The Feature Subspace Method for SMT System Combination

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1096–1104, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP The Feature Subspace Method for SMT System Combination Nan Duan1, Mu Li2, Tong Xiao3, Ming Zhou2 1Tianjin University 2Microsoft Research Asia 3Northeastern University Tianjin, China Beijing, China Shenyang, China {v-naduan,muli,v-toxiao,mingzhou}@microsoft.com Abstract Recently system combination has been shown to be an effective way to improve translation quality over single machine translation sys- tems. In this paper, we present a simple and ef- fective method to systematically derive an en- semble of SMT systems from one baseline li- near SMT model for use in system combina- tion. Each system in the resulting ensemble is based on a feature set derived from the fea- tures of the baseline model (typically a subset of it). We will discuss the principles to deter- mine the feature sets for derived systems, and present in detail the system combination mod- el used in our work. Evaluation is performed on the data sets for NIST 2004 and NIST 2005 Chinese-to-English machine translation tasks. Experimental results show that our method can bring significant improvements to baseline systems with state-of-the-art performance. 1 Introduction Research on Statistical Machine Translation (SMT) has shown substantial progress in recent years. Since the success of phrase-based methods (Och and Ney, 2004; Koehn, 2004), models based on formal syntax (Chiang, 2005) or lin- guistic syntax (Liu et al., 2006; Marcu et al., 2006) have also achieved state-of-the-art perfor- mance. As a result of the increasing numbers of available machine translation systems, studies on system combination have been drawing more and more attention in SMT research. There have been many successful attempts to combine outputs from multiple machine transla- tion systems to further improve translation quali- ty. A system combination model usually takes n- best translations of single systems as input, and depending on the combination strategy, different methods can be used. Sentence-level combina- tion methods directly select hypotheses from original outputs of single SMT systems (Sim et al., 2007; Hildebrand and Vogel, 2008), while phrase-level or word–level combination methods are more complicated and could produce new translations different from any translations in the input (Bangalore et al., 2001; Jayaraman and La- vie, 2005; Matusov et al., 2006; Sim et al., 2007). Among all the factors contributing to the suc- cess of system combination, there is no doubt that the availability of multiple machine transla- tion systems is an indispensable premise. Al- though various approaches to SMT system com- bination have been explored, including enhanced combination model structure (Rosti et al., 2007), better word alignment between translations (Ayan et al., 2008; He et al., 2008) and improved confusion network construction (Rosti et al., 2008), most previous work simply used the en- semble of SMT systems based on different mod- els and paradigms at hand and did not tackle the issue of how to obtain the ensemble in a prin- cipled way. To our knowledge the only work discussed this problem is Macherey and Och (2007), …

Content truncated to 500 words