Multi-engine based Chinese-to-English translation system

Multi-Engine Based Chinese-to-English Translation System Yuncun Zuo Yu Zhou Chengqing Zong National Laboratory of Pattern Recognition, Institute of Automation Chinese Academy of Sciences, Beijing 100080, China {yczuo, yzhou, cqzong }@nlpr.ia.ac.cn Abstract This paper describes a Multi-Engine based Chinese-to- English spoken language translation system. The design and implementation of the system is given in detail. Three different translation engines are employed in the system and a very simple way is proposed to select the best translation from all the outputs generated by them. The evaluation results from IWSLT2004 are reported and analyzed in detail. The results prove that the Multi-Engine based system is practical. 1. Introduction Spoken language translation (SLT) technologies attempt to cross the language barriers between people with different native languages who each want to engage in conversation by using their mother-tongue. The importance of these technologies is increasing because there are many more opportunities for cross-language communication in face-to- face and telephone conversation, especially in the domain of tourism. Our work described in this paper is focused on the translation from Chinese spoken language to English, an important part of the multi-lingual information service system oriented to the 28th Beijing Olympic Games. The remainder of this paper is organized as follows: Section 2 gives the related work of Multi-Engine based spoken language translations; Section 3 describes the overview of our system in detail, and then, Section 4 presents the evaluation results and finally Section 5 gives the conclusion. 2. Related work Many approaches have been proposed and practiced in the SLT research, such as template based method, statistical method and so on. Different translation methods are integrated together to get better results for that each machine translation method has its own strengths and weaknesses. At present, there are two main approaches to use different translation methods in one system. The great difference between the two approaches lies in that the output of the first one is combined by the best translation parts generated by various engines but the output of the second approach is a whole result selected from the results generated by each engine [1, 2, 3, 4]. The center idea of the first one is that: different translation engines run together and each engine generates part rather than whole results of an input sentence. The part results created by one engine may be used by the other engines when they generate new part results. Finally a selector or searcher selects part results from all the created part results to compose a new result for the whole input sentence as the system output. This method is very complex and the builder must be familiar with each translation method. The centre idea of the second one is that: different translation engines run independently and get their own result for a complete input sentence, and then a selector selects the best result as the system output. This method does not create new result. It is comparatively easy to be implemented and the builder need not be familiar with every …

Content truncated to 500 words