Divide and Translate: Improving Long Distance Reordering in Statistical Machine Translation

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 418–427, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Divide and Translate: Improving Long Distance Reordering in Statistical Machine Translation Katsuhito Sudoh, Kevin Duh, Hajime Tsukada, Tsutomu Hirao, Masaaki Nagata NTT Communication Science Laboratories 2-4 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0237, Japan sudoh@cslab.kecl.ntt.co.jp Abstract This paper proposes a novel method for long distance, clause-level reordering in statistical machine translation (SMT). The proposed method separately translates clauses in the source sentence and recon- structs the target sentence using the clause translations with non-terminals. The non- terminals are placeholders of embedded clauses, by which we reduce complicated clause-level reordering into simple word- level reordering. Its translation model is trained using a bilingual corpus with clause-level alignment, which can be au- tomatically annotated by our alignment algorithm with a syntactic parser in the source language. We achieved signifi- cant improvements of 1.4% in BLEU and 1.3% in TER by using Moses, and 2.2% in BLEU and 3.5% in TER by using our hierarchical phrase-based SMT, for the English-to-Japanese translation of re- search paper abstracts in the medical do- main. 1 Introduction One of the common problems of statistical ma- chine translation (SMT) is to overcome the differ- ences in word order between the source and target languages. This reordering problem is especially serious for language pairs with very different word orders, such as English-Japanese. Many previous studies on SMT have addressed the problem by incorporating probabilistic models into SMT re- ordering. This approach faces the very large com- putational cost of searching over many possibili- ties, especially for long sentences. In practice the search can be made tractable by limiting its re- ordering distance, but this also renders long dis- tance movements impossible. Some recent stud- ies avoid the problem by reordering source words prior to decoding. This approach faces difficul- ties when the input phrases are long and require significant word reordering, mainly because their reordering model is not very accurate. In this paper, we propose a novel method for translating long sentences that is different from the above approaches. Problematic long sentences often include embedded clauses1 such as rela- tive clauses. Such an embedded (subordinate) clause can usually be translated almost indepen- dently of words outside the clause. From this viewpoint, we propose a divide-and-conquer ap- proach: we aim to translate the clauses sepa- rately and reconstruct the target sentence using the clause translations. We first segment a source sen- tence into clauses using a syntactic parser. The clauses can include non-terminals as placeholders for nested clauses. Then we translate the clauses with a standard SMT method, in which the non- terminals are reordered as words. Finally we re- construct the target sentence by replacing the non- terminals with their corresponding clause transla- tions. With this method, clause-level reordering is reduced to word-level reordering and can be dealt with efficiently. The models for clause translation are trained using a bilingual corpus with clause- level alignment. We also …

Content truncated to 500 words