CMU System Combination via Hypothesis Selection for WMT’10

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 307–310, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics CMU System Combination via Hypothesis Selection for WMT’10 Almut Silja Hildebrand Carnegie Mellon University Pittsburgh, USA silja@cs.cmu.edu Stephan Vogel Carnegie Mellon University Pittsburgh, USA vogel@cs.cmu.edu Abstract This paper describes the CMU entry for the system combination shared task at WMT’10. Our combination method is hy- pothesis selection, which uses information from n-best lists from the input MT sys- tems, where available. The sentence level features used are independent from the MT systems involved. Compared to the baseline we added source-to-target word alignment based features and trained sys- tem weights to our feature set. We com- bined MT systems for French - English and German - English using provided data only. 1 Introduction For the combination of machine translation sys- tems there have been several approaches described in recent publications. One uses confusion net- works formed along a skeleton sentence to com- bine translation systems as described in (Rosti et al., 2008) and (Karakos et al., 2008). A different approach described in (Heafield et al., 2009) is not keeping the skeleton fixed when aligning the sys- tems. Another approach selects whole hypotheses from a combined n-best list (Hildebrand and Vo- gel, 2008). Our setup follows the latter approach. We com- bine the output from the submitted translation sys- tems, including n-best lists where available, into one joint n-best list, then calculate a set of fea- tures consistently for all hypotheses. We use MER training on the provided development data to de- termine feature weights and re-rank the joint n- best list. We train to maximize BLEU. 2 Features For our entries to the WMT’09 we used the follow- ing feature groups (in parenthesis are the number of separate feature values per group): • Language model scores (3) • Word lexicon scores (6) • Sentence length features (3) • Rank feature (1) • Normalized n-gram agreement (6) • Source-target word alignment features (6) • Trained system weights (no. of systems) The details on language model and word lexi- con scores can be found in (Hildebrand and Vogel, 2008) and details on the rank feature and the nor- malized n-gram agreement can be found in (Hilde- brand and Vogel, 2009). We use three sentence length features, which are the ratio of the hypoth- esis length to the length of the source sentence, the diversion of this ratio from the overall length ratio of the bilingual training data and the differ- ence between the hypothesis length and the av- erage length of the hypotheses in the n-best list for the respective source sentence. The system weights are trained together with the other feature weights during MERT using a binary feature per system. To the feature vector for each hypothe- sis one feature per input system is added; for each hypothesis one of the features is one, indicating which system it came from, all others are zero. 2.1 Source-Target Word Alignment …

Content truncated to 500 words