<?xml version="1.0" standalone="yes"?>
<Paper uid="J94-2001">
  <Title>Tagging English Text with a Probabilistic Model</Title>
  <Section position="8" start_page="165" end_page="166" type="concl">
    <SectionTitle>
8. Conclusion
</SectionTitle>
    <Paragraph position="0"> The results presented in this paper show that estimating the parameters of the model by counting relative frequencies over a very large amount of hand-tagged text lead to the best tagging accuracy.</Paragraph>
    <Paragraph position="1"> Maximum Likelihood training is guaranteed to improve perplexity, but will not necessarily improve tagging accuracy. In our experiments, ML training degrades the performance unless the initial model is already very bad.</Paragraph>
    <Paragraph position="2"> The preceding results suggest that the optimal strategy to build the best possible model for tagging is the following: * get as much tagged (by hand) text as you can afford  Bernard Merialdo Tagging English Text with a Probabilistic Model compute the relative frequencies from this data to build an initial model  get as much untagged text as you can afford starting from M0, perform the Forward-Backward iterations. At each iteration, evaluate the tagging quality of the new model Mi on some held-out tagged text. Stop either when you have reached a preset number of iterations or the model Mi performs worse than Mi-1, whichever occurs first.</Paragraph>
  </Section>
class="xml-element"></Paper>