Reordering with Source Language Collocations

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 1036–1044, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Reordering with Source Language Collocations Zhanyi Liu1,2, Haifeng Wang2, Hua Wu2, Ting Liu1, Sheng Li1 1Harbin Institute of Technology, Harbin, China 2Baidu Inc., Beijing, China {liuzhanyi, wanghaifeng, wu_hua}@baidu.com {tliu, lisheng}@hit.edu.cn Abstract This paper proposes a novel reordering model for statistical machine translation (SMT) by means of modeling the translation orders of the source language collocations. The model is learned from a word-aligned bilingual cor- pus where the collocated words in source sen- tences are automatically detected. During decoding, the model is employed to softly constrain the translation orders of the source language collocations, so as to constrain the translation orders of those source phrases con- taining these collocated words. The experi- mental results show that the proposed method significantly improves the translation quality, achieving the absolute improvements of 1.1~1.4 BLEU score over the baseline me- thods. 1 Introduction Reordering for SMT is first proposed in IBM mod- els (Brown et al., 1993), usually called IBM con- straint model, where the movement of words during translation is modeled. Soon after, Wu (1997) proposed an ITG (Inversion Transduction Grammar) model for SMT, called ITG constraint model, where the reordering of words or phrases is constrained to two kinds: straight and inverted. In order to further improve the reordering perfor- mance, many structure-based methods are pro- posed, including the reordering model in hierarchical phrase-based SMT systems (Chiang, 2005) and syntax-based SMT systems (Zhang et al., 2007; Marton and Resnik, 2008; Ge, 2010; Vis- weswariah et al., 2010). Although the sentence structure has been taken into consideration, these methods don‟t explicitly make use of the strong correlations between words, such as collocations, which can effectively indicate reordering in the target language. In this paper, we propose a novel method to im- prove the reordering for SMT by estimating the reordering score of the source-language colloca- tions (source collocations for short in this paper). Given a bilingual corpus, the collocations in the source sentence are first detected automatically using a monolingual word alignment (MWA) me- thod without employing additional resources (Liu et al., 2009), and then the reordering model based on the detected collocations is learned from the word-aligned bilingual corpus. The source colloca- tion based reordering model is integrated into SMT systems as an additional feature to softly constrain the translation orders of the source collocations in the sentence to be translated, so as to constrain the translation orders of those source phrases contain- ing these collocated words. This method has two advantages: (1) it can au- tomatically detect and leverage collocated words in a sentence, including long-distance collocated words; (2) such a reordering model can be inte- grated into any SMT systems without resorting to any additional resources. We implemented the proposed reordering mod- el in a phrase-based SMT system, and the evalua- tion results show that our method significantly improves translation quality. As compared to the baseline systems, …

Content truncated to 500 words