<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-1013">
  <Title>Probabilistic Parsing for German using Sister-Head Dependencies</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Treebank-based probabilistic parsing has been the subject of intensive research over the past few years, resulting in parsing models that achieve both broad coverage and high parsing accuracy (e.g., Collins 1997; Charniak 2000). However, most of the existing models have been developed for English and trained on the Penn Treebank (Marcus et al., 1993), which raises the question whether these models generalize to other languages, and to annotation schemes that differ from the Penn Treebank markup.</Paragraph>
    <Paragraph position="1"> The present paper addresses this question by proposing a probabilistic parsing model trained on Negra (Skut et al., 1997), a syntactically annotated corpus for German. German has a number of syntactic properties that set it apart from English, and the Negra annotation scheme differs in important respects from the Penn Treebank markup. While Negra has been used to build probabilistic chunkers (Becker and Frank, 2002; Skut and Brants, 1998), the research reported in this paper is the first attempt to develop a probabilistic full parsing model for German trained on a treebank (to our knowledge).</Paragraph>
    <Paragraph position="2"> Lexicalization can increase parsing performance dramatically for English (Carroll and Rooth, 1998; Charniak, 1997, 2000; Collins, 1997), and the lexicalized model proposed by Collins (1997) has been successfully applied to Czech (Collins et al., 1999) and Chinese (Bikel and Chiang, 2000). However, the resulting performance is significantly lower than the performance of the same model for English (see Table 1). Neither Collins et al. (1999) nor Bikel and Chiang (2000) compare the lexicalized model to an unlexicalized baseline model, leaving open the possibility that lexicalization is useful for English, but not for other languages.</Paragraph>
    <Paragraph position="3"> This paper is structured as follows. Section 2 reviews the syntactic properties of German, focusing on its semi-flexible wordorder. Section 3 describes two standard lexicalized models (Carroll and Rooth, 1998; Collins, 1997), as well as an unlexicalized baseline model. Section 4 presents a series of experiments that compare the parsing performance of these three models (and several variants) on Negra. The results show that both lexicalized models fail to out-perform the unlexicalized baseline. This is at odds with what has been reported for English. Learning curves show that the poor performance of the lexicalized models is not due to lack of training data.</Paragraph>
    <Paragraph position="4"> Section 5 presents an error analysis for Collins's (1997) lexicalized model, which shows that the head-head dependencies used in this model fail to cope well with the flat structures in Negra. We propose an alternative model that uses sister-head dependencies instead. This model outperforms the two original lexicalized models, as well as the unlexicalized baseline. Based on this result and on the review of the previous literature (Section 6), we argue (Section 7) that sister-head models are more appropriate for treebanks with very flat structures (such as Negra), typically used to annotate languages with semi-free wordorder (such as German).</Paragraph>
  </Section>
class="xml-element"></Paper>