<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0827">
  <Title>GAMBL, Genetic Algorithm Optimization of Memory-Based WSD</Title>
  <Section position="4" start_page="0" end_page="0" type="intro">
    <SectionTitle>
YESNO
</SectionTitle>
    <Paragraph position="0"> prediction of the first classifier keywords above threshold parameter optimizationand feature selectionwith genetic algorithm heuristic optimizationparameter than onesense?more allows a choice between different statistical and information-theoretic feature and value weighting methods, different neighborhood size and weighting parameters, etc., that should be optimized for each word expert independently. See (Daelemans et al., 2003b) for more information. It has been claimed, e.g. in (Daelemans et al., 1999), that lazy learning has the right bias for learning natural language processing tasks as it makes possible learning from atypical and low-frequency events that are usually discarded by eager learning methods.</Paragraph>
    <Paragraph position="1"> Architecture. Previous work on memory-based WSD includes work from Ng and Lee (1996), Veenstra et al. (2000), Hoste et al. (2002) and Mihalcea (2002). The current design of our WSD system is largely based on Hoste et al. (2002).</Paragraph>
    <Paragraph position="2"> Figure 1 gives an overview of the design of our WSD system: the training text is first linguistically analyzed. For each word-lemma-POS-tag combination, we check if it (i) is in our sense lexicon, (ii) has more than one sense and (iii) has a frequency in the training text above a certain threshold. For all combinations matching these three conditions, we train a word expert module. To all combinations with only one sense, or with more senses and a frequency below the threshold, we assign the default sense, which is respectively the only or most frequent sense in WordNet.</Paragraph>
    <Paragraph position="3"> The word expert module consists of two cascaded memory-based classifiers: the sense predicted by</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Association for Computational Linguistics
</SectionTitle>
      <Paragraph position="0"> for the Semantic Analysis of Text, Barcelona, Spain, July 2004 SENSEVAL-3: Third International Workshop on the Evaluation of Systems the first classifier is used as a feature in the second classifier. The first classifier is trained on keywords selected according to a statistical criterion, and the second one is trained on the prediction of the first and on the local context of the ambiguous word-lemma-POS-tag combination.</Paragraph>
      <Paragraph position="1"> In the remainder of this paper, we will describe the feature construction process from the available information sources (Section 2), the learning and optimization approach (Section 3), and the results (Section 4) and their interpretation.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>