<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-0811">
  <Title>Combining Heterogeneous Classifiers for Word-Sense Disambiguation</Title>
  <Section position="5" start_page="58" end_page="58" type="concl">
    <SectionTitle>
4 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this paper, we have explored ensemble sizes, combination methods, bounds for what can be expected from combinations, factors in the performance of individual classifiers, and methods of improving performance by effective tie-breaking. In accord with much recent work on classifier combination, e.g.</Paragraph>
    <Paragraph position="1"> (Breiman, 1996; Bauer and Kohavi, 1999), we have demonstrated that the combination of classifiers can lead to a substantial performance increase over the individual classifiers within the domain of WSD. In addition, we have shown that highly varying component systems augment each other well and that adding lower-scoring systems can still improve ensemble performance, at least to a certain point. A particular emphasis of our research has been how to make the combination robust to both the wide range of first-order classifier accuracies and to the sparsity of the available training data. Careful but greedy determination of rankings proved to be effective, capturing the highly word-dependent strengths of our classifiers. The resulting system's overall accuracy is very high, despite the medium level of accuracy of the component systems.</Paragraph>
  </Section>
class="xml-element"></Paper>