Collaborative Decoding: Partial Hypothesis Re-ranking Using Translation Consensus between Decoders

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 585–592, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Collaborative Decoding: Partial Hypothesis Re-ranking Using Translation Consensus between Decoders Mu Li1, Nan Duan2, Dongdong Zhang1, Chi-Ho Li1, Ming Zhou1 1Microsoft Research Asia 2Tianjin University Beijing, China Tianjin, China {muli,v-naduan,dozhang,chl,mingzhou}@microsoft.com Abstract This paper presents collaborative decoding (co-decoding), a new method to improve ma- chine translation accuracy by leveraging trans- lation consensus between multiple machine translation decoders. Different from system combination and MBR decoding, which post- process the n-best lists or word lattice of ma- chine translation decoders, in our method mul- tiple machine translation decoders collaborate by exchanging partial translation results. Us- ing an iterative decoding approach, n-gram agreement statistics between translations of multiple decoders are employed to re-rank both full and partial hypothesis explored in decoding. Experimental results on data sets for NIST Chinese-to-English machine translation task show that the co-decoding method can bring significant improvements to all baseline decoders, and the outputs from co-decoding can be used to further improve the result of system combination. 1 Introduction Recent research has shown substantial improve- ments can be achieved by utilizing consensus statistics obtained from outputs of multiple ma- chine translation systems. Translation consensus can be measured either at sentence level or at word level. For example, Minimum Bayes Risk (MBR) (Kumar and Byrne, 2004) decoding over n-best list tries to find a hypothesis with lowest expected loss with respect to all the other transla- tions, which can be viewed as sentence-level consensus-based decoding. Word based methods proposed range from straightforward consensus voting (Bangalore et al., 2001; Matusov et al., 2006) to more complicated word-based system combination model (Rosti et al., 2007; Sim et al., 2007). Typically, the resulting systems take out- puts of individual machine translation systems as input, and build a new confusion network for second-pass decoding. There have been many efforts dedicated to ad- vance the state-of-the-art performance by com- bining multiple systems’ outputs. Most of the work focused on seeking better word alignment for consensus-based confusion network decoding (Matusov et al., 2006) or word-level system combination (He et al., 2008; Ayan et al., 2008). In addition to better alignment, Rosti et al. (2008) introduced an incremental strategy for confusion network construction; and Hildebrand and Vogel (2008) proposed a hypotheses re- ranking model for multiple systems’ outputs with more features including word translation proba- bility and n-gram agreement statistics. A common property of all the work mentioned above is that the combination models work on the basis of n-best translation lists (full hypo- theses) of existing machine translation systems. However, the n-best list only presents a very small portion of the entire search space of a Sta- tistical Machine Translation (SMT) model while a majority of the space, within which there are many potentially good translations, is pruned away in decoding. In fact, due to the limitations of present-day computational resources, a consi- derable number of promising possibilities have to …

Content truncated to 500 words