Fertility-based Source-Language-biased Inversion Transduction Grammar for Word Alignment

Computational Linguistics and Chinese Language Processing Vol. 14, No. 1, March 2009, pp. 1-18 1 © The Association for Computational Linguistics and Chinese Language Processing [Received June 6, 2008; Revised May 12, 2009; Accepted May 15, 2009] Fertility-based Source-Language-biased Inversion Transduction Grammar for Word Alignment Chung-Chi Huang∗ and Jason S. Chang+ Abstract We propose a version of Inversion Transduction Grammar (ITG) model with IBM-style notation of fertility to improve word-alignment performance. In our approach, binary context-free grammar rules of the source language, accompanied by orientation preferences of the target language and fertilities of words, are leveraged to construct a syntax-based statistical translation model. Our model, inherently possessing the characteristics of ITG restrictions and allowing for many consecutive words aligned to one and vice-versa, outperforms the Bracketing Transduction Grammar (BTG) model and GIZA++, a state-of-the-art word aligner, not only in alignment error rate (23% and 14% error reduction) but also in consistent phrase error rate (13% and 9% error reduction). Better performance in these two evaluation metrics suggests that, based on our word alignment result, more accurate phrase pairs may be acquired, leading to better machine translation quality. Keywords: Inversion Transduction Grammar, Syntax-based Statistical Translation Model, Word Alignment. 1. Introduction A statistical translation model is a model which detects word correspondences within sentence pairs, whether relying on lexical information or on syntactic aspects of the involved languages or both. In spite of the fact that methodologies vary, the intention is clear: to obtain better word alignment results so that a better translation model implies better performance in different linguistic applications. Among the methodologies are phrase-based (Och & Ney, 2004; Chiang, 2005; Liu et al., 2006) and syntax-based machine translation systems (Galley et al., 2004; Galley et al., 2006). ∗ CLCLP, TIGP, Academia Sinica, Taipei, Taiwan + Department of Computer Science, NTHU, Hsinchu, Taiwan E-mail: {u901571; jason.jschang}@gmail.com 2 Chung-Chi Huang and Jason S. Chang Since the pioneering work of Brown et al. (1988), a myriad of research projects have focused on the statistical translation model. These could be classified into two main categories: one paying little attention to the grammar of the languages (Vogel et al., 1996; Och & Ney, 2000; Toutanova et al., 2002) and the other explicitly utilizing languages’ structural or syntactic information (Wu, 1997; Yamada & Knight, 2001; Cherry & Lin, 2003; Gildea, 2004; Zhang & Gildea, 2005). With an increasing number of more accurate syntactic analyzers (e.g., part-of-speech tagger and Stanford parser) being developed and in view of the deficiency in modeling grammatical aspects of languages facing IBM-like models, the latter has received increasing attention. Recently, in order to incorporate languages’ syntax, Yamada and Knight (2001) transformed source-language (SL) (e.g., English) parse trees into target-language (TL) (e.g., Japanese) strings, using operations of reordering, inserting, and translating on tree nodes. Instead of accepting monolingual (i.e., SL or TL) parse trees to do the transformation, Wu’s ITG model (1997) first associates production rules (e.g., S→NP VP) commonly shared by two languages with (straight or inverted) word orientation and, based on …

Content truncated to 500 words