<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-0811">
  <Title>Combining Heterogeneous Classifiers for Word-Sense Disambiguation</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The problem of supervised word sense disambiguation (WSD) has been approached using many different classification algorithms, including naive-Bayes, decision trees, decision lists, and memory-based learners. While it is unquestionable that certain algorithms are better suited to the WSD problem than others (for a comparison, see Mooney (1996)), it seems that, given similar input features, various algorithms exhibit roughly similar accuracies.1 This was supported by the SENSEVAL-2 results, where a This paper is based on work supported in part by the Na- null mentations of a single classifier type, such as smoothing or window size, impacted accuracy far more than the choice of classification algorithm.</Paragraph>
    <Paragraph position="1"> large fraction of systems had scores clustered in a fairly narrow region (Senseval-2, 2001).</Paragraph>
    <Paragraph position="2"> We began building our system with 23 supervised WSD systems, each submitted by a student taking the natural language processing course (CS224N) at Stanford University in Spring 2000. Students were free to implement whatever WSD method they chose.</Paragraph>
    <Paragraph position="3"> While most implemented variants of naive-Bayes, others implemented a range of other methods, including n-gram models, vector space models, and memory-based learners. Taken individually, the best of these systems would have turned in an accuracy of 61.2% in the SENSEVAL-2 English lexical sample task (which would have given it 6th place), while others would have produced middling to low performance. In this paper, we investigate how these classifiers behave in combination.</Paragraph>
    <Paragraph position="4"> In section 2, we discuss the first-order classifiers and describe our methods of combination. In section 3, we discuss performance, analyzing what benefit was found from combination, and when. We also discuss aspects of the component systems which substantially influenced overall performance.</Paragraph>
  </Section>
class="xml-element"></Paper>