Learning to Translate with Source and Target Syntax

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1443–1452, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Learning to Translate with Source and Target Syntax David Chiang USC Information Sciences Institute 4676 Admiralty Way, Suite 1001 Marina del Rey, CA 90292 USA chiang@isi.edu Abstract Statistical translation models that try to capture the recursive structure of language have been widely adopted over the last few years. These models make use of vary- ing amounts of information from linguis- tic theory: some use none at all, some use information about the grammar of the tar- get language, some use information about the grammar of the source language. But progress has been slower on translation models that are able to learn the rela- tionship between the grammars of both the source and target language. We dis- cuss the reasons why this has been a chal- lenge, review existing attempts to meet this challenge, and show how some old and new ideas can be combined into a sim- ple approach that uses both source and tar- get syntax for significant improvements in translation accuracy. 1 Introduction Statistical translation models that use synchronous context-free grammars (SCFGs) or related for- malisms to try to capture the recursive structure of language have been widely adopted over the last few years. The simplest of these (Chiang, 2005) make no use of information from syntactic theo- ries or syntactic annotations, whereas others have successfully incorporated syntactic information on the target side (Galley et al., 2004; Galley et al., 2006) or the source side (Liu et al., 2006; Huang et al., 2006). The next obvious step is toward mod- els that make full use of syntactic information on both sides. But the natural generalization to this setting has been found to underperform phrase- based models (Liu et al., 2009; Ambati and Lavie, 2008), and researchers have begun to explore so- lutions (Zhang et al., 2008; Liu et al., 2009). In this paper, we explore the reasons why tree- to-tree translation has been challenging, and how source syntax and target syntax might be used to- gether. Drawing on previous successful attempts to relax syntactic constraints during grammar extrac- tion in various ways (Zhang et al., 2008; Liu et al., 2009; Zollmann and Venugopal, 2006), we com- pare several methods for extracting a synchronous grammar from tree-to-tree data. One confounding factor in such a comparison is that some methods generate many new syntactic categories, making it more difficult to satisfy syntactic constraints at de- coding time. We therefore propose to move these constraints from the formalism into the model, im- plemented as features in the hierarchical phrase- based model Hiero (Chiang, 2005). This aug- mented model is able to learn from data whether to rely on syntax or not, or to revert back to mono- tone phrase-based translation. In experiments on Chinese-English and Arabic- English translation, we find that when both source and target syntax are made available to the model in an unobtrusive way, …

Content truncated to 500 words