<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1819">
  <Title>SS</Title>
  <Section position="8" start_page="22" end_page="22" type="concl">
    <SectionTitle>
6 Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> In this paper, we describe an effective approach to generate rules for Chinese prosodic phrase prediction. The main idea is to extract appropriate features from the linguistic information and to apply rule-learning algorithms to automatically induce rules for predicting prosodic boundary labels. C4.5 and TBL algorithms are experimented in our research. In order to find the most effective features, a series of feature selection experiments is conducted. The acquired rules achieve a best accuracy rate above 90% on test data and outperform the RNN and bigram based methods, which justifies rule-learning as an effective alternative to prosodic phrase prediction.</Paragraph>
    <Paragraph position="1"> But the problem of prosodic phrase prediction is far from solved. The best accuracy rate got by machine is still much lower than that by human. In our future work, the study on this problem will go more deep and wide. Other machine learning methods will be experimented and compared with C4.5 and TBL. Features from deep syntactic, semantic or discourse information will be paid more attention to (Julia and Owen, 2001). And the speech corpus will be enlarged to cover more types of text and speaking styles.</Paragraph>
  </Section>
class="xml-element"></Paper>