Meta-Structure Transformation Model for Statistical Machine Translation

Proceedings of the Second Workshop on Statistical Machine Translation, pages 64–71, Prague, June 2007. c⃝2007 Association for Computational Linguistics 1 Meta-Structure Transformation Model for Statistical Machine Translation Jiadong Sun, Tiejun, Zhao and Huashen Liang MOE-MS Key Lab of National Language Processing and speech Harbin Institute of Technology No. 92, West Da-zhi Street ,Harbin Heilongjiang ,150001 ,China jiadongsun@hit.edu.cn {tjzhao, hsliang }@mtlab.hit.edu.cn Abstract We propose a novel syntax-based model for statistical machine translation in which meta-structure (MS) and meta-structure se- quence (SMS) of a parse tree are defined. In this framework, a parse tree is decom- posed into SMS to deal with the structure divergence and the alignment can be recon- structed at different levels of recombination of MS (RM). RM pairs extracted can per- form the mapping between the sub- structures across languages. As a result, we have got not only the translation for the target language, but an SMS of its parse tree at the same time. Experiments with BLEU metric show that the model signifi- cantly outperforms Pharaoh, a state-art-the- art phrase-based system. 1 Introduction The statistical approach has been widely used in machine translation, which use the noisy-channel- based model. A joint probability model, proposed by Marcu and Wong (2002), is a kind of phrase- based one. Och and Ney (2004) gave a framework of alignment templates for this kind of models. All of the phrase-based models outperformed the word-based models, by automatically learning word and phrase equivalents from bilingual corpus and reordering at the phrase level. But it has been found that phrases longer than three words have little improvement in the performance (Koehn, 2003). Above the phrase level, these models have a simple distortion model that reorders phrases inde- pendently, without consideration of their contents and syntactic information. In recent years, applying different statistical learning methods to structured data has attracted various researchers. Syntax-based MT approaches began with Wu (1997), who introduced the Inver- sion Transduction Grammars. Utilizing syntactic structure as the channel input was introduced into MT by Yamada (2001). Syntax-based models have been presented in different grammar formalisms. The model based on Head-transducer was pre- sented by Alshawi (2000). Daniel Gildea (2003) dealt with the problem of the parse tree isomor- phism with a cloning operation to either tree-to- string or tree-to-tree alignment models. Ding and Palmer (2005) introduced a version of probabilistic extension of Synchronous Dependency Insertion Grammars (SDIG) to deal with the pervasive structure divergence. All these approaches don’t model the translation process, but formalize a model that generates two languages at the same time, which can be considered as some kind of tree transducers. Graehl and Knight (2004) described the use of tree transducers for natural language processing and addressed the training problems for this kind of transducers. In this paper, we define a model based on the MS decomposition of the parse trees for statistical machine translation, which can capture structural variations and has a proven generation capacity. During the translation process of our model, the parse tree of the …

Content truncated to 500 words