Joint Parsing and Translation

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 707–715, Beijing, August 2010 Joint Parsing and Translation Yang Liu and Qun Liu Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences {yliu,liuqun} @ict.ac.cn Abstract Tree-based translation models, which ex- ploit the linguistic syntax of source lan- guage, usually separate decoding into two steps: parsing and translation. Although this separation makes tree-based decoding simple and efficient, its translation perfor- mance is usually limited by the number of parse trees offered by parser. Alter- natively, we propose to parse and trans- late jointly by casting tree-based transla- tion as parsing. Given a source-language sentence, our joint decoder produces a parse tree on the source side and a transla- tion on the target side simultaneously. By combining translation and parsing mod- els in a discriminative framework, our ap- proach significantly outperforms a forest- based tree-to-string system by 1.1 ab- solute BLEU points on the NIST 2005 Chinese-English test set. As a parser, our joint decoder achieves an F1 score of 80.6% on the Penn Chinese Treebank. 1 Introduction Recent several years have witnessed the rapid development of syntax-based translation models (Chiang, 2007; Galley et al., 2006; Shen et al., 2008; Quirk et al., 2005; Liu et al., 2006; Huang et al., 2006; Eisner, 2003; Zhang et al., 2008; Chi- ang, 2010), which incorporate formal or linguis- tic syntax into translation process. Depending on whether modeling the linguistic syntax of source language or not, we divide them into two cate- gories: string-based and tree-based models. 1 1Mi et al. (2008) also distinguish between string-based and tree-based models but depending on the type of input. source target parse+translate string tree string source target string parse tree translate string (a) (b) Figure 1: Tree-based decoding: (a) separate pars- ing and translation versus (b) joint parsing and translation. String-based models include string-to-string (Chiang, 2007) and string-to-tree (Galley et al., 2006; Shen et al., 2008). Regardless of the syn- tactic information on the source side, they treat decoding as a parsing problem: the decoder parses a source-language sentence using the source pro- jection of a synchronous grammar while building the target sub-translations in parallel. Tree-based models include tree-to-string (Liu et al., 2006; Huang et al., 2006) and tree-to-tree (Quirk et al., 2005; Eisner, 2003; Zhang et al., 2008; Chiang, 2010). These models explicitly use source parse trees and divide decoding into two separate steps: parsing and translation. parser first parses a source-language sentence intoA a parse tree, and then a decoder converts the tree to a translation on the target side (see Figure 1(a)). Figure 2 gives a training example for tree-to- string translation, which consists of a Chinese tree, an English sentence, and the word align- ment between them. Romanized Chinese words are given to facilitate identification. Table 1 shows 707 �� � �9 �1 � �! NR P NR VV AS NN NPB NPB NPB PP VPB VP IP bushi yu shalong juxing le huitan Bush held a …

Content truncated to 500 words