<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-1304">
  <Title>Coaxing Confidences from an Old Friend: Probabilistic Classifications from Transformation Rule Lists</Title>
  <Section position="7" start_page="87" end_page="87" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this paper we presented a novel way to convert transformation rule lists, a common paradigm in natural language processing, into a form that is equivalent in its classification behavior, but is capable of providing probability estimates. Using this approach, favorable properties of transformation rule lists that makes them popular for language processing are retained, while the many advantages of a probabilistic system axe gained.</Paragraph>
    <Paragraph position="1"> To demonstrate the efficacy of this approach, the resulting probabilities were tested in three ways: directly measuring the modeling accuracy on the test set via cross entropy, testing the goodness of the output probabilities in a active learning algorithm, and observing the rejection curves attained from these probability estimates.</Paragraph>
    <Paragraph position="2"> The experiments clearly demonstrate that the resulting probabilities perform at least as well as the ones generated by C4.5 decision trees, resulting in better performance in all cases. This proves that the resulting probabilistic classifier is as least as good as other state-of-the-art probabilistic models.</Paragraph>
    <Paragraph position="3"> The positive results obtained suggest that the probabilistic classifier obtained from transformation rule lists can be successfully used in machine learning algorithms that require soft-decision classifters, such as boosting or voting. Future research will include testing the behavior of the system under AdaBoost (Freund and Schapire, 1997). We also intend to investigate the effects that other decision tree growth and smoothing techniques may have on continued refinement of the converted rule list.</Paragraph>
  </Section>
class="xml-element"></Paper>