This master dissertation is based on the research about the reordering problem instatistical machine translation (SMT). Specifically, we focus on reordering problem ofhierarchical phrase-based translation model, since its reordering ability mainly relies onthe hierarchical rules. Statistical model are integrated to guide the hierarchical rule selec-tion for better translation performance. A hierarchical rule contains both the source sideand the target side. Many translation errors come from improper pattern matching of thesource side. Moreover, improper target rule selection may also lead to unacceptable re-orderings. Previous work just pays attention on the selection of either the source side of ahierarchical rule or the target side of a hierarchical rule rather than consider both of themsimultaneously. This paper presents a joint rule selection model to predict the selection ofhierarchical rules, which correlates the two tasks more closely and conducts the search ina relative larger space. The proposed model is estimated based on four sub-models whererich context knowledge from both source and target sides is leveraged. Our method can beeasily incorporated into the practical SMT systems with the log-linear model framework.The experimental results show that our method can not only yield significant gains overthe baseline, but also bring absolute improvements over others'work in performance. |