<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1638">
  <Title>Better Informed Training of Latent Syntactic Features</Title>
  <Section position="9" start_page="324" end_page="324" type="concl">
    <SectionTitle>
7 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have discussed &amp;quot;informed&amp;quot; techniques for inducing latent syntactic features. Our INHERIT model tries to constrain the way in which features are passed through the tree. The motivation for this approach is twofold: First, we wanted to capture the linguistic insight that features follow certain patterns in propagating through the tree. Second, we wanted to make it statistically feasible and computationally tractable to increase L to higher values than in the PCFG-LA model. The hope was that the learning process could then make finer distinctions and learn more fine-grained information.</Paragraph>
    <Paragraph position="1"> However, it turned out that the higher values of L did not compensate for the perhaps overly con10This affects EM training only by requiring a convex optimization at the M step (Riezler, 1998).</Paragraph>
    <Paragraph position="2"> strained model. The results on English parsing rather suggest that it is the similarity in degrees of freedom (e.g., INHERIT with L=3x3x3 and PCFG-LA with L=2x2) that produces comparable results.</Paragraph>
    <Paragraph position="3"> Substantial gains were achieved by using markovization and splitting only selected nonterminals. With these techniques we reach a parsing accuracy similar to Matsuzaki et al. (2005), but with an order of magnitude less parameters, resulting in more efficient parsing. We hope to get more wins in future by using more sophisticated annealing techniques and log-linear modeling techniques.</Paragraph>
  </Section>
class="xml-element"></Paper>