<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3222">
  <Title>The Leaf Projection Path View of Parse Trees: Exploring String Kernels for HPSG Parse Selection</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> We proposed a new representation of parse trees that allows us to connect more tightly tree structures to the words of the sentence. Additionally this representation allows for the natural extension of string kernels to kernels on trees. The major source of accuracy improvement for our models was this representation, as even with bi-gram features, the performance was higher than previously achieved. We were able to improve on these results by using more sophisticated Subsequence kernels and by our Repetition kernel which captures some salient properties of word projection paths.</Paragraph>
    <Paragraph position="1"> In future work, we aim to explore the definition of new string kernels that are more suitable for this particular application and apply these ideas to Penn Treebank parse trees. We also plan to explore annotation with more features from HPSG signs.</Paragraph>
  </Section>
class="xml-element"></Paper>