<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1019">
  <Title>Modelling the substitutability of discourse connectives</Title>
  <Section position="7" start_page="154" end_page="154" type="evalu">
    <SectionTitle>
4.3.2 Results
</SectionTitle>
    <Paragraph position="0"> A leave-one-out cross validation procedure was used. For each triple &lt;p,q,qprime&gt; , the data concerning the pairs p,q and p,qprime were held back, and the remaining data used to construct the models. The results are shown in Table 4. For comparison, a random baseline classifier achieves 50% accuracy.</Paragraph>
    <Paragraph position="1"> The results demonstrate the utility of the new variance-based function V . The new variance-based function V is better than KL divergence at distinguishing HYPONYMY from SYNONYMY (kh2 =</Paragraph>
    <Paragraph position="3"> worse on the coarser grained task. This is consistent with the expectations of Table 1. The two classifiers were also combined by making a naive Bayes assumption. This gave an accuracy of 76.1% on the first task, which is significantly better than just using KL divergence (kh2 = 5.65,df = 1,p &lt; 0.05), and not significantly worse than using V . The combination's accuracy on the second task was 76.2%, which is about the same as using KL divergence.</Paragraph>
    <Paragraph position="4"> This shows that combining similarity- and variance-based measures can be useful can improve overall performance.</Paragraph>
  </Section>
class="xml-element"></Paper>