<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2403">
  <Title>A Semantic Kernel for Predicate Argument Classification</Title>
  <Section position="6" start_page="1" end_page="1" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this paper, we have experimented an original kernel based on semantic structures from PropBank corpus. The results have shown that: + the Semantic Kernel (SK) can be adopted to classify predicate arguments defined in PropBank; + SVMs using SK performs better than SVMs trained with the linear kernel of standard features; and + the higher gradient in the accuracy/training percentage plots shows that SK generalizes better than the linear kernel.</Paragraph>
    <Paragraph position="1"> Finally, SK suggests that some features, contained in the fragments of semantic structures, should be backported in a flat feature space. Conversely, the good performance of the linear kernel suggests that standard features, e.g. Head Word, Predicate Word should be emphasized in the definition of a convolution kernel for argument classification. Moreover, other selections of predicate/argument substructures (able to capture different linguistic relations) as well as kernel combinations (e.g. flat features with SK) could furthermore improve semantic shallow parsing.</Paragraph>
  </Section>
class="xml-element"></Paper>