| Machine translation(MT)is the process of using software to translate text from one natural language to another,which is an important tool for cross-lingual communication in the information era.After decades of effort machine translation has improved much,but the translation quality of machine translation systems is still not satisfactory for many applications.One reason for this is that there exist several key challenges and problems to machine translation that still require better solutions.The reordering problem,which comes from the differences of word order and syntax structure between languages,is one of such problems,and the ability of selecting the correct word order has a significant impact on the translation quality of a machine translation system.In this thesis we look into the reordering problem from the pre-ordering approach.After analyzing the linguistic phenomena related to reordering,the properties of machine translation systems and previous research on reordering for MT,we proposed two pre-ordering methods for phrase-based machine translation systems.First,we propose a novel word reordering method based on word alignments.Our method extracts bilingual structural information for reordering from automatically word-aligned sentence pairs for training reorder models.It does not require parsers trained on monolingual treebanks,thus can be used on under-resourced languages.Second,in order to achieve better performance for resourceful language pairs,we propose a syntax-based pre-ordering method that utilize syntactic and linguistic information.Based on the proposed probabilistic framework for syntactic pre-ordering,our method is capable of incorporating hand-crafted rules and rule-based methods into the framework to achieve better reordering performance and translation quality.Experiments show that both our two proposed pre-ordering methods are effective in reordering source sentences to resemble target word order,and in improving translation quality. |
PDF Full Text Request