Improving Arabic-Chinese Statistical Machine Translation using English as Pivot Language

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 173–181, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics Improving Arabic-Chinese Statistical Machine Translation using English as Pivot Language Nizar Habash Jun Hu Center for Computational Learning Systems Computer Science Department Columbia University Columbia University New York, NY 10115, USA New York, NY 10115, USA habash@ccls.columbia.edu jh2740@columbia.edu Abstract We present a comparison of two approaches for Arabic-Chinese machine translation using Eng- lish as a pivot language: sentence pivoting and phrase-table pivoting. Our results show that using English as a pivot in either approach out- performs direct translation from Arabic to Chi- nese. Our best result is the phrase-pivot system which scores higher than direct translation by 1.1 BLEU points. An error analysis of our best system shows that we successfully handle many complex Arabic-Chinese syntactic variations. 1 Introduction Arabic and Chinese are two languages with a very large global presence; however, there has not been, to our knowledge, any work on MT for this pair. Given the cost involved in creat- ing parallel corpora for Arabic and Chinese and given that there are lots of available resources (in particular parallel corpora) for Arabic and English and for Chinese and English, we are interested in exploring the role English might serve as a pivot (or bridge) language. In this paper we explore different ways of pivoting through English to translate Arabic to Chinese. Our work is similar to previous research on pivot languages except in that our three lan- guages (source, pivot and target) are very dif- ferent and from completely unrelated families. We focus our experiments on a trilingual paral- lel corpus to keep all conditions experimentally clean. Our results show that using English as a pivot language for translating Arabic to Chinese actually outperforms direct translation. We be- lieve this may be a result of English being a sort of middle ground between Arabic and Chinese in terms of different linguistic features and, in particular, word order. Section 2 describes previous work. Section 3 discusses relevant linguistic issues of Arabic, Chinese and English. Section 4 describes our system and different pivoting techniques. And Section 5 presents our experimental results. 2 Previous Work There has been a lot of work on translation from Chinese to English (Wang et al., 2007; Crego and Mariño, 2007; Carpuat and Wu, 2007; among others) and from Arabic to Eng- lish (Sadat and Habash, 2006, Al-Onaizan and Papineni, 2006; among others). There is also a fair amount of work on translation into Chinese from Japanese, Korean and English (Isahara et al., 2007; Kim et al., 2002; Ye et al., 2007; among others). In 2008, the National Institute of Standards and Technology (NIST) MT Evaluation competition introduced English- Chinese as a new evaluation track.1 Much work has been done on exploiting multi- lingual corpora for MT or related tasks such as lexical induction or word alignment. Schafer and Yarowsky (2002) induced translation lexi- cons for languages without common parallel corpora …

Content truncated to 500 words