<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3211">
  <Title>Mixing Weak Learners in Semantic Parsing</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> The results documented in these experiments are very promising and mandate further research. The final classification accuracy of the Random Forest was 88.3%, just 0.6% behind the SVM results (Pradhan et al., 2003) and 4.6% higher than the next best results (Surdeanu et al., 2003) - results that were based on a number of additional features.</Paragraph>
    <Paragraph position="1"> We defined several modifications to the RF algorithm that increased accuracy. These improvements are important for any application with high dimensional categorical inputs, which includes many NLP tasks. We introduced new features which provided a 1.1% improvement in accuracy over the best results using features from the literature. We also introduced a technique to reduce the dimensionality of the feature space, resulting in a reduction to just 3% of the original feature space size. This could be an important enabler for handling larger datasets and improving the efficiency of feature and model selection.</Paragraph>
  </Section>
class="xml-element"></Paper>