This paper investigates the methods of target language generation in Interchange-Format (IF) based spoken language translation system. On the one hand, the paper introduces the researches on Chinese generation in specific domain. On the other hand, it makes researches on the generation theories and methods that improve the portability and domain portability of a language generator. The research work described in this thesis includes the following parks: First, analysis on the characteristics of the spoken Chinese and research on the methods of domain specific Chinese generation. This generation method adopts an integrative architecture, and is divided into two layers: sentence layer and phrase layer. The template-based method is employed in the Chinese generator. The variables and functions that need further processare imported into the templates. Such method achieves more efficiency than feature based pipeline-structure generator. In addition, it is more flexible than common template-based generation. Second, investigation of the generation strategies. The flexibility, efficiency and robustness are three important factors in the spoken language translation systems. Towards these requirements, our generator employs a hybrid approach in combination of template-based and feature-based generation methods. For those fixed expressions the templates containing variables are used to improve the efficiency. For the other flexible expressions, the feature-based generation method is employed to fulfill the requirement of flexibility. The template method is efficient, but it is inflexible and has poor domain portability. The feature-based method is of the advantages of generality and flexibility, but it is inefficient. Therefore, the combination of the two methods makes the generator have the better tradeoff between the efficiency and the flexibility. Regarding the requirement of robustness, we take measures both in micro-planning and surface generation. Since the Chinese doesn't have much requirement for the participant members of a sentence, the restrictions on the participant members are ignored in the Chinese sentence generation. For English generation, we also loose the limitations on the participant members of a sentence. Furthermore, default values are given for some participant members under the permits of the domain knowledge. Third, probing into the approach to the micro-planning. The micro-planning module is designed according to the characteristics of IF. In our system, there are three sub-tasks for micro planning: sentence planning, phrase planning and lexical selection. The ultimate result for micro-planning is a kind of syntactic and semantic structure that is suitable for sentence generation. The sentence type is determined in the sentence planning according to IF and the domain knowledge, and the predicate-argument frame of the sentence, including the main v
修改评论