NTT System Description for the WMT2006 Shared Task

Proceedings of the Workshop on Statistical Machine Translation, pages 122–125, New York City, June 2006. c⃝2006 Association for Computational Linguistics NTT System Description for the WMT2006 Shared Task Taro Watanabe Hajime Tsukada Hideki Isozaki NTT Communication Science Laboratories 2-4 Hikaridai, Seika-cho, Soraku-gun, Kyoto, Japan 619-0237 {taro,tsukada,isozaki}@kecl.ntt.co.jp Abstract We present two translation systems ex- perimented for the shared-task of “Work- shop on Statistical Machine Translation,” a phrase-based model and a hierarchical phrase-based model. The former uses a phrasal unit for translation, whereas the latter is conceptualized as a synchronous- CFG in which phrases are hierarchically combined using non-terminals. Experi- ments showed that the hierarchical phrase- based model performed very comparable to the phrase-based model. We also report a phrase/rule extraction technique differ- entiating tokenization of corpora. 1 Introduction We contrasted two translation methods for the Workshop on Statistical Machine Translation (WMT2006) shared-task. One is a phrase-based translation in which a phrasal unit is employed for translation (Koehn et al., 2003). The other is a hierarchical phrase-based translation in which translation is realized as a set of paired production rules (Chiang, 2005). Section 2 discusses those two models and details extraction algorithms, decoding algorithms and feature functions. We also explored three types of corpus pre- processing in Section 3. As expected, different tokenization would lead to different word align- ments which, in turn, resulted in the divergence of the extracted phrase/rule size. In our method, phrase/rule translation pairs extracted from three distinctly word-aligned corpora are aggregated into one large phrase/rule translation table. The experi- ments and the final translation results are presented in Section 4. 2 Translation Models We used a log-linear approach (Och and Ney, 2002) in which a foreign language sentence f J 1 = f1, f2, ... fJ is translated into another language, i.e. English, eI 1 = e1, e2, ..., eI by seeking a maximum likelihood solution of ˆeI 1 = argmax eI 1 Pr(eI 1| f J 1 ) (1) = argmax eI 1 exp ��M m=1 λmhm(eI 1, f J 1 ) � � e′I′ 1 exp ��M m=1 λmhm(e′I′ 1 , f J 1 ) �(2) In this framework, the posterior probability Pr(eI 1| f J 1 ) is directly maximized using a log-linear combination of feature functions hm(eI 1, f J 1 ), such as a ngram language model or a translation model. When decoding, the denominator is dropped since it depends only on f J 1 . Feature function scaling factors λm are optimized based on a maximum likelihood approach (Och and Ney, 2002) or on a direct error minimization approach (Och, 2003). This modeling allows the integration of various feature functions depending on the scenario of how a translation is constituted. In a phrase-based statistical translation (Koehn et al., 2003), a bilingual text is decomposed as K phrase translation pairs (¯e1, ¯f¯a1), (¯e2, ¯f¯a2), ...: The in- put foreign sentence is segmented into phrases ¯f K 1 , 122 mapped into corresponding English ¯eK 1 , then, re- ordered to form the …

Content truncated to 500 words