<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-1023">
  <Title>Weakly Supervised Natural Language Learning Without Redundant Views</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> We have investigated single-view algorithms (selftraining and EM) as an alternative to multi-view algorithms (co-training) for weakly supervised learning for problems that do not appear to have a natural feature split.</Paragraph>
    <Paragraph position="1"> Experimental results on two coreference data sets indicate that self-training outperforms co-training under various parameter settings and is comparatively less sensitive to parameter changes. While weakly supervised EM is not able to outperform co-training, we introduce a variation of EM, FS-EM, for boosting the performance of EM via feature selection. Like self-training, FS-EM easily outperforms co-training.</Paragraph>
    <Paragraph position="2"> Co-training algorithms such as CoBoost (Collins and Singer, 1999) and Greedy Agreement (Abney, 2002) that explicitly trade classifier agreement on unlabeled data against error on labeled data may be more robust to the underlying assumptions of co-training and can conceivably perform better than the Blum and Mitchell algorithm for problems without a natural feature split.9 Other less studied single-view weakly supervised algorithms in the NLP community such as co-training with different learning algorithms (Goldman and Zhou, 2000) and graph mincuts (Blum and Chawla, 2001) can be similarly applied to these problems to further test our original hypothesis. We plan to explore these possibilities in future research.</Paragraph>
  </Section>
class="xml-element"></Paper>