The UZH System Combination System for WMT 2011

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 166–170, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics The UZH System Combination System for WMT 2011 Rico Sennrich Institute of Computational Linguistics University of Zurich Binzm¨uhlestr. 14 CH-8050 Z¨urich sennrich@cl.uzh.ch Abstract This paper describes the UZH system that was used for the WMT 2011 system combination shared task submission. We participated in the system combination task for the translation directions DE–EN and EN–DE. The system uses Moses as a backbone, with the outputs of the 2–3 best individual systems being inte- grated through additional phrase tables. The system compares well to other system com- bination submissions, with no other submis- sion being significantly better. A BLEU-based comparison to the individual systems, how- ever, indicates that it achieves no significant gains over the best individual system. 1 Introduction For our submission to the WMT 2011 shared task, we built a system with the multi-engine MT ap- proach described in (Sennrich, 2011), which builds on the idea by (Chen et al., 2007). A Moses SMT system (Koehn et al., 2007) is used as a backbone, trained on the WMT 2011 training data. Translation hypotheses by other systems are integrated through a second phrase table. In this second phrase ta- ble, the phrase translation probabilities and lexical weights are computed based on the word and phrase frequencies in both the translation hypotheses and a parallel training corpus. On the evaluation data in (Sennrich, 2011), this system significantly outper- formed MEMT (Heafield and Lavie, 2010), which was among the best-performing system combination tools at WMT 2010 (Callison-Burch et al., 2010). In this paper, we apply the same approach to a dif- ferent translation scenario, namely the WMT 2011 shared task. We fail to significantly outperform the best individual system in terms of BLEU score. In section 2, we describe our system combination ap- proach. In section 3, we present the results, and dis- cuss possible reasons why the system fails to show the same performance gains as in the translation task on which it was evaluated initially. 2 System Description We participated in the system combination task DE– EN and EN–DE. Since the combination is achieved by integrating translation hypotheses into an existing Moses system, which we will call the primary sys- tem, we first describe the methods and data used for training this primary system. Then, we describe how the translation hypotheses are selected out of the in- dividual system submissions and integrated into the Moses system. 2.1 Primary System For the training of the primary systems, we mostly followed the baseline instructions for the transla- tion task1. We use news-commentary and Europarl as parallel training data. The language models are a linear interpolation of the news-commentary, Eu- roparl and news corpora, optimized for minimal cross-entropy on the newstest2008 data sets in the respective target language. Additionally, we prune the primary phrase table using statistical significance tests, as described by (Johnson et al., 2007). For the translation direction DE–EN, …

Content truncated to 500 words