CMU System Combination for WMT‘09

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 47–50, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics CMU System Combination for WMT’09 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’09. Our combination method is hy- pothesis selection, which uses information from n-best lists from several MT systems. The sentence level features are indepen- dent from the MT systems involved. To compensate for various n-best list sizes in the workshop shared task including first- best-only entries, we normalize one of our high-impact features for varying sub-list size. We combined restricted data track entries in French - English, German - En- glish and Hungarian - English using pro- vided data only. 1 Introduction For the combination of machine translation sys- tems there have been two main approaches de- scribed in recent publications. One uses confusion network decoding to combine translation systems as described in (Rosti et al., 2008) and (Karakos et al., 2008). The other approach selects whole hy- potheses from a combined n-best list (Hildebrand and Vogel, 2008). Our setup follows the approach described in (Hildebrand and Vogel, 2008). We combine the output from the available translation systems into one joint n-best list, then calculate a set of fea- tures consistently for all hypotheses. We use MER training on a development set to determine feature weights and re-rank the joint n-best list. 2 Features For our entries to the WMT’09 we used the fol- lowing feature groups: • Language model score • Word lexicon scores • Sentence length features • Rank feature • Normalized n-gram agreement The details on language model and word lexi- con scores can be found in (Hildebrand and Vogel, 2008). We use two sentence length features, which are the ratio of the hypothesis length to the length of the source sentence and the difference between the hypothesis length and the average length of the hypotheses in the n-best list for the respec- tive source sentence. We also use the rank of the hypothesis in the original system’s n-best list as a feature. 2.1 Normalized N-gram Agreement The participants of the WMT’09 shared transla- tion task provided output from their translation systems in various sizes. Most submission were 1st-best translation only, some submitted 10-best up to 300-best lists. In preliminary experiments we saw that adding a high scoring 1st-best translation to a joint n-best list composed of several larger n-best lists does not yield the desired improvement. This might be due to the fact, that hypotheses within an n-best list originating from one single system (sub-list) tend to be much more similar to each other than to hy- potheses from another system. This leads to hy- potheses from larger sub-lists scoring higher in the n-best list based features, e.g. because they collect more n-gram matches within their sub-list, which ”supports” them the more the larger it …

Content truncated to 500 words