Fast Consensus Hypothesis Regeneration for Machine Translation

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 11–16, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Fast Consensus Hypothesis Regeneration for Machine Translation Boxing Chen, George Foster and Roland Kuhn National Research Council Canada 283 Alexandre-Taché Boulevard, Gatineau (Québec), Canada J8X 3X7 {Boxing.Chen, George.Foster, Roland.Kuhn}@nrc.ca Abstract This paper presents a fast consensus hy- pothesis regeneration approach for ma- chine translation. It combines the advan- tages of feature-based fast consensus de- coding and hypothesis regeneration. Our approach is more efficient than previous work on hypothesis regeneration, and it explores a wider search space than con- sensus decoding, resulting in improved performance. Experimental results show consistent improvements across language pairs, and an improvement of up to 0.72 BLEU is obtained over a competitive single-pass baseline on the Chinese-to- English NIST task. 1 Introduction State-of-the-art statistical machine translation (SMT) systems are often described as a two-pass process. In the first pass, decoding algorithms are applied to generate either a translation N-best list or a translation forest. Then in the second pass, various re-ranking algorithms are adopted to compute the final translation. The re-ranking al- gorithms include rescoring (Och et al., 2004) and Minimum Bayes-Risk (MBR) decoding (Kumar and Byrne, 2004; Zhang and Gildea, 2008; Tromble et al., 2008). Rescoring uses more so- phisticated additional feature functions to score the hypotheses. MBR decoding directly incorpo- rates the evaluation metrics (i.e., loss function), into the decision criterion, so it is effective in tuning the MT performance for a specific loss function. In particular, sentence-level BLEU loss function gives gains on BLEU (Kumar and Byrne, 2004). The naïve MBR algorithm computes the loss function between every pair of k hypotheses, needing O(k2) comparisons. Therefore, only small number k is applicable. Very recently, De- Nero et al. (2009) proposed a fast consensus de- coding (FCD) algorithm in which the similarity scores are computed based on the feature expec- tations over the translation N-best list or transla- tion forest. It is equivalent to MBR decoding when using a linear similarity function, such as unigram precision. Re-ranking approaches improve performance on an N-best list whose contents are fixed. A complementary strategy is to augment the con- tents of an N-best list in order to broaden the search space. Chen et al (2008) have proposed a three-pass SMT process, in which a hypothesis regeneration pass is added between the decoding and rescoring passes. New hypotheses are gener- ated based on the original N-best hypotheses through n-gram expansion, confusion-network decoding or re-decoding. All three hypothesis regeneration methods obtained decent and com- parable improvements in conjunction with the same rescoring model. However, since the final translation candidates in this approach are pro- duced from different methods, local feature func- tions (such as translation models and reordering models) of each hypothesis are not directly com- parable and rescoring must exploit rich global feature functions to compensate for the loss of local feature functions. Thus this approach is de- pendent on the use of computationally expensive …

Content truncated to 500 words