Mixture Model-based Minimum Bayes Risk Decoding using Multiple Machine Translation Systems

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 313–321, Beijing, August 2010 Mixture Model-based Minimum Bayes Risk Decoding using Multiple Machine Translation Systems Nan Duan1 School of Computer Science and Technology Tianjin University v-naduan@microsoft.com Mu Li, Dongdong Zhang, Ming Zhou Microsoft Research Asia muli@microsoft.com dozhang@microsoft.com mingzhou@microsoft.com Abstract We present Mixture Model-based Mini- mum Bayes Risk (MMMBR) decoding, an approach that makes use of multiple SMT systems to improve translation ac- curacy. Unlike existing MBR decoding methods defined on the basis of single SMT systems, an MMMBR decoder re- ranks translation outputs in the combined search space of multiple systems using the MBR decision rule and a mixture dis- tribution of component SMT models for translation hypotheses. MMMBR decod- ing is a general method that is indepen- dent of specific SMT models and can be applied to various commonly used search spaces. Experimental results on the NIST Chinese-to-English MT evaluation tasks show that our approach brings significant improvements to single system-based MBR decoding and outperforms a state- of-the-art system combination method. 1 1 Introduction Minimum Bayes Risk (MBR) decoding is be- coming more and more popular in recent Statis- tical Machine Translation (SMT) research. This approach requires a second-pass decoding pro- cedure to re-rank translation hypotheses by risk scores computed based on model’s distribution. Kumar and Byrne (2004) first introduced MBR decoding to SMT field and developed it on the N-best list translations. Their work has shown that MBR decoding performs better than Maximum a Posteriori (MAP) decoding for dif- ferent evaluation criteria. After that, many dedi- 1 This work has been done while the author was visiting Microsoft Research Asia. cated efforts have been made to improve the per- formances of SMT systems by utilizing MBR- inspired methods. Tromble et al. (2008) pro- posed a linear approximation to BLEU score (log-BLEU) as a new loss function in MBR de- coding and extended it from N-best lists to lat- tices, and Kumar et al. (2009) presented more efficient algorithms for MBR decoding on both lattices and hypergraphs to alleviate the high computational cost problem in Tromble et al.’s work. DeNero et al. (2009) proposed a fast con- sensus decoding algorithm for MBR for both linear and non-linear similarity measures. All work mentioned above share a common setting: an MBR decoder is built based on one and only one MAP decoder. On the other hand, recent research has shown that substantial im- provements can be achieved by utilizing consen- sus statistics over multiple SMT systems (Rosti et al., 2007; Li et al., 2009a; Li et al., 2009b; Liu et al., 2009). It could be desirable to adapt MBR decoding to multiple SMT systems as well. In this paper, we present Mixture Model- based Minimum Bayes Risk (MMMBR) decoding, an approach that makes use of multiple SMT systems to improve translation performance. In this work, we can take advantage of a larger search space for hypothesis selection, and em- ploy an improved probability distribution over translation hypotheses based on …

Content truncated to 500 words