<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0827">
  <Title>GAMBL, Genetic Algorithm Optimization of Memory-Based WSD</Title>
  <Section position="11" start_page="0" end_page="0" type="ackno">
    <SectionTitle>
PREDICTION TYPE AW LS
</SectionTitle>
    <Paragraph position="0"> predictions from keyword classifier 59 74 predictions from old data classifier / 65 words in local context 59 58 POS-tags of local context 55 65 chunk+relation tags of local context 67 72 formation seems to contribute to the disambiguation process: Table 4 list for each type of feature the percentage of times it was selected by the GA. Though Table 4 is an not exhaustive comparison of the different types of features, we nevertheless see that the GA selects syntactic and grammatical information more often than plain words or POS-tags.</Paragraph>
    <Paragraph position="1"> Finally, Table 4 also suggests that our cascaded approach to combine two different information sources is quite successful: the predictions from the previous classifier(s) are very often selected, especially in the LS task, where the prediction from the keyword classifier is most often selected.</Paragraph>
  </Section>
class="xml-element"></Paper>