<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-2013">
  <Title>Named Entity Extraction with Conditional Markov Models and Classifiers</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> We presented a simple, knowledge-poor named entity recognizer using standard components. Our decomposition into extraction and classification phases was motivated by the common syntactic regularities and the ambiguous status of some named entities. We have shown that the conditional next-tag model used for extraction is not unprincipled (a criticism brought forward by McCallum et al.</Paragraph>
    <Paragraph position="1"> (2000) against next-tag classifiers that do not output probabilities), but arises naturally from a conditional sequence model and plausible independence assumptions. This extraction model achieves fairly high accuracy (and just as observed by Punyakanok and Roth (2001) it outperforms a joint generative Markov model). A separate classification step makes it easy to use sentence-level features and large amounts of contexts. Such features would be difficult to integrated into standard models, the major exception being conditional random fields (Lafferty et al., 2001), compared to which the approach proposed here is much simpler.</Paragraph>
  </Section>
class="xml-element"></Paper>