The CASIA Phrase-Based Machine Translation System

The CASIA Phrase-Based Machine Translation System Wei Pang, Zhendong Yang, Zhenbiao Chen, Wei Wei, Bo Xu, Chengqing Zong Institute of Automation Chinese Academy of Sciences wpang@hitic.ia.ac.cn Abstract In this paper we propose a phrase-based translation system. In the system, we use phrase translation model instead of word-based model. An improved method of computing phrase translation probability is studied. We translate numeral phrases first by using a standard templates depository. We develop a phrase-based decoder that employs a beam search algorithm. To make the result more reasonable, we apply those words with fertility probability of zero. We improve the previously proposed tracing back algorithm to get the best path. Some experiments concerned are presented. 1 Introduction Statistical machine translation is a promising approach for large vocabulary text translation. In the early 90s, IBM developed Candide system. Since then, many statistical machine translation systems were proposed [2][3]. These systems apply a translation model to capture the relationship between the source and target languages, and use a language model to drive the search process. The primary IBM model was purely word-based. To get more complex structure, better lexical choice and more reliable local reordering, the phrase-based statistical machine translation systems were proposed. Yamada and Knight [4] used phrase translation in a syntax-based translation system; March and Wong [5] introduced a joint-probability model for phrase translation; CMU and IBM also improved their systems with phrase translation capability. Our system applies a phrase-based translation model to capture the corresponding relationship between two languages. We propose a formula to compute the phrase translation probability through word alignment. The phrase-based decoder we developed employs a beam search algorithm, similar to the one in [6]. We applied zero fertility words in the target language. Because the translation quality largely depends on the accuracy of phrase-to-phrase translation pairs extracted from bilingual corpora, we propose a different tracing back algorithm to find the best path. Four methods are studied to extract bilingual phrase pairs. We describe these methods and phrase-based translation model in Section 2. Section 3 explains the method of numeral phrase translation. Section 4 outlines the architecture of the decoder that combines the translation model, distortion model, language model to generate target sentence. In Section 5, we present a series of experiments in which Chinese sentences are translated into English sentences, and analyze the results of these experiments. We make a summary in Section 6. 2 Phrase Translation Model Our system is based on a phrase translation model, which is different from the original IBM model. The phrase we mention here is composed of a series of words that perhaps possess no syntax or semantic meanings. In addition, a word can also be treated as a phrase, so the word-based model is included in the phrase-based model. There are different methods of getting phrase pairs from a bilingual corpus .We used four methods as follows: 2-1 Extracting directly through IBM Word-Based model By using IBM model 4, we get a series of target language words that correspond to …

Content truncated to 500 words