<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1819">
  <Title>SS</Title>
  <Section position="7" start_page="22" end_page="22" type="evalu">
    <SectionTitle>
5 Evaluation Results
</SectionTitle>
    <Paragraph position="0"> To evaluate the generalization ability of the acquired rules, 5-fold cross validation tests are executed on the corpus for both C4.5 and TBL.</Paragraph>
    <Paragraph position="1"> We reimplemented the RNN algorithm and POS bigram statistical model to predict prosodic word boundary on the same corpus for comparison. Since our corpus is not large enough for HMM training and the CART method is also decision-tree based as C4.5, we didn't realize them in our experiments. The evaluation results are shown in Table 5.</Paragraph>
    <Paragraph position="2"> Both the C4.5 rules and the TBL rules outperform the RNN algorithm and POS bigram method because the overall accuracy</Paragraph>
    <Section position="1" start_page="22" end_page="22" type="sub_section">
      <SectionTitle>
rates Acc
</SectionTitle>
      <Paragraph position="0"> of the rule based methods are higher.</Paragraph>
      <Paragraph position="1"> TBL achieves comparable accuracy with C4.5 induction, which demonstrates that the design of transformation rule templates is successful.  in Table 5, we discover that prosodic word boundaries can be more accurately predicted than prosodic phrase ones. It can be explained as follows. Prosodic word is the smallest prosodic unit in the prosodic hierarchy, which has more relation with the word level features such as POS, word length etc. Prosodic phrase is a larger prosodic unit less related to word level features, thus it cannot be predicted accurately using these features.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>