Auto word alignment based Chinese-English EBMT

Auto Word Alignment Based Chinese-English EBMT* Yang Muyun, Zhao Tiejun, Liu Haijie, Shi Xiaosheng and Jiang Hongfei Research Center for Language Technology, School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China. {ymy,tjzhao,hjliu,xshshi,hfjiang}@mtlab.hit.edu.cn * Supported by the High Technology Research and Development Program of China (2002AA117010-09) and National Natural Science Foundation of China (60375019) Abstract We present a bidirectional Example-Based Machine Translation (EBMT) system for Chinese—English. The prerequisite is a bilingual aligned corpus of Chinese—English sentences, and we describe the example extraction efforts purely based on word alignment. The whole system is designed to be language independent and as automatic as possible for construction. We present initial experiments which show that our algorithm can successfully generate better translations for the domain in question than the baseline rule based system. 1. Introduction The Olympic Games will be held in Beijing, China in 2008. It is clear that a great deal of translation will be required from Chinese to English, and vice versa. This paper describes our bi-directional Example-Based Machine Translation (EBMT) system for Chinese—English, which is purely based on the word alignment information of a given bilingual corpus. Among the vast issues for a translation system, the followings are emphasized in our system design: (1) Automatic construction: Manual knowledge composition is not desired for system building except for some public existed knowledge basses (e.g. dictionary). The whole process of translation knowledge acquisition should be as automatic as possible. (2) Sub-sentential translation example focus: Linguistically there are infinite sentences. A translation system is desired to capture the translation correspondences under the sentence level, hoping to recombine them for proper translations. (3) Linguistic light approach: Current deep linguistic analysis tools (like parser et al) are not reliable enough. So, if necessary, we would just choose shallow linguistic analyser which causes somewhat less information loss. (4) Adaptability: Since Olympics demands multi languages besides English and Chinese, the method is kept language independent as possible so that the system success (if it did!) could be readily extended to the translations between Chinese and other languages. For the same reason, domain specific advantages are not exploited in the the system. 2. Auto Word Alignment Based EBMT To meet the concerns mentioned above, we propose an EBMT method based solely on the word alignment information of the Chinese English parallel corpus. Basically speaking, the whole process of system construction can be fully automatic as long as a translation dictionary and a word aligned bilingual corpus is provided. Figure 1 describes the architecture of the whole system. Figure 1: AWA based EBMT architecture. Roughly speaking, the whole system can be understood as tow parts: the training process (left half of Fig.1) and the translation process (right half). The training process automatically extracts the translation example base from a word aligned Chinese English bilingual corpus. The key component is the word alignment based example extraction algorithm. The translation process will pre-process the input sentence (tokenization, numerical processing, Chinese word segmentation et al), feed it into the translation …

Content truncated to 500 words