<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1005">
  <Title>Antecedent Recovery: Experiments with a Trace Tagger</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> One of the main contributions of this paper is that a two-step pre-processing approach to finding EEs outperforms both post-processing and in-processing.</Paragraph>
    <Paragraph position="1"> We found the pre-processing technique was successful because it used features not explicitly incorporated into the other models.</Paragraph>
    <Paragraph position="2"> Furthermore, we found that the result presented in Dienes and Dubey (2003), i.e. pre-processing is better for antecedent recovery than unlexicalized in-processing, also holds when comparing lexicalized models. However, comparing the lexicalized pre-processing system to the unlexicalized one, we find that although lexicalization results in much better trees, there is only a slight improvement in antecedent recovery.</Paragraph>
    <Paragraph position="3"> Third, we present a generalization of Model 3 of Collins (1999) to handle a broader range of EEs.</Paragraph>
    <Paragraph position="4"> While this particular model was not able to outperform the pre-processing method, it can be further developed into a parsing model which can handle non-local dependencies by incorporating the local cues we found relevant.</Paragraph>
    <Paragraph position="5"> In particular, a local window of five words, accompanied by the gap+ threads proved to be crucial.</Paragraph>
    <Paragraph position="6"> Thus we claim that, in order to detect long-distance dependencies, a robust stochastic parser should integrate lexical information as well as local cues cutting across phrase boundaries by either incorporating them into the probability model or using them in the beam-search.</Paragraph>
  </Section>
class="xml-element"></Paper>