NUDT machine translation system for IWSLT2007

NUDT Machine Translation System for IWSLT2007 Wen-Han Chao1, Zhou-Jun Li2 1School of Computer Science, National University of Defense Technology, China cwh2k@163.com 2School of Computer Science and Engineering, Beihang University, China lizj@buaa.edu.cn Abstract In this paper, we describe our machine translation system which was used for the Chinese-to-English task in the IWSLT2007 evaluation campaign. The system is a statistical machine translation (SMT) system, while containing an example-based decoder. In this way, it will help to solve the re-ordering problem and other problems for spoken language MT, such as lots of omissions, idioms etc. We report the results of the system for the provided evaluation sets. 1. Introduction The state-of-the-art statistical machine translation (SMT) model [1][2] is the log-linear model [3], which provides a framework to incorporate any useful knowledge for machine translation, such as translation model, language model etc. In a SMT system, one important problem is the re- ordering between words and phrases, especially when the source language and target language are very different in word order, such as Chinese and English. For the spoken language translation, the re-ordering problem will be more crucial, since the spoken language is more flexible in word order. In addition, lots of omissions and idioms make the translation more difficult. In this paper, we present our hybrid translation system, which is a SMT system, while using an example-based decoder, which will use the translation examples to keep the translation structure, i.e. constraint the reordering, and make the omitted words having the chance to be translated. In our system, each translation example is a triple (C, E, TA), where C represents the Chinese sentence, E the English sentence, and TA is the word alignment between C and E, which satisfies the inversion transduction grammar (ITG) [4] constraint, i.e. the TA forms a constituent structure tree. This paper is organized as follows. In Section 2, we describe the various components in our system, especially the word aligner and decoder. In section 3, we report the experimental results of Chinese-English translation, and we conclude in section 4 and provide avenues for further research. 2. System Description Our machine translation system is a modular MT engine, which mainly consists of the following components: • Word Alignment: taking the bilingual sentence-aligned training corpus as input, obtains the Viterbi word alignment for each sentence pair, in our system, the word alignment must satisfy the ITG constraint. • Phrase Pair Extracting: taking the bilingual word- aligned training corpus as input, extracts the valid phrase pairs and builds the translation model and the reordering model. • Decoder: given a Chinese sentence as input, search the best translation using the word-aligned corpus and the translation model, reordering model and language model. 2.1. Word Alignment The word alignment [5] is the base of the SMT system. In our system, the word alignment for each sentence pair is used to build translation model and reordering model, and also used to provide the valid translation example. In our system, the word alignment needs to satisfy the …

Content truncated to 500 words