<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2016">
  <Title>Rapid Parser Development: A Machine Learning Approach for Korean</Title>
  <Section position="8" start_page="122" end_page="122" type="concl">
    <SectionTitle>
7 Conclusions
</SectionTitle>
    <Paragraph position="0"> Comparisons with related work are unfortunately very problematic, because the corpora are different and are sometimes not even described in other work. In most cases Korean research groups also use other evaluation metrics, particularly dependency accuracy, which is often used in dependency structure approaches. Training on about 40,000 sentences (Collins, 1997) achieves a crossing brackets rate of 1.07, a better value than our 1.63 value for regular parsing or the 1.13 value assuming perfect segmentation/tagging, but even for similar text types, comparisons across languages are of course problematic.</Paragraph>
    <Paragraph position="1"> It is clear to us that with more training sentences, and with more features and background knowledge to better leverage the increased number of training sentences, accuracy rates can still be improved significantly. But we believe that the reduction of parser development time from two years or more down to three months is in many cases already very valuable, even if the accuracy has not 'maxed out' yet. And given the experience we have gained from this project, we hope this research to be only a first step to an even steeper development time reduction.</Paragraph>
    <Paragraph position="2"> A particularly promising research direction for this is to harness knowledge and training resources across languages.</Paragraph>
  </Section>
class="xml-element"></Paper>