CMU Multi-Engine Machine Translation for WMT 2010

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 301–306, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics CMU Multi-Engine Machine Translation for WMT 2010 Kenneth Heafield Carnegie Mellon University Pittsburgh, PA, USA. heafield@cs.cmu.edu Alon Lavie Carnegie Mellon University Pittsburgh, PA, USA. alavie@cs.cmu.edu Abstract This paper describes our submission, cmu-heafield-combo, to the WMT 2010 machine translation system combi- nation task. Using constrained resources, we participated in all nine language pairs, namely translating English to and from Czech, French, German, and Spanish as well as combining English translations from multiple languages. Combination proceeds by aligning all pairs of system outputs then navigating the aligned out- puts from left to right where each path is a candidate combination. Candidate com- binations are scored by their length, agree- ment with the underlying systems, and a language model. On tuning data, improve- ment in BLEU over the best system de- pends on the language pair and ranges from 0.89% to 5.57% with mean 2.37%. 1 Introduction System combination merges the output of sev- eral machine translation systems into a sin- gle improved output. Our system combina- tion scheme, submitted to the Workshop on Sta- tistical Machine Translation (WMT) 2010 as cmu-heafield-combo, is an improvement over our previous system (Heafield et al., 2009), called cmu-combo in WMT 2009. The scheme consists of aligning 1-best outputs from each sys- tem using the METEOR (Denkowski and Lavie, 2010) aligner, identifying candidate combinations by forming left-to-right paths through the aligned system outputs, and scoring these candidates us- ing a battery of features. Improvements this year include unigram paraphrase alignment, support for all target languages, new features, language mod- eling without pruning, and more parameter opti- mization. This paper describes our scheme with emphasis on improved areas. 2 Related Work Confusion networks (Rosti et al., 2008) are the most popular form of system combination. In this approach, a single system output acts as a back- bone to which the other outputs are aligned. This backbone determines word order while other out- puts vote for substitution, deletion, and insertion operations. Essentially, the backbone is edited to produce a combined output which largely pre- serves word order. Our approach differs in that we allow paths to switch between sentences, effec- tively permitting the backbone to switch at every word. Other system combination techniques typically use TER (Snover et al., 2006) or ITGs (Karakos et al., 2008) to align system outputs, meaning they depend solely on positional information to find approximate matches; we explicitly use stem, synonym, and paraphrase data to find alignments. Our use of paraphrases is similar to Leusch et al. (2009), though they learn a monolingual phrase table while we apply cross-lingual pivoting (Ban- nard and Callison-Burch, 2005). 3 Alignment System outputs are aligned at the token level using a variant of the METEOR (Denkowski and Lavie, 2010) aligner. This identifies, in decreasing order of priority: exact, stem, synonym, and unigram paraphrase matches. Stems (Porter, 2001) are available for all languages except Czech, …

Content truncated to 500 words