<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-1011">
  <Title>Online Learning of Approximate Dependency Parsing Algorithms</Title>
  <Section position="7" start_page="86" end_page="87" type="concl">
    <SectionTitle>
6 Discussion
</SectionTitle>
    <Paragraph position="0"> We described approximate dependency parsing algorithms that support higher-order features and multiple parents. We showed that these approximations can be combined with online learning to achieve fast parsing with competitive parsing accuracy. These results show that the gain from allowing richer representations outweighs the loss from approximate parsing and further shows the robustness of online learning algorithms with approximate inference.</Paragraph>
    <Paragraph position="1"> The approximations we have presented are very simple. Theystart withareasonably good baseline and make small transformations until the score of the structure converges. These approximations work because freer-word order languages we studied are still primarily projective, making the approximate starting point close to the goal parse.</Paragraph>
    <Paragraph position="2"> However, we would like to investigate the benefits for parsing of more principled approaches to approximate learning and inference techniques such as the learning as search optimization framework of (Daum'e and Marcu, 2005). This framework will possibly allow us to include effectively more global features over the dependency structure than  those in our current second-order model.</Paragraph>
  </Section>
class="xml-element"></Paper>