Improving Alignment of System Combination by Using Multi-objective Optimization

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 535–544, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Improving Alignment of System Combination by Using Multi-objective Optimization Tian Xia+, Zongcheng Ji∗, Shaodan Zhai+, Yidong Chen++, Qun Liu∗, Shaojun Wang+ ++ Xiamen University, Xiamen 361005, P.R. China + Wright State University, 3640 Colonel Glenn Hwy, Dayton, OH 45435, USA ∗ Institute of Computing Technology, Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {jizongcheng, liuqun}@ict.ac.cn and ydchen@xmu.edu.cn {xia.7, zhai.6, shaojun.wang}@wright.edu Abstract This paper proposes a multi-objective opti- mization framework which supports heteroge- neous information sources to improve align- ment in machine translation system combi- nation techniques. In this area, most of techniques usually utilize confusion networks (CN) as their central data structure to com- pact an exponential number of an potential hy- potheses, and because better hypothesis align- ment may benefit constructing better quality confusion networks, it is natural to add more useful information to improve alignment re- sults. However, these information may be het- erogeneous, so the widely-used Viterbi algo- rithm for searching the best alignment may not apply here. In the multi-objective opti- mization framework, each information source is viewed as an independent objective, and a new goal of improving all objectives can be searched by mature algorithms. The so- lutions from this framework, termed Pareto optimal solutions, are then combined to con- struct confusion networks. Experiments on two Chinese-to-English translation datasets show significant improvements, 0.97 and 1.06 BLEU points over a strong Indirected Hidden Markov Model-based (IHMM) system, and 4.75 and 3.53 points over the best single ma- chine translation systems. 1 Introduction System combination (SC) techniques have the power of boosting translation quality in BLEU by several percent over the best among all input ma- chine translation systems (Bangalore et al., 2001; Matusov et al., 2006; Sim et al., 2007; Rosti et al., 2007b; Rosti et al., 2007a; Huang and Papineni, 2007; He et al., 2008; Rosti et al., 2008; He and Toutanova, 2009; Li et al., 2009; Feng et al., 2009; Pauls et al., 2009). A central data structure in the SC is the confusion network, and its quality greatly affects the final performance. He et al. (2008) pro- posed a new hypothesis alignment algorithm for constructing high-quality confusion networks called Indirect Hidden Markov Model (IHMM), which does better in synonym matching compared with the classic translation edit rate (TER) based algo- rithm (Rosti et al., 2007b; Rosti et al., 2008; Sim et al., 2007). Now, current state-of-the-art SC systems have been using IHMM or variants in their align- ment algorithms more or less (Li et al., 2009; Feng et al., 2009). Our motivation derives from an observation that in an ideal alignment of a pair of sentences, many-to- many alignments often exist. For instance, “be about to” has the same meaning with “be on the point of”. Because Hidden Markov Model based align- ment algorithms, e.g. IHMM for system combina- tion, HMM in GIZA++ software for statistical …

Content truncated to 500 words