<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0428">
  <Title>Named Entity Recognition with Character-Level Models</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> The primary argument of this paper is that character sub-strings are a valuable, and, we believe, underexploited source of model features. In an HMM with an admittedly very local sequence model, switching from a word model to a character model gave an error reduction of about 30%. In the final, much richer chained maxent setting, the reduction from the best model minus a4 -gram features to the reported best model was about 25% - smaller, but still substantial. This paper also again demonstrates how the ease of incorporating features into a discriminative maxent model allows for productive feature engineering.</Paragraph>
    <Paragraph position="1">  decision point: deciding the classification of Grace.</Paragraph>
  </Section>
class="xml-element"></Paper>