Zero-resource Dependency Parsing: Boosting Delexicalized Cross-lingual Transfer with Linguistic Knowledge

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 119–130, Osaka, Japan, December 11-17 2016. Zero-resource Dependency Parsing: Boosting Delexicalized Cross-lingual Transfer with Linguistic Knowledge Lauriane Aufrant1,2 and Guillaume Wisniewski1 and Franc¸ois Yvon1 1LIMSI, CNRS, Univ. Paris-Sud, Universit´e Paris-Saclay, 91 405 Orsay, France 2DGA, 60 boulevard du G´en´eral Martial Valin, 75 509 Paris, France {lauriane.aufrant,guillaume.wisniewski,francois.yvon}@limsi.fr Abstract This paper studies cross-lingual transfer for dependency parsing, focusing on very low-resource settings where delexicalized transfer is the only fully automatic option. We show how to boost parsing performance by rewriting the source sentences so as to better match the linguistic regu- larities of the target language. We contrast a data-driven approach with an approach relying on linguistically motivated rules automatically extracted from the World Atlas of Language Struc- tures. Our findings are backed up by experiments involving 40 languages. They show that both approaches greatly outperform the baseline, the knowledge-driven method yielding the best accu- racies, with average improvements of +2.9 UAS, and up to +90 UAS (absolute) on some frequent PoS configurations. 1 Introduction The need to automatically process an increasing number of languages has made obvious the extreme de- pendency of standard development pipelines on in-domain, annotated resources that are required to train efficient statistical models. However, for most languages, annotated corpora only exist for a restricted number of domains, when they exist at all. In response to such low-resource scenario, four main strate- gies have been considered. The first is to hire experts and handcraft these resources, possibly with the help of active learning techniques: Garrette and Baldridge (2013) show that this strategy can be effec- tive and probably cheaper than expected. An alternative is to use models learned on some resource-rich source language(s) to process a low-resource target language; note that this is only possible once the source and target data have been mapped into a shared representation space (Zeman and Resnik, 2008). When source-target parallel corpora are available, a third approach projects annotations across languages via alignment links (Yarowsky and Ngai, 2001; Hwa et al., 2005; Lacroix et al., 2016). A variant using artificial parallel corpora, obtained via Machine Translation, is suggested and discussed by Tiedemann et al. (2014). In this work, we focus on the problem of learning dependency parsers for an under-resourced lan- guage and consider the delexicalized transfer approach of Zeman and Resnik (2008), in which the shared source-target representation is obtained by replacing all tokens by their PoS (assuming a common tagset). Thanks to this language-independent representation, a model trained with annotated sentences in a source language can be readily applied to parse sentences in any other language. Delexicalized tech- niques are especially useful in very low-resource settings, in which existing parallel corpora are likely to be too small or even non-existing. The development of cross-linguistically homogeneous and consis- tent schemes for PoS labels (Petrov et al., 2012) and, more recently, for dependency trees (McDonald et al., 2013) has been of great help to improve the applicability and effectiveness …

Content truncated to 500 words