Acquisition Of Tree-To-String Alignment Based On Phrase-Syntax Structure

Translation template is one kind of the most important source knowledge in machine translation systems. The quality and scale of the templates can directly influence the performance machine translation system. So, it becomes a hot point in current research that to acquired high grade translation template more efficient from corpus automatically.In this paper, we propose a Tree-to-String Alignment Template which based on phrasal structure, describes the alignment between the source parse tree and the target string. Massive syntax structures,structure tags and variables was introduced to the template, which made syntax-based models enable to process the non-continual phrase and has the generalization ability. Depending on the type of decoder, our template can be used in the models of syntax-based statistic machine translation systems, example-based machine translation systems and rule-based machine translation systems.Based on this, we propose a method of acquired this template automatically from the unannotated bilingual corpora and single tree bank corpora. This approach is semi-supervised, statistical, and data-driven, which extract the translation template with comprehensive utilization of two aspect's information. One is to post-order traversal the syntax tree, to extract the candidate template information based on the word alignment result, which include the source language syntax sub-tree, the corresponding goal string of language and the alignment information; One is to extract tree structure information from single sentence. The preliminary experimental result is show that our method is effective. This templates can help assist in machine translation and improves translation quality.