<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0624">
  <Title>Sparse Bayesian Classification of Predicate Arguments</Title>
  <Section position="6" start_page="179" end_page="179" type="concl">
    <SectionTitle>
5 Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> We have provided an application of Relevance Vector Machines to a large-scale NLP task. The resulting classifiers are drastically smaller that those produced by the SV training methods. On the other hand, the classification accuracy is lower, probably because of the use of lexicalized features.</Paragraph>
    <Paragraph position="1"> The results on the Brown test set shows that the genre has a significant impact on the performance.</Paragraph>
    <Paragraph position="2"> An evaluation of the contribution of six parse tree  path features suggests that dependency tree paths are more useful for semantic role labeling than the traditional constituent tree path.</Paragraph>
    <Paragraph position="3"> In the future, we will investigate if it is possible to incorporate the g parameter into the probability model, thus eliminating the need for cross-validation completely. In addition, the training algorithm will need to be redesigned to scale up to larger training sets. The learning paradigm is still young and optimized methods (such as for SVM) have yet to appear. One possible direction is the greedy method described in (Tipping and Faul, 2003).</Paragraph>
  </Section>
class="xml-element"></Paper>