A Dependency Treelet String Correspondence Model for Statistical Machine Translation

Proceedings of the Second Workshop on Statistical Machine Translation, pages 40–47, Prague, June 2007. c⃝2007 Association for Computational Linguistics A Dependency Treelet String Correspondence Model for Statistical Machine Translation Deyi Xiong, Qun Liu and Shouxun Lin Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences Beijing, China, 100080 {dyxiong, liuqun, sxlin}@ict.ac.cn Abstract This paper describes a novel model using dependency structures on the source side for syntax-based statistical machine transla- tion: Dependency Treelet String Correspon- dence Model (DTSC). The DTSC model maps source dependency structures to tar- get strings. In this model translation pairs of source treelets and target strings with their word alignments are learned automatically from the parsed and aligned corpus. The DTSC model allows source treelets and tar- get strings with variables so that the model can generalize to handle dependency struc- tures with the same head word but with dif- ferent modifiers and arguments. Addition- ally, target strings can be also discontinuous by using gaps which are corresponding to the uncovered nodes which are not included in the source treelets. A chart-style decod- ing algorithm with two basic operations– substituting and attaching–is designed for the DTSC model. We argue that the DTSC model proposed here is capable of lexical- ization, generalization, and handling discon- tinuous phrases which are very desirable for machine translation. We finally evaluate our current implementation of a simplified ver- sion of DTSC for statistical machine trans- lation. 1 Introduction Over the last several years, various statistical syntax- based models were proposed to extend traditional word/phrase based models in statistical machine translation (SMT) (Lin, 2004; Chiang, 2005; Ding et al., 2005; Quirk et al., 2005; Marcu et al., 2006; Liu et al., 2006). It is believed that these models can improve the quality of SMT significantly. Com- pared with phrase-based models, syntax-based mod- els lead to better reordering and higher flexibility by introducing hierarchical structures and variables which make syntax-based models capable of hierar- chical reordering and generalization. Due to these advantages, syntax-based approaches are becoming an active area of research in machine translation. In this paper, we propose a novel model based on dependency structures: Dependency Treelet String Correspondence Model (DTSC). The DTSC model maps source dependency structures to target strings. It just needs a source language parser. In contrast to the work by Lin (2004) and by Quirk et al. (2005), the DTSC model does not need to generate target language dependency structures using source struc- tures and word alignments. On the source side, we extract treelets which are any connected subgraphs and consistent with word alignments. While on the target side, we allow the aligned target sequences to be generalized and discontinuous by introducing variables and gaps. The variables on the target side are aligned to the corresponding variables of treelets, while gaps between words or variables are corre- sponding to the uncovered nodes which are not in- cluded by treelets. To complete the translation pro- cess, we design two basic operations for the decod- …

Content truncated to 500 words