The RWTH System Combination System for WMT 2011

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 152–158, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics The RWTH System Combination System for WMT 2011 Gregor Leusch, Markus Freitag, and Hermann Ney RWTH Aachen University Aachen, Germany {leusch,freitag,ney}@cs.rwth-aachen.de Abstract RWTH participated in the System Combi- nation task of the Sixth Workshop on Sta- tistical Machine Translation (WMT 2011). For three language pairs, we combined 6 to 14 systems into a single consen- sus translation. A three-level meta- combination scheme combining six dif- ferent system combination setups with three different engines was applied on the French–English language pair. Depend- ing on the language pair, improvements versus the best single system are in the range of +1.9% and +2.5% abs. on BLEU, and between −1.8% and −2.4% abs. on TER. Novel techniques compared with RWTH’s submission to WMT 2010 include two additional system combina- tion engines, an additional word alignment technique, meta combination, and addi- tional optimization techniques. 1 Introduction RWTH’s main approach to System Combination (SC) for Machine Translation (MT) is a refined version of the ROVER approach in Automatic Speech Recognition (ASR) (Fiscus, 1997), with additional steps to cope with reordering between different hypotheses, and to use true casing infor- mation from the input hypotheses. The basic con- cept of the approach has been described by Ma- tusov et al. (2006). Several improvements have been added later (Matusov et al., 2008). This ap- proach includes an enhanced alignment and re- ordering framework. In contrast to existing ap- proaches (Jayaraman and Lavie, 2005; Rosti et al., 2007b), the context of the whole corpus rather than a single sentence is considered in this itera- tive, unsupervised procedure, yielding a more reli- able alignment. Majority voting on the generated lattice is performed using prior weights for each system as well as other statistical models such as a special n-gram language model. True cas- ing is considered a separate step in RWTH’s ap- proach, which also takes the input hypotheses into account. The pipeline, and consequently the de- scription of the main pipeline given in this paper, is based on our pipeline for WMT 2010 (Leusch and Ney, 2010), with extensions as described. When necessary, we denote this pipeline as Align-to- Lattice, or A2L . For the French–English task, we used two ad- ditional system combination engines for the first time: The first one uses the same alignments as A2L, but generates lattices in the OpenFST frame- work (Allauzen et al., 2007). The OpenFST de- coder (fstshortestpath) is then used to find the best path (consensus translation) in this lattice. Analogously, we call this engine A2FST . The sec- ond additional engine, which we call SCUNC, uses a TER-based alignment, similar to the approach by Rosti et al. (2007b). Instead of a lattice rescor- ing, finding the consensus translation is consid- ered a per-node classification problem: For each slot, which one is the “correct” one (i.e. will give the “best” output)? This approach is inspired by iROVER (Hillard et …

Content truncated to 500 words