<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-2006">
  <Title>Finding non-local dependencies: beyond pattern matching</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
7 Conclusions and future work
</SectionTitle>
    <Paragraph position="0"> We have presented an algorithm for recovering long-distance dependencies in local dependency structures. We extend the pattern matching approach of Johnson (2002) with machine learning techniques, and use dependency structures instead of constituency trees. Evaluation on the Penn Treebank shows an increase in accuracy.</Paragraph>
    <Paragraph position="1"> However, we do not have yet satisfactory results when working on a parser output. The conversion algorithm and the dependency labels we use are largely based on the Penn Treebank annotation, and it seems difficult to use them with the output of a parser.</Paragraph>
    <Paragraph position="2"> A parsing accuracy evaluation scheme based on grammatical relations (GR), presented in (Briscoe et al., 2002), provides a set of dependency labels (grammatical relations) and a manually annotated dependency corpus. Non-local dependencies are also annotated there, although no explicit difference is made between local and non-local dependencies.</Paragraph>
    <Paragraph position="3"> Since our classification algorithm does not depend on a particular set of dependency labels, we can also use the set of labels described by Briscoe et al, if we convert Penn Treebank to a GR-based dependency treebank and use it as the training corpus. This will allow us to make the patterns independent of the Penn Treebank annotation details and simplify testing the algorithm with a parser'u output. We will also be able to use the flexible and parameterizable scoring schemes discussed in (Briscoe et al., 2002).</Paragraph>
    <Paragraph position="4"> We also plan to develop the approach by using iteration of our non-local relations extraction algorithm, i.e., by running the algorithm, inserting the found non-local dependencies, running it again etc., until no new dependencies are found. While raising an important and interesting issue of the order in which we examine our patterns, we believe that this will allow us to handle very long extraction chains, like the one in sentence &amp;quot;Aichi revised its tax calculations after being challenged for allegedly failing to report. . . &amp;quot;, where Aichi is a (non-local) dependent of five verbs. Iteration of the algorithm will also help to increase the coverage (which is 93,7% with our 16 non-iterated patterns).</Paragraph>
  </Section>
class="xml-element"></Paper>