A Dependency Edge-based Transfer Model for Statistical Machine Translation

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1103–1113, Dublin, Ireland, August 23-29 2014. A Dependency Edge-based Transfer Model for Statistical Machine Translation Hongshen Chen†§ Jun Xie† Fandong Meng†§ Wenbin Jiang† Qun Liu‡† †Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences §University of Chinese Academy of Sciences {chenhongshen,xiejun,mengfandong,jiangwenbin}@ict.ac.cn ‡CNGL, School of Computing, Dublin City University qliu@computing.dcu.ie Abstract Previous models in syntax-based statistical machine translation usually resort to some kinds of synchronous procedures, few of these works are based on the analysis-transfer-generation methodology. In this paper, we present a statistical implementation of the analysis-transfer- generation methodology in rule-based translation. The procedures of syntax analysis, syntax transfer and language generation are modeled independently in order to break the synchronous constraint, resorting to dependency structures with dependency edges as atomic manipulating units. Large-scale experiments on Chinese to English translation show that our model exhibits state-of-the-art performance by significantly outperforming the phrase-based model. The statis- tical transfer-generation method results in significantly better performance with much smaller models. 1 Introduction Researches in statistical machine translation have been flourishing in recent years. Statistical translation methods can be divided into word-based (Brown et al., 1993), phrase-based (Marcu and Wong, 2002; Koehn et al., 2003) and syntax-based models (Yamada and Knight, 2001; Graehl and Knight, 2004; Chiang, 2005; Liu et al., 2006; Mi et al., 2008; Huang et al., 2006; Lin, 2004; Ding and Palmer, 2004; Quirk et al., 2005; Shen et al., 2008; Xie et al., 2011; Meng et al., 2013). Compared with word-based and phrase-based methods, syntax-based models perform better in long distance reordering and enjoy higher generalization capability by leveraging the hierarchical structures in natural languages, and achieve the state-of-the-art performance in these years. Most syntax-based models (except for Lin (2004) ) utilize some kinds of synchronous generation procedures which directly model the structural correspondence between two languages. In contrast, the analysis-transfer-generation methodology in rule-based translation solves the machine translation problem in a more divided scheme, where the processing procedures of analysis, structural transfer and language generation are modeled separately. The analysis-transfer-generation strategy can tolerate higher non-isomorphism between languages if with a more general transformation unit and it can facilitate elaborating engineering of each processing procedure, however, there isn’t a statistical transfer model that shows the comparable performance with the current state-of-the-art SMT model so far. In this paper, we propose a novel statistical analysis-transfer-generation model for machine transla- tion, to integrate the advantages of the transfer-generation scheme and the statistical modeling. The procedures of transfer and generation are modeled on dependency structures with dependency edges as atomic manipulating units. First, the source sentence is parsed by a dependency parser. Then, the source dependency structure is transferred into a target structure by translation rules, which composed of the source and target edges. Last, the target sentence is finally generated from the target edges which are used as intermediate syntactic structures. By directly modeling the edge, the most basic unit in the dependency tree, …

Content truncated to 500 words