Parser Adaptation and Projection with Quasi-Synchronous Grammar Features

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 822–831, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Parser Adaptation and Projection with Quasi-Synchronous Grammar Features∗ David A. Smith Department of Computer Science University of Massachusetts Amherst Amherst, MA 01003, USA dasmith@cs.umass.edu Jason Eisner Department of Computer Science Johns Hopkins University Baltimore, MD 21218, USA jason@cs.jhu.edu Abstract We connect two scenarios in structured learning: adapting a parser trained on one corpus to another annotation style, and projecting syntactic annotations from one language to another. We propose quasi- synchronous grammar (QG) features for these structured learning tasks. That is, we score a aligned pair of source and target trees based on local features of the trees and the alignment. Our quasi-synchronous model assigns positive probability to any alignment of any trees, in contrast to a syn- chronous grammar, which would insist on some form of structural parallelism. In monolingual dependency parser adap- tation, we achieve high accuracy in trans- lating among multiple annotation styles for the same sentence. On the more difficult problem of cross-lingual parser projection, we learn a dependency parser for a target language by using bilin- gual text, an English parser, and auto- matic word alignments. Our experiments show that unsupervised QG projection im- proves on parses trained using only high- precision projected annotations and far outperforms, by more than 35% absolute dependency accuracy, learning an unsu- pervised parser from raw target-language text alone. When a few target-language parse trees are available, projection gives a boost equivalent to doubling the number of target-language trees. ∗The first author would like to thank the Center for Intel- ligent Information Retrieval at UMass Amherst. We would also like to thank Noah Smith and Rebecca Hwa for helpful discussions and the anonymous reviewers for their sugges- tions for improving the paper. 1 Introduction 1.1 Parser Adaptation Consider the problem of learning a dependency parser, which must produce a directed tree whose vertices are the words of a given sentence. There are many differing conventions for representing syntactic relations in dependency trees. Say that we wish to output parses in the Prague style and so have annotated a small target corpus—e.g., 100 sentences—with those conventions. A parser trained on those hundred sentences will achieve mediocre dependency accuracy (the proportion of words that attach to their correct parent). But what if we also had a large number of trees in the CoNLL style (the source corpus)? Ide- ally they should help train our parser. But unfor- tunately, a parser that learned to produce perfect CoNLL-style trees would, for example, get both links “wrong” when its coordination constructions were evaluated against a Prague-style gold stan- dard (Figure 1). If it were just a matter of this one construction, the obvious solution would be to write a few rules by hand to transform the large source training cor- pus into the target style. Suppose, however, that there were many more ways that our corpora dif- fered. Then we would like to learn …

Content truncated to 500 words