<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1013">
  <Title>Pseudo-Projective Dependency Parsing</Title>
  <Section position="2" start_page="0" end_page="99" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> It is sometimes claimed that one of the advantages of dependency grammar over approaches based on constituency is that it allows a more adequate treatment of languages with variable word order, where discontinuous syntactic constructions are more common than in languages like English (Mel'Vcuk, 1988; Covington, 1990). However, this argument is only plausible if the formal framework allows non-projective dependency structures, i.e. structures where a head and its dependents may correspond to a discontinuous constituent. From the point of view of computational implementation this can be problematic, since the inclusion of non-projective structures makes the parsing problem more complex and therefore compromises efficiency and in practice also accuracy and robustness. Thus, most broad-coverage parsers based on dependency grammar have been restricted to projective structures.</Paragraph>
    <Paragraph position="1"> This is true of the widely used link grammar parser for English (Sleator and Temperley, 1993), which uses a dependency grammar of sorts, the probabilistic dependency parser of Eisner (1996), and more recently proposed deterministic dependency parsers (Yamada and Matsumoto, 2003; Nivre et al., 2004).</Paragraph>
    <Paragraph position="2"> It is also true of the adaptation of the Collins parser for Czech (Collins et al., 1999) and the finite-state dependency parser for Turkish by Oflazer (2003).</Paragraph>
    <Paragraph position="3"> This is in contrast to dependency treebanks, e.g.</Paragraph>
    <Paragraph position="4"> Prague Dependency Treebank (HajiVc et al., 2001b), Danish Dependency Treebank (Kromann, 2003), and the METU Treebank of Turkish (Oflazer et al., 2003), which generally allow annotations with non-projective dependency structures. The fact that projective dependency parsers can never exactly reproduce the analyses found in non-projective treebanks is often neglected because of the relative scarcity of problematic constructions. While the proportion of sentences containing non-projective dependencies is often 15-25%, the total proportion of non-projective arcs is normally only 1-2%. As long as the main evaluation metric is dependency accuracy per word, with state-of-the-art accuracy mostly below 90%, the penalty for not handling non-projective constructions is almost negligible. Still, from a theoretical point of view, projective parsing of non-projective structures has the drawback that it rules out perfect accuracy even as an asymptotic goal.</Paragraph>
    <Paragraph position="5">  (&amp;quot;Only one of them concerns quality.&amp;quot;)</Paragraph>
    <Paragraph position="7"> There exist a few robust broad-coverage parsers that produce non-projective dependency structures, notably Tapanainen and J&amp;quot;arvinen (1997) and Wang and Harper (2004) for English, Foth et al. (2004) for German, and Holan (2004) for Czech. In addition, there are several approaches to non-projective dependency parsing that are still to be evaluated in the large (Covington, 1990; Kahane et al., 1998; Duchier and Debusmann, 2001; Holan et al., 2001; Hellwig, 2003). Finally, since non-projective constructions often involve long-distance dependencies, the problem is closely related to the recovery of empty categories and non-local dependencies in constituency-based parsing (Johnson, 2002; Dienes and Dubey, 2003; Jijkoun and de Rijke, 2004; Cahill et al., 2004; Levy and Manning, 2004; Campbell, 2004).</Paragraph>
    <Paragraph position="8"> In this paper, we show how non-projective dependency parsing can be achieved by combining a data-driven projective parser with special graph transformation techniques. First, the training data for the parser is projectivized by applying a minimal number of lifting operations (Kahane et al., 1998) and encoding information about these lifts in arc labels.</Paragraph>
    <Paragraph position="9"> When the parser is trained on the transformed data, it will ideally learn not only to construct projective dependency structures but also to assign arc labels that encode information about lifts. By applying an inverse transformation to the output of the parser, arcs with non-standard labels can be lowered to their proper place in the dependency graph, giving rise 1The dependency graph has been modified to make the final period a dependent of the main verb instead of being a dependent of a special root node for the sentence.</Paragraph>
    <Paragraph position="10"> to non-projective structures. We call this pseudo-projective dependency parsing, since it is based on a notion of pseudo-projectivity (Kahane et al., 1998).</Paragraph>
    <Paragraph position="11"> The rest of the paper is structured as follows.</Paragraph>
    <Paragraph position="12"> In section 2 we introduce the graph transformation techniques used to projectivize and deprojectivize dependency graphs, and in section 3 we describe the data-driven dependency parser that is the core of our system. We then evaluate the approach in two steps.</Paragraph>
    <Paragraph position="13"> First, in section 4, we evaluate the graph transformation techniques in themselves, with data from the Prague Dependency Treebank and the Danish Dependency Treebank. In section 5, we then evaluate the entire parsing system by training and evaluating on data from the Prague Dependency Treebank.</Paragraph>
  </Section>
class="xml-element"></Paper>