Annotation Projection-based Representation Learning for Cross-lingual Dependency Parsing

Proceedings of the 19th Conference on Computational Language Learning, pages 73–82, Beijing, China, July 30-31, 2015. c⃝2015 Association for Computational Linguistics Annotation Projection-based Representation Learning for Cross-lingual Dependency Parsing Min Xiao and Yuhong Guo Department of Computer and Information Sciences Temple University, Philadelphia, PA 19122, USA {minxiao,yuhong}@temple.edu Abstract Cross-lingual dependency parsing aims to train a dependency parser for an annotation-scarce target language by ex- ploiting annotated training data from an annotation-rich source language, which is of great importance in the field of nat- ural language processing. In this paper, we propose to address cross-lingual de- pendency parsing by inducing latent cross- lingual data representations via matrix completion and annotation projections on a large amount of unlabeled parallel sen- tences. To evaluate the proposed learn- ing technique, we conduct experiments on a set of cross-lingual dependency parsing tasks with nine different languages. The experimental results demonstrate the effi- cacy of the proposed learning method for cross-lingual dependency parsing. 1 Introduction The natural language processing (NLP) commu- nity has witnessed an enormous development of multilingual resources, which draws increasing at- tention to developing cross-lingual NLP adapta- tion systems. Cross-lingual dependency parsing aims to train a dependency parser for a target lan- guage where labeled data is rare or unavailable by exploiting the abundant annotated data from a source language. Cross-lingual dependency pars- ing can effectively reduce the expensive manual annotation effort in individual languages and has been increasingly studied in the multilingual com- munity. Previous works have demonstrated the success of cross-lingual dependency parsing for a variety of languages (Durrett et al., 2012; McDon- ald et al., 2013; T¨ackstr¨om et al., 2013; Søgaard and Wulff, 2012). One fundamental issue of cross-lingual depen- dency parsing lies in how to effectively transfer the annotation information from the source language domain to the target language domain. Due to the language divergence over the word-level represen- tations and the sentence structures, simply training a monolingual dependency parser on the labeled source language data without adaptation learn- ing will fail to produce a dependency parser that works in the target language domain. To tackle this problem, a variety of works in the literature have designed better algorithms to exploit the an- notated resources in the source languages, includ- ing the cross-lingual annotation projection meth- ods (Hwa et al., 2005; Smith and Eisner, 2009; Zhao et al., 2009), the cross-lingual direct trans- fer with linguistic constraints methods (Ganchev et al., 2009; Naseem et al., 2010; Naseem et al., 2012), and the cross-lingual representation learn- ing methods (Durrett et al., 2012; T¨ackstr¨om et al., 2012; Zhang et al., 2012). In this work, we propose a novel representation learning method to address cross-lingual depen- dency parsing, which exploits annotation projec- tions on a large amount of unlabeled parallel sen- tences to induce latent cross-lingual features via matrix completion. It combines the advantages of the cross-lingual annotation projection meth- ods, which project labeled information into the tar- get language domain, and the cross-lingual rep- resentation learning methods, which learn latent interlingual …

Content truncated to 500 words