Feature-Rich Discriminative Phrase Rescoring for SMT

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 492–500, Beijing, August 2010 Feature-Rich Discriminative Phrase Rescoring for SMT Fei Huang and Bing Xiang IBM T. J. Watson Research Center {huangfe, bxiang}@us.ibm.com Abstract This paper proposes a new approach to phrase rescoring for statistical machine translation (SMT). A set of novel fea- tures capturing the translingual equiva- lence between a source and a target phrase pair are introduced. These features are combined with linear regression model and neural network to predict the quality score of the phrase translation pair. These phrase scores are used to dis- criminatively rescore the baseline MT system’s phrase library: boost good phrase translations while prune bad ones. This approach not only significantly im- proves machine translation quality, but also reduces the model size by a consid- erable margin. 1 Introduction Statistical Machine Translation (SMT) systems, including phrase-based (Och and Ney 2002; Koehn et. al. 2003), syntax-based (Yamada and Knight 2001; Galley et. al. 2004) or hybrid sys- tems (Chiang 2005; Zollmann and Venugopal 2006), are typically built with bilingual phrase pairs, which are extracted from parallel sentences with word alignment. Due to the noises in the bilingual sentence pairs and errors from auto- matic word alignment, the extracted phrase pairs may contain errors, such as • dropping content words (the $num countries ,||个:<null>), • length mismatch • (along the lines of the || 的:of) content irrelevance (the next $num years, || 水平:level 方面:aspect 所:<null>) These incorrect phrase pairs compete with cor- rect phrase pairs during the decoding process, and are often selected when their counts are high (if they contain systematic alignment errors) or certain model costs are low (for example, when some source content words are translated into target function words in an incorrect phrase pair, the language model cost of the incorrect pair may be small, making it more likely that the pair will be selected for the final translation). As a result, the translation quality is degraded when these incorrect phrase pairs are selected. Various approaches have been proposed over the past decade for the purpose of improving the phrase pair quality for SMT. For example, a term weight based model was presented in (Zhao, et al., 2004) to rescore phrase translation pairs. It models the translation probability with similari- ties between the query (source phrase) and document (target phrase). Significant improve- ment was obtained in the translation performance. In (Johnson, et al., 2007; Yang and Zheng, 2009), a statistical significance test was used to heavily prune the phrase table and thus achieved higher precision and better MT performance. In (Deng, et al., 2008), a generic phrase train- ing algorithm was proposed with the focus on phrase extraction. Multiple feature functions are utilized based on information metrics or word alignment. The feature parameters are optimized to directly maximize the end-to-end system per- formance. Significant improvement was reported for a small MT task. But when the phrase table is large, such as in a large-scale SMT system, the computational cost …

Content truncated to 500 words