Tree-based Translation without using Parse Trees

Proceedings of COLING 2012: Technical Papers, pages 3037–3054, COLING 2012, Mumbai, December 2012. Tree­based Translation without Using Parse Trees Feifei ZHAI, Jiajun ZHANG, Yu ZHOU and Chengqing ZONG National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China {ffzhai, jjzhang, yzhou, cqzong}@nlpr.ia.ac.cn ABSTRACT Parse trees are indispensable to the existing tree-based translation models. However, there exist two major challenges in utilizing parse trees: 1) For most language pairs, it is hard to get parse trees due to the lack of syntactic resources for training. 2) Numerous parse trees are not compatible with word alignment which is generally learned by GIZA++. Therefore, a number of useful translation rules are often excluded. To overcome these two problems, in this paper we make a great effort to bypass the parse trees and induce effective unsupervised trees for tree- based translation models. Our unsupervised trees depend only on the word alignment without utilizing any syntactic resource or linguistic parser. Hence, they are very beneficial for the translation between resource-poor languages. Our experimental results have shown that the string-to-tree translation system using our unsupervised trees significantly outperforms the string- to-tree system using parse trees. KEYWORDS : Tree-based translation; Unsupervised tree; EM algorithm. 3037 1 Introduction Recently, tree-based models1 have been widely studied in statistical machine translation (SMT). The existing tree-based models include string-to-tree models (Galley et al., 2006; Marcu et al., 2006; Shen et al., 2008), tree-to-string models (Quirk et al., 2005; Liu et al., 2006; Huang et al., 2006;), and tree-to-tree models (Eisner, 2003; Ding and Palmer, 2005; Cowan et al., 2006; Zhang et al., 2008; Liu et al., 2009). Due to the effective use of syntactic information, tree-based models have achieved comparable (Liu et al., 2009) and even better performance over phrase-based models (Marcu et al., 2006). source sentence word alignment target sentence target parse tree rule extraction parser word alignment (a) target sentence source sentence word alignment target unsuperveised tree rule extraction (b) FIGURE 1 – Rule extraction for string-to-tree translation model: (a) using parse trees versus (b) using our unsupervised trees. In the existing tree-based translation models, parse trees are essential to extracting translation rules. FIGURE 1(a) illustrates the rule extraction process of string-to-tree translation model. The parse tree is usually generated by a linguistic parser which is trained on a manually annotated corpus, such as Treebank. However, the manually annotated corpus is always too inadequate to fully display the strengths of tree-based models. In particular, traditional tree-based systems can not work at all for language pairs without any syntactic resource, which has greatly restricted their application. From FIGURE 1(a), we can also discover that syntactic parsing is completely independent of word alignment. The separation of parser and alignment leads to a severe incompatibility problem between them. Together with the widely existing parsing errors, numerous useful translation rules are excluded during rule extraction. To overcome the above two problems of current tree-based models, in this paper, we give up using parse trees and induce better alternatives for …

Content truncated to 500 words