<?xml version="1.0" standalone="yes"?>
<Paper uid="P01-1038">
  <Title>Generation of VP Ellipsis: A Corpus-Based Approach</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have found that the decision to elide VPs is statistically correlated with several factors, including distance between antecedent and candidate VPs by word or sentence, and the presence or absence of syntactic and discourse relations. These findings provide a strong foundation on which to build algorithms for the generation of VPE. We have explored several possible algorithms with the help of a machine learning system, and we have found that these automatically derived algorithms perform well on cross-validation tests.</Paragraph>
    <Paragraph position="1"> We have also seen that the decision whether or not to elide can be better made later in the generation process: the more features are available, the better. It is perhaps not surprising that the decision cannot be made very well just after after text planning: it is well known that VPE is subject to syntactic constraints, and the relevant information is not yet available. It is perhaps more surprising that the surface-oriented features appear to contribute to the quality of the decision, pushing the decision past the realization phase. One possible explanation is that there are in fact other features, which we have not yet identified, and for which the surface-oriented features are standins. If this is the case, further work will allow us to define algorithms so that the decision on VPE can be made after sentence planning. However, it is also possible that decisions about VPE (and related pronominal constraints) cannot be made before the text is linearized, presumably because of the processing limitations of the hearer/reader (and of the speaker/writer). Walker (1996) has argued in favor of the importance of limited attention in processing discourse phenomena, and the surface-oriented features can be argued to model such cognitive constraints.</Paragraph>
  </Section>
class="xml-element"></Paper>