<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2007">
  <Title>Noun Phrase Recognition by System Combination</Title>
  <Section position="7" start_page="53" end_page="53" type="concl">
    <SectionTitle>
5 Concluding remarks
</SectionTitle>
    <Paragraph position="0"> We have put forward a method for recognizing noun phrases by combining the results of a memory-based classifier applied to different representations of the data. We have examined different combination techniques and each of them performed significantly better than the best individual classifier. We have chosen to work with majority voting because it does not require tuning data and thus enables the individual classifiers to use all the training data. This approach was applied to three standard data sets for base noun phrase recognition and arbitrary noun phrase recognition. For all data sets majority voting improved the best result for that data set known to US.</Paragraph>
    <Paragraph position="1"> Varying data representations is not the only way for generating different classifiers for combination purposes. We have also tried dividing the training data in partitions (bagging) and working with artificial training data generated by a crossover-like operator borrowed from genetic algorithm theory. With our memory-based classifier applied to this data, we have been unable to generate a combination which improved the performance of its best member. Another approach would be to use different classification algorithms and combine the results. We are working on this but we are still to overcome the practical problems which prevent us from obtaining acceptable results with the other learning algorithms.</Paragraph>
  </Section>
class="xml-element"></Paper>