Rich Bitext Projection Features for Parse Reranking

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 282–290, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Rich bitext projection features for parse reranking Alexander Fraser Renjing Wang Institute for Natural Language Processing University of Stuttgart {fraser,wangrg}@ims.uni-stuttgart.de Hinrich Sch¨utze Abstract Many different types of features have been shown to improve accuracy in parse reranking. A class of features that thus far has not been considered is based on a pro- jection of the syntactic structure of a trans- lation of the text to be parsed. The intu- ition for using this type of bitext projec- tion feature is that ambiguous structures in one language often correspond to un- ambiguous structures in another. We show that reranking based on bitext projection features increases parsing accuracy signif- icantly. 1 Introduction Parallel text or bitext is an important knowledge source for solving many problems such as ma- chine translation, cross-language information re- trieval, and the projection of linguistic resources from one language to another. In this paper, we show that bitext-based features are effective in ad- dressing another NLP problem, increasing the ac- curacy of statistical parsing. We pursue this ap- proach for a number of reasons. First, one lim- iting factor for syntactic approaches to statistical machine translation is parse quality (Quirk and Corston-Oliver, 2006). Improved parses of bi- text should result in improved machine translation. Second, as more and more texts are available in several languages, it will be increasingly the case that a text to be parsed is itself part of a bitext. Third, we hope that the improved parses of bitext will serve as higher quality training data for im- proving monolingual parsing using a process sim- ilar to self-training (McClosky et al., 2006). It is well known that different languages encode different types of grammatical information (agree- ment, case, tense etc.) and that what can be left unspecified in one language must be made explicit NP NP NP DT a NN baby CC and NP DT a NN woman SBAR who had gray hair Figure 1: English parse with high attachment in another. This information can be used for syn- tactic disambiguation. However, it is surprisingly hard to do this well. We use parses and alignments that are automatically generated and hence imper- fect. German parse quality is considered to be worse than English parse quality, and the annota- tion style is different, e.g., NP structure in German is flatter. We conduct our research in the framework of N-best parse reranking, but apply it to bitext and add only features based on syntactic projection from German to English. We test the idea that, generally, English parses with more isomorphism with respect to the projected German parse are bet- ter. The system takes as input (i) English sen- tences with a list of automatically generated syn- tactic parses, (ii) a translation of the English sen- tences into German, (iii) an automatically gen- erated parse of the German translation, and …

Content truncated to 500 words