A Lexicalized Reordering Model for Hierarchical Phrase-based Translation

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1144–1153, Dublin, Ireland, August 23-29 2014. A Lexicalized Reordering Model for Hierarchical Phrase-based Translation1 Hailong Cao1, Dongdong Zhang2, Mu Li2, Ming Zhou2 and Tiejun Zhao1 1Harbin Institute of Technology, Harbin, P.R. China 2Microsoft Research Asia, Beijing, P.R. China {hailong, tjzhao}@mtlab.hit.edu.cn {Dongdong.Zhang, muli, mingzhou}@microsoft.com Abstract Lexicalized reordering model plays a central role in phrase-based statistical machine translation sys- tems. The reordering model specifies the orientation for each phrase and calculates its probability con- ditioned on the phrase. In this paper, we describe the necessity and the challenge of introducing such a reordering model for hierarchical phrase-based translation. To deal with the challenge, we propose a novel lexicalized reordering model which is built directly on synchronous rules. For each target phrase contained in a rule, we calculate its orientation probability conditioned on the rule. We test our model on both small and large scale data. On NIST machine translation test sets, our reordering model achieved a 0.6-1.2 BLEU point improvements for Chinese-English translation over a strong baseline hierarchical phrase-based system. 1 Introduction In statistical machine translation, the problem of reordering source language into the word order of the target language remains a central research topic. Statistical phrase-based translation models (Och and Ney, 2004; Koehn et al., 2003) are good at local reordering, or the reordering of words within the phrase, since the order is specified by phrasal translations. However, phrase-based models remain weak at long-distance reordering, or the reordering of the phrases. To improve the reordering of the phrases, two types of models have been developed. The first one is lexicalized reordering models (Tillman, 2004; Huang et al., 2005; Al-Onaizan and Papineni, 2006; Nagata et al., 2006; Xiong et al., 2006; Zens and Ney, 2006; Koehn et al., 2007; Gal- ley and Manning, 2008; Cherry et al., 2012) which predict reordering by taking advantage of lexical information. The model in (Koehn et al., 2007) distinguishes three orientations with respect to the pre- vious and the next phrase—monotone (M), swap (S) and discontinuous (D). For example, we can ex- tract a phrase pair “xiayou ||| the lower reach of” whose orientations with respect to the previous and the next phrase are D and S respectively, as shown in Figure 1. Such a model is simple and effective, and has become a standard component of phrase-based systems such as MOSES. Figure 1. Phrase orientations for Chinese-English translation. The other is a hierarchical phrase-based (HPB) translation model (Chiang, 2007) based on synchro- nous grammar. In the HPB model, a synchronous grammar rule may contain both terminals (words) and nonterminals (sub-phrases). The order of terminals and nonterminal are specified by the rule. For This work is licenced under a Creative Commons Attribution 4.0 International License. Page numbers and proceedings footer are added by the organizers. License details: http://creativecommons.org/licenses/by/4.0/ 1144 example, the translation rule <X xiayou, the lower reach of X > specifies that the translation of sub phrase X before “xiayou” should …

Content truncated to 500 words