignments between the hy- potheses are calculated. The hypotheses are then reordered to match the word order of a selected primary hypothesis. From this, we create a confu- sion network (CN), which we then rescore using Figure 1: The system combination architecture. system prior weights and a language model (LM). The single best path in this CN then constitutes the consensus translation. 2.1 Word Alignment The proposed alignment approach is a statistical one. It takes advantage of multiple translations for a whole corpus to compute a consensus translation for each sentence in this corpus. It also takes ad- vantage of the fact that the sentences to be aligned are in the same language. For each source sentence F in the test corpus, we select one of its translations En, n=1, . . . , M, as the primary hypothesis. Then we align the sec- ondary hypotheses Em(m = 1, . . . , M; n ̸= m) with En to match the word order in En. Since it is not clear which hypothesis should be primary, i. e. has the “best” word order, we let every hypothesis play the role …

Content truncated to 500 words