Treebank Translation for Cross-Lingual Parser Induction

Proceedings of the Eighteenth Conference on Computational Language Learning, pages 130–140, Baltimore, Maryland USA, June 26-27 2014. c⃝2014 Association for Computational Linguistics Treebank Translation for Cross-Lingual Parser Induction J¨org Tiedemann Dep. of Linguistics and Philology Uppsala University jorg.tiedemann@lingfil.uu.se ˇZeljko Agi´c Linguistics Department University of Potsdam zagic@uni-potsdam.de Joakim Nivre Dep. of Linguistics and Philology Uppsala University joakim.nivre@lingfil.uu.se Abstract Cross-lingual learning has become a popu- lar approach to facilitate the development of resources and tools for low-density lan- guages. Its underlying idea is to make use of existing tools and annotations in resource-rich languages to create similar tools and resources for resource-poor lan- guages. Typically, this is achieved by either projecting annotations across parallel cor- pora, or by transferring models from one or more source languages to a target language. In this paper, we explore a third strategy by using machine translation to create syn- thetic training data from the original source- side annotations. Specifically, we apply this technique to dependency parsing, us- ing a cross-lingually unified treebank for adequate evaluation. Our approach draws on annotation projection but avoids the use of noisy source-side annotation of an unre- lated parallel corpus and instead relies on manual treebank annotation in combination with statistical machine translation, which makes it possible to train fully lexicalized parsers. We show that this approach signif- icantly outperforms delexicalized transfer parsing. 1 Introduction The lack of resources and tools is a serious problem for the majority of the world’s languages (Bender, 2013). Many applications require robust tools and the development of language-specific resources is expensive and time consuming. Furthermore, many tasks such as data-driven syntactic parsing require strong supervision to achieve reasonable results for real-world applications, since the performance of fully unsupervised methods lags behind by a large margin in comparison with the state of the art. Cross-lingual learning has been proposed as one possible solution to quickly create initial tools for languages that lack the appropriate resources (Ganchev and Das, 2013). By and large, there are two main strategies that have been proposed in the literature: annotation projection and model transfer. 1.1 Previous Cross-Lingual Approaches Annotation projection relies on the mapping of lin- guistic annotation across languages using paral- lel corpora and automatic alignment as basic re- sources (Yarowsky et al., 2001; Hwa et al., 2005; T¨ackstr¨om et al., 2013a). Tools that exist for the source language are used to annotate the source side of the corpus and projection heuristics are then applied to map the annotation through word align- ment onto the corresponding target language text. Target language tools can then be trained on the projected annotation assuming that the mapping is sufficiently correct. Less frequent, but also possi- ble, is the scenario where the source side of the cor- pus contains manual annotation (Agi´c et al., 2012). This addresses the problem created by projecting noisy annotations, but it presupposes parallel cor- pora with manual annotation, which are rarely avail- able, and expensive and time-consuming to pro- duce. Model transfer instead relies on universal fea- tures and …

Content truncated to 500 words