Syntactic Transfer Using a Bilingual Lexicon

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1–11, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Syntactic Transfer Using a Bilingual Lexicon Greg Durrett, Adam Pauls, and Dan Klein Computer Science Division University of California, Berkeley {gdurrett,adpauls,klein}@cs.berkeley.edu Abstract We consider the problem of using a bilingual dictionary to transfer lexico-syntactic infor- mation from a resource-rich source language to a resource-poor target language. In con- trast to past work that used bitexts to trans- fer analyses of specific sentences at the token level, we instead use features to transfer the behavior of words at a type level. In a dis- criminative dependency parsing framework, our approach produces gains across a range of target languages, using two different low- resource training methodologies (one weakly supervised and one indirectly supervised) and two different dictionary sources (one manu- ally constructed and one automatically con- structed). 1 Introduction Building a high-performing parser for a language with no existing treebank is still an open problem. Methods that use no supervision at all (Klein and Manning, 2004) or small amounts of manual su- pervision (Haghighi and Klein, 2006; Cohen and Smith, 2009; Naseem et al., 2010; Berg-Kirkpatrick and Klein, 2010) have been extensively studied, but still do not perform well enough to be deployed in practice. Projection of dependency links across aligned bitexts (Hwa et al., 2005; Ganchev et al., 2009; Smith and Eisner, 2009) gives better perfor- mance, but crucially depends on the existence of large, in-domain bitexts. A more generally appli- cable class of methods exploits the notion of univer- sal part of speech tags (Petrov et al., 2011; Das and ... the senators demand strict new ethics rules ... DT NNS VBP JJ JJ NNS NNS Gewerkschaften verlangen Verzicht auf die Reform NN VVFIN NN APPR ART NN Unions demand abandonment on the reform Figure 1: Sentences in English and German both contain- ing words that mean “demand.” The fact that the English demand takes nouns on its left and right indicates that the German verlangen should do the same, correctly suggest- ing attachments to Verzicht and Gewerkschaften. Petrov, 2011) to train parsers that can run on any lan- guage with no adaptation (McDonald et al., 2011) or unsupervised adaptation (Cohen et al., 2011). While these universal parsers currently constitute the highest-performing methods for languages with- out treebanks, they are inherently limited by operat- ing at the coarse POS level, as lexical features are vital to supervised parsing models. In this work, we consider augmenting delexical- ized parsers by transferring syntactic information through a bilingual lexicon at the word type level. These parsers are delexicalized in the sense that, al- though they receive target language words as input, their feature sets do not include indicators on those words. This setting is appropriate when there is too little target language data to learn lexical features di- rectly. Our main approach is to add features which are lexical in the sense …

Content truncated to 500 words