基于句法结构和篇章结构的统计机器翻译关键技术研究

Structured knowledge of natural languages at syntactic and discourse levels is one kind of the most important features of natural languages. Statistical syntax-based machine translation treats machine translation as a process of converting source structures to target ones at single sentence level. Therefore, cross-language structure divergence and cross-sentence discourse information deficiency are two fundamental issues for the state-of-the-art statistical syntax-based machine translation. This proposal aims to focus on the following researches to address the two fundamental issues aforementioned: 1) from bilingual alignment and mapping viewpoints, study weakly-supervised and unsupervised bilingual grammar induction algorithms, automatically construct machine translation-oriented bilingual grammar rule systems and mapping mechanisms in order to solve the cross-language structure divergence issue; 2) establish the discourse-based model for statistical machine translation in order to impose consistent discourse structures and fundamental discourse features between source and target languages, and thus fill in the discourse information deficiency in the state-of-the-art statistical machine translation research. The research achievements from this proposal will establish the foundation of theory and technology for next generation syntax-based statistical machine translation, and provide guidance to deep semantic structures-based statistical machine translation in the future.

语言的结构性，包括句法和篇章等结构信息，是语言的基本特征之一。统计句法机器翻译把翻译看作是一个从源语言到目标语言的句法结构转换过程，而这种转换过程目前是以单一句子为基本单位进行的。因此，跨语言句法结构差异性和跨句子篇章结构信息的缺失是目前统计句法机器翻译面临的两个核心问题。本项目拟围绕这两个问题展开以下创新性研究：1）从双语映射角度出发，研究弱指导和无指导的双语句法推导算法，自动构建适合机器翻译的双语句法规则体系和映射机制，进而解决跨语言句法结构差异性；2）建立篇章级翻译模型，使得源语言篇章结构信息和篇章基本特征在目标语言中得到体现，使生成的目标语言具有良好的篇章一致性，进而填补机器翻译中篇章信息缺失这一研究空白。本项目预期成果不仅可以为下一代统计句法机器翻译奠定理论和技术基础，同时也可以为未来基于深层语义结构信息的统计机器翻译提供借鉴。