Multi-Source Transfer of Delexicalized Dependency Parsers

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 62–72, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Multi-Source Transfer of Delexicalized Dependency Parsers Ryan McDonald Google New York, NY ryanmcd@google.com Slav Petrov Google New York, NY slav@google.com Keith Hall Google Z¨urich kbhall@google.com Abstract We present a simple method for transferring dependency parsers from source languages with labeled training data to target languages without labeled training data. We first demon- strate that delexicalized parsers can be di- rectly transferred between languages, produc- ing significantly higher accuracies than unsu- pervised parsers. We then use a constraint driven learning algorithm where constraints are drawn from parallel corpora to project the final parser. Unlike previous work on project- ing syntactic resources, we show that simple methods for introducing multiple source lan- guages can significantly improve the overall quality of the resulting parsers. The projected parsers from our system result in state-of-the- art performance when compared to previously studied unsupervised and projected parsing systems across eight different languages. 1 Introduction Statistical parsing has been one of the most active ar- eas of research in the computational linguistics com- munity since the construction of the Penn Treebank (Marcus et al., 1993). This includes work on phrase- structure parsing (Collins, 1997; Charniak, 2000; Petrov et al., 2006), dependency parsing (McDonald et al., 2005; Nivre et al., 2006) as well as a num- ber of other formalisms (Clark and Curran, 2004; Wang and Harper, 2004; Shen and Joshi, 2008). As underlying modeling techniques have improved, these parsers have begun to converge to high lev- els of accuracy for English newswire text. Subse- quently, researchers have begun to look at both port- ing these parsers to new domains (Gildea, 2001; Mc- Closky et al., 2006; Petrov et al., 2010) and con- structing parsers for new languages (Collins et al., 1999; Buchholz and Marsi, 2006; Nivre et al., 2007). One major obstacle in building statistical parsers for new languages is that they often lack the manu- ally annotated resources available for English. This observation has led to a vast amount of research on unsupervised grammar induction (Carroll and Charniak, 1992; Klein and Manning, 2004; Smith and Eisner, 2005; Cohen and Smith, 2009; Berg- Kirkpatrick and Klein, 2010; Naseem et al., 2010; Spitkovsky et al., 2010; Blunsom and Cohn, 2010). Grammar induction systems have seen large ad- vances in quality, but parsing accuracies still signif- icantly lag behind those of supervised systems. Fur- thermore, they are often trained and evaluated under idealized conditions, e.g., only on short sentences or assuming the existence of gold-standard part-of- speech (POS) tags.1 The reason for these assump- tions is clear. Unsupervised grammar induction is difficult given the complexity of the analysis space. These assumptions help to give the model traction. The study of unsupervised grammar induction has many merits. Most notably, it increases our under- standing of how computers (and possibly humans) learn in the absence of any explicit feedback. How- ever, the gold POS tag assumption …

Content truncated to 500 words