Transforming Trees to Improve Syntactic Convergence

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 863–872, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Transforming Trees to Improve Syntactic Convergence David Burkett and Dan Klein Computer Science Division University of California, Berkeley {dburkett,klein}@cs.berkeley.edu Abstract We describe a transformation-based learning method for learning a sequence of mono- lingual tree transformations that improve the agreement between constituent trees and word alignments in bilingual corpora. Using the manually annotated English Chinese Transla- tion Treebank, we show how our method au- tomatically discovers transformations that ac- commodate differences in English and Chi- nese syntax. Furthermore, when transforma- tions are learned on automatically generated trees and alignments from the same domain as the training data for a syntactic MT system, the transformed trees achieve a 0.9 BLEU im- provement over baseline trees. 1 Introduction Monolingually, many Treebank conventions are more or less equally good. For example, the En- glish WSJ treebank (Marcus et al., 1993) attaches verbs to objects rather than to subjects, and it at- taches prepositional modifiers outside of all quan- tifiers and determiners. The former matches most linguistic theories while the latter does not, but to a monolingual parser, these conventions are equally learnable. However, once bilingual data is involved, such treebank conventions entail constraints on rule extraction that may not be borne out by semantic alignments. To the extent that there are simply di- vergences in the syntactic structure of the two lan- guages, it will often be impossible to construct syn- tax trees that are simultaneously in full agreement with monolingual linguistic theories and with the alignments between sentences in both languages. To see this, consider the English tree in Figure 1a, taken from the English side of the English Chi- nese Translation Treebank (Bies et al., 2007). The lowest VP in this tree is headed by ‘select,’ which aligns to the Chinese verb ‘ 挑选.’ However, ‘ 挑 选’ also aligns to the other half of the English in- finitive, ‘to,’ which, following common English lin- guistic theory, is outside the VP. Because of this violating alignment, many syntactic machine trans- lation systems (Galley et al., 2004; Huang et al., 2006) won’t extract any translation rules for this constituent. However, by applying a simple trans- formation to the English tree to set up the infinitive as its own constituent, we get the tree in Figure 1b, which may be less well-motivated linguistically, but which corresponds better to the Chinese-mediated semantics and permits the extraction of many more syntactic MT rules. In this work, we develop a method based on transformation-based learning (Brill, 1995) for au- tomatically acquiring a sequence of tree transforma- tions of the sort in Figure 1. Once the transformation sequence has been learned, it can be deterministi- cally applied to any parsed sentences, yielding new parse trees with constituency structures that agree better with the bilingual alignments yet remain con- sistent across the corpus. In particular, we use this …

Content truncated to 500 words