<?xml version="1.0" standalone="yes"?>
<Paper uid="N04-1018">
  <Title>Detecting Structural Metadata with Decision Trees and Transformation-Based Learning</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have demonstrated a two-tiered system that detects various types of dis uencies in spontaneous speech. In the rst tier, a decision tree model utilizes multiple knowledge sources to predict interword boundary events.</Paragraph>
    <Paragraph position="1"> Then the system employs a transformation-based learning algorithm to identify the extent and type of dis uencies. Experimental results show that the large variance and noise inherent in prosodic features makes them much less effective than lexical features for reference data; however, in the presence of word recognition errors prevalent in automatic transcripts of spontaneous speech, prosodic features have more value. Performance differences for the various score combination methods were small, but combining decision tree and HE-LM scores with a weight optimized on dev data is slightly better for edit dis uencies. Transformation-based learning is an effective way to tag llers and edit regions after boundary events are tagged, but the best performance is obtained when training with automatically predicted SU and IP boundary events.</Paragraph>
    <Paragraph position="2"> As this is a new task, error rates are relatively high (though signi cantly better than chance), but this approach achieved competitive results on the Fall 2003 NIST Rich Transcription Evaluation, and there are many directions for future improvements.</Paragraph>
  </Section>
class="xml-element"></Paper>