Unsupervised Syntactic Alignment with Inversion Transduction Grammars

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 118–126, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Unsupervised Syntactic Alignment with Inversion Transduction Grammars Adam Pauls Dan Klein Computer Science Division University of California at Berkeley {adpauls,klein}@cs.berkeley.edu David Chiang Kevin Knight Information Sciences Institute University of Southern California {chiang,knight}@isi.edu Abstract Syntactic machine translation systems cur- rently use word alignments to infer syntactic correspondences between the source and tar- get languages. Instead, we propose an un- supervised ITG alignment model that directly aligns syntactic structures. Our model aligns spans in a source sentence to nodes in a target parse tree. We show that our model produces syntactically consistent analyses where possi- ble, while being robust in the face of syntactic divergence. Alignment quality and end-to-end translation experiments demonstrate that this consistency yields higher quality alignments than our baseline. 1 Introduction Syntactic machine translation has advanced signif- icantly in recent years, and multiple variants cur- rently achieve state-of-the-art translation quality. Many of these systems exploit linguistically-derived syntactic information either on the target side (Gal- ley et al., 2006), the source side (Huang et al., 2006), or both (Liu et al., 2009). Still others induce their syntax from the data (Chiang, 2005). Despite differ- ences in detail, the vast majority of syntactic meth- ods share a critical dependence on word alignments. In particular, they infer syntactic correspondences between the source and target languages through word alignment patterns, sometimes in combination with constraints from parser outputs. However, word alignments are not perfect indi- cators of syntactic alignment, and syntactic systems are very sensitive to word alignment behavior. Even a single spurious word alignment can invalidate a large number of otherwise extractable rules, while unaligned words can result in an exponentially large set of extractable rules to choose from. Researchers have worked to incorporate syntactic information into word alignments, resulting in improvements to both alignment quality (Cherry and Lin, 2006; DeN- ero and Klein, 2007), and translation quality (May and Knight, 2007; Fossum et al., 2008). In this paper, we remove the dependence on word alignments and instead directly model the syntactic correspondences in the data, in a manner broadly similar to Yamada and Knight (2001). In particu- lar, we propose an unsupervised model that aligns nodes of a parse tree (or forest) in one language to spans of a sentence in another. Our model is an in- stance of the inversion transduction grammar (ITG) formalism (Wu, 1997), constrained in such a way that one side of the synchronous derivation respects a syntactic parse. Our model is best suited to sys- tems which use source- or target-side trees only. The design of our model is such that, for divergent structures, a structurally integrated backoff to flatter word-level (or null) analyses is available. There- fore, our model is empirically robust to the case where syntactic divergence between languages pre- vents syntactically accurate ITG derivations. We show that, with appropriate pruning, our model can be efficiently trained …

Content truncated to 500 words