<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0733">
  <Title>Text Chunking by System Combination</Title>
  <Section position="5" start_page="152" end_page="152" type="metho">
    <SectionTitle>
4 Concluding Remarks
</SectionTitle>
    <Paragraph position="0"> We have evaluated three methods for recognizing non-recursive non-overlapping text chunks of arbitrary syntactical categories. In each method a memory-based learner was trained to recognize chunks represented in five different ways. We have examined nine different methods for combining the five results. A 10-fold cross-validation experiment on the training data of the CoNLL-2000 shared task revealed that (1) the combined results were better than the individual results, (2) the combination methods perform equally well and (3) the best performances of the three processing methods were similar. We have selected the double-pass method with majority voting for processing the CoNLL-2000 shared task data. This method outperformed an earlier text chunking study for most chunk types, despite the fact that it used about 80% less training data.</Paragraph>
  </Section>
class="xml-element"></Paper>