Parsing the SynTagRus Treebank of Russian

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 641–648 Manchester, August 2008 Parsing the SYNTAGRUS Treebank of Russian Joakim Nivre V¨axj¨o University and Uppsala University joakim.nivre@vxu.se Igor M. Boguslavsky Universidad Polit´ecnica de Madrid Departamento de Inteligencia Artificial igor@opera.dia.fi.upm.es Leonid L. Iomdin Russian Academy of Sciences Institute for Information Transmission Problems iomdin@iitp.ru Abstract We present the first results on parsing the SYNTAGRUS treebank of Russian with a data-driven dependency parser, achieving a labeled attachment score of over 82% and an unlabeled attachment score of 89%. A feature analysis shows that high parsing accuracy is crucially dependent on the use of both lexical and morphological features. We conjecture that the latter result can be generalized to richly inflected languages in general, provided that sufficient amounts of training data are available. 1 Introduction Dependency-based syntactic parsing has become increasingly popular in computational linguistics in recent years. One of the reasons for the growing interest is apparently the belief that dependency- based representations should be more suitable for languages that exhibit free or flexible word order and where most of the clues to syntactic structure are found in lexical and morphological features, rather than in syntactic categories and word order configurations. Some support for this view can be found in the results from the CoNLL shared tasks on dependency parsing in 2006 and 2007, where a variety of data-driven methods for dependency parsing have been applied with encouraging results to languages of great typological diversity (Buch- holz and Marsi, 2006; Nivre et al., 2007a). However, there are still important differences in parsing accuracy for different language types. For © Joakim Nivre, Igor M. Boguslavsky, and Leonid L. Iomdin, 2008. Licensed under the Creative Com- mons Attribution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. example, Nivre et al. (2007a) observe that the lan- guages included in the 2007 CoNLL shared task can be divided into three distinct groups with re- spect to top accuracy scores, with relatively low accuracy for richly inflected languages like Arabic and Basque, medium accuracy for agglutinating languages like Hungarian and Turkish, and high accuracy for more configurational languages like English and Chinese. A complicating factor in this kind of comparison is the fact that the syntactic an- notation in treebanks varies across languages, in such a way that it is very difficult to tease apart the impact on parsing accuracy of linguistic structure, on the one hand, and linguistic annotation, on the other. It is also worth noting that the majority of the data sets used in the CoNLL shared tasks are not derived from treebanks with genuine depen- dency annotation, but have been obtained through conversion from other kinds of annotation. And the data sets that do come with original depen- dency annotation are generally fairly small, with less than 100,000 words available for training, the notable exception of course being the Prague De- pendency Treebank of Czech (Hajiˇc et al., 2001), which is one of the largest and most widely used treebanks in …

Content truncated to 500 words