<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1126">
  <Title>Mapping Collocational Properties into Machine Learning Features</Title>
  <Section position="9" start_page="231" end_page="232" type="concl">
    <SectionTitle>
8 Conclusions
</SectionTitle>
    <Paragraph position="0"> We performed extensive experimentation investigating the interactions among collocational property, feature organization, and machine learning algorithm. We found a highly significant interaction between collocational property and feature organization, which is extremely consistent across the machine learning algorithms experimented with. The results obtained with the per-class organization and the highly  definitive collocations (i.e., the SP collzcations) are significantly better than any experiment using either&amp;quot; the lower quality collocations or the over-range organization.</Paragraph>
    <Paragraph position="1"> The per-class organizations allow us to take advantage of the lower frequency, higher quality collocations; with the over-range organizations, the results are no better than with the lower quality ones. Our analysis shows, however, that merely using a per-class organization with high-quality collocations is not sufficient to realize the potential benefits: a larger number of collocations are needed for increased results.</Paragraph>
    <Paragraph position="2"> Very importantly, using the per-class organizations with the lower quality collocations proved costly--the results decreased by over 10%. Choices must be made in how collocations are selected and organized in any event. A main lesson from these experiments is that inappropriate organizations must be avoided for the particular type of property at hand.</Paragraph>
    <Paragraph position="3"> In continuing work, we are investigating interactions with additional experimental parameters. The goals of this paper were to investigate issues relevant for many NLP applications in a uniform framework, and to shed some light on interactions between collocational properties and how they are represented as features in machine learning algorithms.</Paragraph>
  </Section>
class="xml-element"></Paper>