<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1041">
  <Title>Using Probabilistic Models as Predictors for a Symbolic Parser</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> There seems to be an upper limit for the level of quality that can be achieved by a parser if it is confined to information drawn from a single source. Stochastic parsers for English trained on the Penn Treebank have peaked their performance around 90% (Charniak, 2000). Parsing of German seems to be even harder and parsers trained on the NEGRA corpus or an enriched version of it still perform considerably worse. On the other hand, a great number of shallow components like taggers, chunkers, supertaggers, as well as general or specialized attachment predictors have been developed that might provide additional information to further improve the quality of a parser's output, as long as their contributions are in some sense complementory. Despite these prospects, such possibilities have rarely been investigated so far.</Paragraph>
    <Paragraph position="1"> To estimate the degree to which the desired synergy between heterogeneous knowledge sources can be achieved, we have established an experimental framework for syntactic analysis which allows us to plug in a wide variety of external predictor components, and to integrate their contributions as additional evidence in the general decision-making on the optimal structural interpretation. We refer to this approach as hybrid parsing because it combines different kinds of linguistic models, which have been acquired in totally different ways, ranging from manually compiled rule sets to statistically trained components.</Paragraph>
    <Paragraph position="2"> In this paper we investigate the benefit of external predictor components for the parsing quality which can be obtained with a rule-based grammar. For that purpose we trained a range of predictor components and integrated their output into the parser by means of soft constraints. Accordingly, the goal of our research was not to extensively optimize the predictor components themselves, but to quantify their contribution to the overall parsing quality. The results of these experiments not only lead to a better understanding of the utility of the different knowledge sources, but also allow us to derive empirically based priorities for further improving them. We are able to show that the potential of WCDG for information fusion is strong enough to accomodate even rather unreliable information from a wide range of predictor components. Using this potential we were able to reach a quality level for dependency parsing German which is unprecendented so far.</Paragraph>
  </Section>
class="xml-element"></Paper>