Linguistic features in data-driven dependency parsing

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 25–32 Manchester, August 2008 Linguistic features in data-driven dependency parsing Lilja Øvrelid NLP-unit, Dept. of Swedish University of Gothenburg Sweden lilja.ovrelid@svenska.gu.se Abstract This article investigates the effect of a set of linguistically motivated features on ar- gument disambiguation in data-driven de- pendency parsing of Swedish. We present results from experiments with gold stan- dard features, such as animacy, definite- ness and finiteness, as well as correspond- ing experiments where these features have been acquired automatically and show significant improvements both in overall parse results and in the analysis of specific argument relations, such as subjects, ob- jects and predicatives. 1 Introduction Data-driven dependency parsing has recently re- ceived extensive attention in the parsing commu- nity and impressive results have been obtained for a range of languages (Nivre et al., 2007). Even with high overall parsing accuracy, however, data- driven parsers often make errors in the assign- ment of argument relations such as subject and object and the exact influence of data-derived fea- tures on the parsing accuracy for specific linguistic constructions is still relatively poorly understood. There are a number of studies that investigate the influence of different features or representational choices on overall parsing accuracy, (Bod, 1998; Klein and Manning, 2003). There are also attempts at a more fine-grained analysis of accuracy, target- ing specific linguistic constructions or grammati- cal functions (Carroll and Briscoe, 2002; K¨ubler and Proki´c, 2006; McDonald and Nivre, 2007). c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. But there are few studies that combine the two per- spectives and try to tease apart the influence of dif- ferent features on the analysis of specific construc- tions, let alone motivated by a thorough linguistic analysis. In this paper, we investigate the influence of a set of linguistically motivated features on parse re- sults for Swedish, and in particular on the analysis of argument relations such as subjects, objects and subject predicatives. Motivated by an error anal- ysis of the best performing parser for Swedish in the CoNLL-X shared task, we extend the feature model employed by the parser with a set of lin- guistically motivated features and go on to show how these features may be acquired automatically. We then present results from corresponding parse experiments with automatic features. The rest of the paper is structured as follows. In section 2 we present relevant properties of Swedish morphosyntax, as well as the treebank and parser employed in the experiments. Section 3 presents an error analysis of the baseline parser and we go on to motivate a set of linguistic features in sec- tion 4, which are employed in a set of experiments with gold standard features, discussed in section 5. Section 6 presents the automatic acquisition of these features, with a particular focus on animacy classification and in section 7 we report parse ex- periments with automatic features. 2 Parsing Swedish Before we turn …

Content truncated to 500 words