Regenerating Hypotheses for Statistical Machine Translation

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 105–112 Manchester, August 2008 Regenerating Hypotheses for Statistical Machine Translation Boxing Chen, Min Zhang, Aiti Aw and Haizhou Li Department of Human Language Technology Institute for Infocomm Research 21 Heng Mui Keng Terrace, 119613, Singapore {bxchen, mzhang, aaiti, hli}@i2r.a-star.edu.sg Abstract This paper studies three techniques that improve the quality of N-best hypotheses through additional regeneration process. Unlike the multi-system consensus ap- proach where multiple translation sys- tems are used, our improvement is achieved through the expansion of the N- best hypotheses from a single system. We explore three different methods to im- plement the regeneration process: re- decoding, n-gram expansion, and confu- sion network-based regeneration. Ex- periments on Chinese-to-English NIST and IWSLT tasks show that all three methods obtain consistent improvements. Moreover, the combination of the three strategies achieves further improvements and outperforms the baseline by 0.81 BLEU-score on IWSLT’06, 0.57 on NIST’03, 0.61 on NIST’05 test set re- spectively. 1 Introduction State-of-the-art Statistical Machine Translation (SMT) systems usually adopt a two-pass search strategy (Och, 2003; Koehn, et al., 2003) as shown in Figure 1. In the first pass, a decoding algorithm is applied to generate an N-best list of translation hypotheses, while in the second pass, the final translation is selected by rescoring and re-ranking the N-best translations through addi- tional feature functions. The fundamental as- sumption behind using a second pass is that the generated N-best list may contain better transla- © 2008. Licensed under the Creative Commons Attri- bution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc- sa/3.0/). Some rights reserved. tions than the best choice found by the decoder. Therefore, the performance of a two-pass SMT system can be improved from two aspects, i.e. scoring models and the quality of the N-best hy- potheses. Rescoring pass improves the performance of machine translation by enhancing the scoring models with more global sophisticated and dis- criminative feature functions. The idea for apply- ing two passes instead of one is that some global feature functions cannot be easily decomposed into local scores and computed during decoding. Furthermore, rescoring allows some feature func- tions, such as word and n-gram posterior prob- abilities, to be estimated on the N-best list (Uef- fing, 2003; Chen et al., 2005; Zens and Ney, 2006). In this two-pass method, translation perform- ance hinges on the N-best hypotheses that are generated in the first pass (since rescoring occurs on these), so adding the translation candidates generated by other MT systems to these hypothe- ses could potentially improve the performance. This technique is called system combination (Bangalore et al., 2001; Matusov et al., 2006; Sim et al., 2007; Rosti et al., 2007a; Rosti et al., 2007b). We have instead chosen to regenerate new hy- potheses from the original N-best list, a tech- nique which we call regeneration. Regeneration is an intermediate pass between decoding and rescoring as depicted in Figure 2. Given the original N-best list (N-best1) generated by the decoder, this regeneration pass creates new trans- lation …

Content truncated to 500 words