The XMU SMT system for IWSLT 2007

The XMU SMT System for IWSLT 2007 Yidong Chen, Xiaodong Shi, Changle Zhou Department of Cognitive Science, School of Information Sciences and Technologies, Xiamen University, Xiamen, Fujian, P. R. China {ydchen, mandel, dozero}@xmu.edu.cn Abstract In this paper, an overview of the XMU statistical machine translation (SMT) system for the 2007 IWSLT Speech Translation Evaluation is given. Our system is a phrase-based system with a reordering model based on chunking and reordering of source language. In this year’s evaluation, we participated in the open data track for Clean Transcripts for the Chinese-English translation direction. The system ranked the 12th among the 15 participating systems. 1. Introduction This paper describes the system which participated in the 2007 IWSLT Speech Translation Evaluation of Department of Cognitive Science, Xiamen University. The system is a phrase-based Statistical Machine Translation (SMT) system with a reordering model based on chunking and reordering of source language. The steps of our model could be stated briefly as follows: • At training time: Firstly, align the sentence pairs and get word alignment matrixes. Secondly, chunk the source sentences extract chunk reordering information according to the word alignment matrixes. Thirdly, reorder the source sentences using the chunk reordering information. Finally, train the baseline phrase-based system with the reordered source sentences and the original target sentences. • At decoding time: Firstly, chunk the test sentences. Secondly, reorder the test sentences with the chunk reordering information. Finally, translate the reordered sentences with the baseline phrase-based decoder in monotonic order. This paper is organized as follows. Section 2 describes data preparation. Section 3 gives an overview of the translation model. In section 4, experiments and the results are reported. And finally, section 5 concludes. 2. Preparing the Data 2.1. Preprocessing Data preprocessing is not a trivial task for machine translation system. Our experiments showed that good preprocessing model can result in better translation quality. data Three types of preprocessing were performed on the Chinese part of the training data: • Segmentation and Tagging: To transform Chinese characters into Chinese words and POS tags of each word. • Chunking: To chunk the POS-tagged Chinese sentences. We used a Chinese chunker based on CRF++1 in this evaluation. • SBC case to DBC case: To replace numbers, English characters or punctuations in SBC case in Chinese by their DBC case. For instance, "１", "Ａ" and "。" would respectively be replaced by "1", "A" and ".". For the English part of the training data, also two types of preprocessing were performed: • Tokenization: To separate punctuations from words in English sentences. • Truecasing of the first word of an English sentence: To transform the uppercase version of the beginning words of English sentences into their lowercase version if their lowercase version occur more often. 2.2. Word Alignment To achieve n-to-n word alignment, we first run GIZA++ up to IBM model 4 in both translation directions to get an initial word alignment, and then apply “grow-diag-final” method [1] to refine it. This process could be addressed in detail as …

Content truncated to 500 words