Generalizing Tree Transformations for Inductive Dependency Parsing

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 968–975, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Generalizing Tree Transformations for Inductive Dependency Parsing Jens Nilsson∗ Joakim Nivre∗† ∗V¨axj¨o University, School of Mathematics and Systems Engineering, Sweden †Uppsala University, Dept. of Linguistics and Philology, Sweden {jni,nivre,jha}@msi.vxu.se Johan Hall∗ Abstract Previous studies in data-driven dependency parsing have shown that tree transformations can improve parsing accuracy for specific parsers and data sets. We investigate to what extent this can be generalized across languages/treebanks and parsers, focusing on pseudo-projective parsing, as a way of capturing non-projective dependencies, and transformations used to facilitate parsing of coordinate structures and verb groups. The results indicate that the beneficial effect of pseudo-projective parsing is independent of parsing strategy but sensitive to language or treebank specific properties. By contrast, the construction specific transformations appear to be more sensitive to parsing strategy but have a constant positive effect over several languages. 1 Introduction Treebank parsers are trained on syntactically anno- tated sentences and a major part of their success can be attributed to extensive manipulations of the train- ing data as well as the output of the parser, usually in the form of various tree transformations. This can be seen in state-of-the-art constituency-based parsers such as Collins (1999), Charniak (2000), and Petrov et al. (2006), and the effects of different trans- formations have been studied by Johnson (1998), Klein and Manning (2003), and Bikel (2004). Corre- sponding manipulations in the form of tree transfor- mations for dependency-based parsers have recently gained more interest (Nivre and Nilsson, 2005; Hall and Nov´ak, 2005; McDonald and Pereira, 2006; Nilsson et al., 2006) but are still less studied, partly because constituency-based parsing has dominated the field for a long time, and partly because depen- dency structures have less structure to manipulate than constituent structures. Most of the studies in this tradition focus on a par- ticular parsing model and a particular data set, which means that it is difficult to say whether the effect of a given transformation is dependent on a partic- ular parsing strategy or on properties of a particu- lar language or treebank, or both. The aim of this study is to further investigate some tree transforma- tion techniques previously proposed for data-driven dependency parsing, with the specific aim of trying to generalize results across languages/treebanks and parsers. More precisely, we want to establish, first of all, whether the transformation as such makes specific assumptions about the language, treebank or parser and, secondly, whether the improved pars- ing accuracy that is due to a given transformation is constant across different languages, treebanks, and parsers. The three types of syntactic phenomena that will be studied here are non-projectivity, coordination and verb groups, which in different ways pose prob- lems for dependency parsers. We will focus on tree transformations that combine preprocessing with post-processing, and where the parser is treated as a black box, such as the pseudo-projective parsing technique proposed by Nivre and Nilsson …

Content truncated to 500 words