Grammar Comparison Study for Translational Equivalence Modeling and Statistical Machine Translation

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 1097–1104 Manchester, August 2008 Grammar Comparison Study for Translational Equivalence Modeling and Statistical Machine Translation Min Zhang1, Hongfei Jiang2, Haizhou Li1, Aiti Aw1 and Sheng Li2 1Institute for Infocomm Research, Singapore 2Harbin Institute of Technology, China {mzhang, hli, aaiti}@i2r.a-star.edu.sg {hfjiang, lisheng}@mtlab.hit.edu.cn Abstract This paper presents a general platform, namely synchronous tree sequence sub- stitution grammar (STSSG), for the grammar comparison study in Transla- tional Equivalence Modeling (TEM) and Statistical Machine Translation (SMT). Under the STSSG platform, we compare the expressive abilities of various gram- mars through synchronous parsing and a real translation platform on a variety of Chinese-English bilingual corpora. Ex- perimental results show that the STSSG is able to better explain the data in paral- lel corpora than other grammars. Our study further finds that the complexity of structure divergence is much higher than suggested in literature, which imposes a big challenge to syntactic transformation- based SMT. 1 Introduction Translational equivalence is a mathematical rela- tion that holds between linguistic expressions with the same meaning (Wellington et al., 2006). The common explicit representations of this rela- tion are word alignments, phrase alignments and structure alignments between bilingual sentences. Translational Equivalence Modeling (TEM) is a process to describe and build these alignments using mathematical models. Thus, the study of TEM is highly relevant to Statistical Machine Translation (SMT). Grammar is the most important infrastructure for TEM and SMT since translation models’ ex- pressive and generative abilities are mainly de- © 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. termined by the grammar. Many grammars, such as finite-state grammars (FSG), bracket/inversion transduction grammars (BTG/ITG) (Wu, 1997), context-free grammar (CFG), tree substitution grammar (TSG) (Comon et al., 2007) and their synchronous versions, have been explored in SMT. Based on these grammars, a great number of SMT models have been recently proposed, including string-to-string model (Synchronous FSG) (Brown et al., 1993; Koehn et al., 2003), tree-to-string model (TSG-string) (Huang et al., 2006; Liu et al., 2006; Liu et al., 2007), string-to- tree model (string-CFG/TSG) (Yamada and Knight, 2001; Galley et al., 2006; Marcu et al., 2006), tree-to-tree model (Synchronous CFG/TSG, Data-Oriented Translation) (Chiang, 2005; Cowan et al., 2006; Eisner, 2003; Ding and Palmer, 2005; Zhang et al., 2007; Bod, 2007; Quirk wt al., 2005; Poutsma, 2000; Hearne and Way, 2003) and so on. Although many achievements have been ob- tained by these advances, it is still unclear which of these important pursuits is able to best explain human translation data, as each has its advan- tages and disadvantages. Therefore, it has great meaning in both theory and practice to do com- parison studies among these grammars and SMT models to see which of them are capable of better describing parallel translation data. This is a fun- damental issue worth exploring in multilingual information processing. However, little effort in previous work has been put in this point. To ad- dress this issue, in …

Content truncated to 500 words