<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1071">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics A Progressive Feature Selection Algorithm for Ultra Large Feature Spaces</Title>
  <Section position="7" start_page="566" end_page="566" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> This paper presents our progressive feature selection algorithm that greatly extends the feature space for conditional maximum entropy modeling. The new algorithm is able to select features from feature space in the order of tens of millions in practice, i.e., 8 times the maximal size previous algorithms are able to process, and unlimited space size in theory. Experiments on edit region identification task have shown that the increased feature space leads to 17.66% relative improvement (or 3.85% absolute) over the best result reported by Kahn et al. (2005), and 10.65% relative improvement (or 2.14% absolute) over the new baseline SGC algorithm with all the variables from Zhang and Weng (2005).</Paragraph>
    <Paragraph position="1"> We also show that symbolic prosody labels together with confidence scores are useful in edit region identification task.</Paragraph>
    <Paragraph position="2"> In addition, the improvements in the edit identification lead to a relative 20% error reduction in parsing disfluent sentences when gold edits are used as the upper bound.</Paragraph>
  </Section>
class="xml-element"></Paper>