<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1515">
  <Title>Vancouver, October 2005. c(c)2005 Association for Computational Linguistics Constituent Parsing by Classification</Title>
  <Section position="6" start_page="149" end_page="149" type="concl">
    <SectionTitle>
4 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this work, we presented a near state-of-the-art approach to constituency parsing which overcomes some of the limitations of other discriminative parsers. Like Yamada and Matsumoto (2003) and Sagae and Lavie (2005), our parser is driven by classifiers. Even though these classifiers themselves never do any parsing during training, they can be combined into an effective parser. We also presented a beam search method under the objective function of maximizing the minimum confidence.</Paragraph>
    <Paragraph position="1"> To ensure efficiency, some discriminative parsers place stringent requirements on which types of features are permitted. Our approach requires no such restrictions and our scoring function can, in principle, use arbitrary information from the history to evaluate constituent inferences. Even though our features may be of too fine granularity to discriminate through linear combination, discriminatively trained decisions trees determine useful feature combinations automatically, so adding new features requires minimal human effort.</Paragraph>
    <Paragraph position="2"> Training discriminative parsers is notoriously slow, especially if it requires generating examples by repeatedly parsing the treebank (Collins &amp; Roark, 2004; Taskar et al., 2004). Although training time is still a concern in our setup, the situation is ameliorated by generating training examples in advance and inducing one-vs-all classifiers in parallel, a technique similar in spirit to the POS-tag parallelization in Yamada and Matsumoto (2003) and Sagae and Lavie (2005).</Paragraph>
    <Paragraph position="3"> This parser serves as a proof-of-concept, in that we have not fully exploited the possibilities of engineering intricate features or trying more complex search methods. Its flexibility offers many opportunities for improvement, which we leave to future work.</Paragraph>
  </Section>
class="xml-element"></Paper>