<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2048">
  <Title>Exploring the Potential of Intractable Parsers</Title>
  <Section position="9" start_page="375" end_page="375" type="concl">
    <SectionTitle>
7 Discussion
</SectionTitle>
    <Paragraph position="0"> This project began with a question: can we develop a history-based parsing framework that is simple, general, and effective? We sought to provide a versatile probabilistic framework that would be free from the constraints that dynamic programming places on PCFG-based approaches.</Paragraph>
    <Paragraph position="1"> The work presented in this paper gives favorable evidence that more exible (and worst-case intractable) probabilistic approaches can indeed perform well in practice, both in terms of running time and parsing quality.</Paragraph>
    <Paragraph position="2"> We can extend this research in multiple directions. First, the set of features we selected were chosen with simplicity in mind, to see how well a simple and unadorned set of features would work, given our probabilistic model. A next step would be a more carefully considered feature set. For instance, although lexical information was used, it was employed in only a most basic sense. There was no attempt to use head information, which has been so successful in PCFG parsing methods.</Paragraph>
    <Paragraph position="3"> Another parameter to experiment with is the model order, i.e. the order in which the model variables are assigned. In this work, we explored only one speci c order (the left-to-right, leaves-to-head assignment) but in principle there are many other feasible orders. For instance, one could try a top-down approach, or a bottom-up approach in which internal nodes are assigned immediately after all of their descendants' values have been determined.</Paragraph>
    <Paragraph position="4"> Throughout this paper, we strove to present the model in a very general manner. There is no reason why this framework cannot be tried in other application areas that rely on dynamic programming techniques to perform hierarchical labeling, such as phrase-based machine translation. Applying this framework to such application areas, as well as developing a general-purpose parser based on HLPs, are the subject of our continuing work.</Paragraph>
  </Section>
class="xml-element"></Paper>