Chinese-English Statistical Machine Translation by Parsing

Statistical machine translation (SMT) has evolved from the word-based level to higher levels of abstraction. Currently the best known systems are phrased-based, and recent research has started to explore tree-based systems with syntactical information. This thesis aims to study large-scale Chinese-English SMT using a syntactic tree-based model. From the engineering point of view, SMT systems are very complex to build. However, existing pieces of software can bring the work load to a manageable level for this thesis. Using the GenPar framework and other software, this thesis studies Chinese-English SMT by parsing by large-scale experiments. This is the first application of GenPar on Chinese-English SMT. The experiments show that the accuracy of Chinese-English SMT by parsing is comparable to existing SMT by parsing of other language pairs. However, the accuracy of current MT methods is still largely below human translation, and is influenced by the difference between training and testing data, such as the writing style and domain. Two important factors in the SMT by parsing model are studied, and it is observed that though the accuracy of word-to-word alignment influences the translation accuracy, the mono-lingual English and Chinese grammars do not have a significant impact on the results. From the above observations, advantages and weaknesses of the SMT model are analysed, and possible future improvements for Chinese-English SMT are suggested. This thesis is organised in three main parts. The first chapter presents the introduction and overview of the thesis. The second and third chapters summarise the related theories by literature review, giving a detailed exposition of the theory of SMT and SMT by parsing. The last two chapters report the novel experiments of Chinese-English SMT by generalised parsing. By discussing the experimental output, the last chapter summarises this thesis and proposes further work.