<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-1037">
  <Title>Parametric Models of Linguistic Count Data</Title>
  <Section position="6" start_page="525" end_page="525" type="concl">
    <SectionTitle>
4 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have presented theoretical and empirical evidence for zero-inflation among linguistic count data.</Paragraph>
    <Paragraph position="1"> Zero-inflated models can account for increased variation at least as well as overdispersed models on standard document classification tasks. Given the computational advantages of simple zero-inflated models, theycanandshouldbeusedinplaceofstandard models. For document classification, an event model based on a zero-inflated binomial distribution outperforms conventional Bernoulli and binomial models.</Paragraph>
  </Section>
class="xml-element"></Paper>