Description of the JHU System Combination Scheme for WMT 2011

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 171–176, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Description of the JHU System Combination Scheme for WMT 2011 Daguang Xu Johns Hopkins University Baltimore, USA dxu5@jhu.edu Yuan Cao Johns Hopkins University Baltimore, USA yuan.cao@jhu.edu Damianos Karakos Johns Hopkins University Baltimore, USA damianos@jhu.edu Abstract This paper describes the JHU system combi- nation scheme used in WMT-11. The JHU system combination is based on confusion network alignment, and inherited the frame- work developed by (Karakos et al., 2008). We improved our core system combination al- gorithm by making use of TER-plus, which was originally designed for string alignment, for alignment of confusion networks. Exper- imental results on French-English, German- English, Czech-English and Spanish-English combination tasks show significant improve- ments on BLEU and TER by up to 2 points on average, compared to the best individual sys- tem output, and improvements compared with the results produced by ITG which we used in WMT-10. 1 Introduction System combination aims to improve the translation quality by combining the outputs from multiple in- dividual MT systems. The state-of-the-art system combination methodologies can be roughly catego- rized as follows (Karakos et al., 2010): 1. Confusion network based: confusion network is a form of lattice with the constraint that all paths need to pass through all nodes. An exam- ple of a confusion network is shown in Figure 1. Here, the set of arcs between two consecutive nodes represents a bin, the number following a word is the count of this word in its bin, and 0 1 this/10 2 is/7 was/3 3 a/8 one/2 4 dog/9 cat/1 5/0 ./10 Figure 1: Example confusion network. The total count in each bin is 10. each bin has the same size. The basic method- ology of system combination based on confu- sion network includes the following steps: (a) Choose one system output as the “skeleton”, which roughly decides the word order. (b) Align further system outputs to the skeleton, thus forming a confusion network. (c) Rescore the final confusion network using a language model, then pick the best path as the output of combination. A textual representation (where each line con- tains the words and counts of each bin) is usu- ally the most convenient for machine process- ing. 2. Joint optimization based: unlike building con- fusion network, this method considers all sys- tem outputs at once instead of incrementally. Then a log-linear model is used to derive costs, followed by a search algorithm to explore the combination space (Jayaraman et al., 2005; Heafield et al., 2009; He et al., 2009). 3. Hypothesis selection based: this method only includes algorithms that output one of the input translations, and no word selection from mul- tiple systems is performed. Typical algorithms can be found in (Rosti et al., 2007). 171 This paper describes the JHU system com- bination submitted to the Sixth Workshop on Statistical Machine Translation (WMT-11) (http://statmt.org/wmt11/index.html ). The JHU system combination is confusion network based …

Content truncated to 500 words