<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0633">
  <Title>Semantic Role Labeling Using Lexical Statistical Information</Title>
  <Section position="6" start_page="215" end_page="215" type="concl">
    <SectionTitle>
4 Conclusion
</SectionTitle>
    <Paragraph position="0"> Semantic role labeling is a difficult task, and accordingly, how to achieve an accurate and robust performance is still an open question. In our work we used a limited set of syntactic tree based distance and size metrics coupled with raw lexical statistics, and showed that such 'lazy learning' configuration can still achieve a reasonable performance.</Paragraph>
    <Paragraph position="1"> We concentrated on reducing the complexity given by the number and dimensionality of the instances to be classified during learning. This is the core motivation behind performing tree pruning and statistical feature encoding. This also helped us to avoid the use of sparse features such as the explicit path in the parse tree between the candidate constituent and the predicate, and the predicate's sub-categorization rule (cf. e.g. Pradhan et al. (2004)). Future work will concentrate on benchmarking this approach within alternative architectures (i.e.</Paragraph>
    <Paragraph position="2"> two-phase with filtering) and different learning schemes (i.e. vector-based methods such as Support Vector Machines and Artificial Neural Networks).</Paragraph>
    <Paragraph position="3"> Acknowledgements: This work has been funded by the Klaus Tschira Foundation, Heidelberg, Germany. The first author has been supported by a KTF grant (09.003.2004).</Paragraph>
  </Section>
class="xml-element"></Paper>