A Novel Reordering Model Based on Multi-layer Phrase for Statistical Machine Translation

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 447–455, Beijing, August 2010 A Novel Reordering Model Based on Multi-layer Phrase for Sta- tistical Machine Translation Yanqing He1, Yu Zhou2, Chengqing Zong2, Huilin Wang1 1Institute of Scientific and Technical Information of China {heyq,wanghl}@istic.ac.cn 2Institute of Automation, Chinese Academy of Sciences {yzhou,cqzong}@nlpr.ia.ac.cn Abstract Phrase reordering is of great importance for statistical machine translation. Ac- cording to the movement of phrase trans- lation, the pattern of phrase reordering can be divided into three classes: mono- tone, BTG (Bracket Transduction Grammar) and hierarchy. It is a good way to use different styles of reordering models to reorder different phrases ac- cording to the characteristics of both the reordering models and phrases itself. In this paper a novel reordering model based on multi-layer phrase (PRML) is proposed, where the source sentence is segmented into different layers of phras- es on which different reordering models are applied to get the final translation. This model has some advantages: differ- ent styles of phrase reordering models are easily incorporated together; when a complicated reordering model is em- ployed, it can be limited in a smaller scope and replaced with an easier reor- dering model in larger scope. So this model better trade-offs the translation speed and performance simultaneously. 1 Introduction In statistical machine translation (SMT), phrase reordering is a complicated problem. According to the type of phrases, the existing phrase reor- dering models are divided into two categories: contiguous phrase-based reordering models and non-contiguous phrase-based reordering models. Contiguous phrase-based reordering models are designed to reorder contiguous phrases. In such type of reordering models, a contiguous phrase is reordered as a unit and the movements of phrase don’t involve insertions inside the oth- er phrases. Some of these models are content- independent, such as distortion models (Och and Ney, 2004; Koehn et al., 2003) which penalize translation according to jump distance of phrases, and flat reordering model (Wu, 1995; Zens et al., 2004)which assigns constant probabilities for monotone order and non-monotone order. These reordering models are simple and the contents of phrases have not been considered. So it’s hard to obtain a satisfactory translation performance. Some lexicalized reordering models (Och et al., 2004; Tillmann 2004, Kumar and Byrne, 2005, Koehn et al., 2005) learn local orientations (mo- notone or non-monotone) with probabilities for each bilingual phrase from training data. These models are phrase-dependent, so improvements over content-independent reordering models are obtained. However, many parameters need to be estimated. Non-contiguous phrase-based reordering models are proposed to process non-contiguous phrases and the movements of phrase involve insertion operations. This type of reordering models mainly includes all kinds of syntax- based models where more structural information is employed to obtain a more flexible phrase movement. Linguistically syntactic approaches (Yamada and Knight, 2001; Galley et al., 2004, 2006; Marcu et al., 2006; Liu et al., 2006; Shie- ber et al., 1990; Eisner, 2003; Quirk et al., 2005; Ding and Palmer, 2005) employ linguistically syntactic information to enhance their reordering …

Content truncated to 500 words