<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-1051">
  <Title>PORTING PUNDIT TO THE RESOURCE MANAGEMENT DOMAIN</Title>
  <Section position="9" start_page="280" end_page="281" type="concl">
    <SectionTitle>
FUTURE DIRECTIONS
</SectionTitle>
    <Paragraph position="0"> There are several directions that we plan to pursue. The first is to complete our experiments using selectional constraints to prune parses. A second general area that we will focus on over the next few months is the notion of how to train the system, that is, using the training set to customize the system to the given domain automatically. In particular, we plan to experiment with a &amp;quot;minimal&amp;quot; lexicon, to determine if we can improve our results by pruning out unneeded syntactic class information (e.g., just having alert entered as a noun for this domain). If pruning the lexicon improves our performance significantly, then we will experiment with various ways to use the  training data to tune the system (in some automatic way) to &amp;quot;specialize&amp;quot; the lexicon to the particular application. Similarly, we plan to investigate techniques for using the training corpus to tune the parser to a new domain.</Paragraph>
    <Paragraph position="1"> Our ultimate objective is to couple PUNDIT to a speech recognition system. To achieve this, we must focus not only on obtaining the correct parse, but on ruling out incorrect parses. So far, most development work has focused on extending coverage, and not on tightening the grammar to prevent overgeneration. Clearly, it is critical to address this problem if we plan to use a broad-coverage natural language system for spoken language understanding. This will also include developing metrics to measure overgeneration.</Paragraph>
    <Paragraph position="2"> Finally, we expect to add the rules to support the in-depth semantic coverage that we have produced for our message domains. Overall, we are optimistic that by adding semantic constraints, plus extending the syntactic coverage in some quite limited ways, we will be able to exceed a 90~ correct analysis rate on the test data, which brings the system within the bounds of a realistically useful system.</Paragraph>
  </Section>
class="xml-element"></Paper>