<?xml version="1.0" standalone="yes"?>
<Paper uid="J86-3002">
  <Title>DISCOVERY PROCEDURES FOR SUBLANGUAGE SELECTIONAL PATTERNS: INITIAL EXPERIMENTS</Title>
  <Section position="10" start_page="27" end_page="27" type="concl">
    <SectionTitle>
7 CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> Overall, the experiments we have conducted using our discovery procedure are encouraging but not conclusive.</Paragraph>
    <Paragraph position="1"> The selectional patterns gathered from a limited text sample - when coupled with a procedure for restriction relaxation - do about as well as manually prepared selectional patterns. Furthermore, the growth curves for the selectional patterns suggest that a corpus several times larger would yield a more complete set of patterns and thus better performance in parsing.</Paragraph>
    <Paragraph position="2"> We have learned that such a procedure requires substantial human interaction and we intend, before advancing to a larger corpus, to restructure the system to facilitate this interaction. The present system is basically organized for batch processing; interaction takes place by editing intermediate files. Our next step will be to move to an interactive environment that supports the following capabilities: * isolating parse ambiguities and homographs and prompting the user to choose the appropriate reading/meaning; * displaying new selectional patterns the first time they are encountered; * supporting simultaneous inspection and manipulation of text, parse tree, and selectional patterns.</Paragraph>
    <Paragraph position="3"> In all of this interaction, however, the user is still acting only as a monitor of the patterns generated. We are still faced with the difficult issue of how to bootstrap the system into a new domain. In the absence of selectional patterns, choosing the correct parse can become a tedious and time-consuming procedure, requiring extensive interaction with both a domain expert and a linguist. It is clearly not a realistic method of building up a set of patterns sufficient for semi-automated processing ~3f the type described above.</Paragraph>
    <Paragraph position="4"> Combining the text-based approach with elicitation procedures offers a more practical method of acquiring an initial set of domain knowledge. An expert could provide some initial word classes and a partial set of relationships, from which to generate selectional patterns. A sample of text would then provide additional examples, with the expert available to elaborate on further patterns. For example, a system being developed at BBN 4 uses a hierarchy of sublanguage classes; given a selectional pattern, it asks the user to generalize it by replacing classes with superclasses where possible. This initial period of intensive interaction with an expert would provide a sufficient pattern base so that the text-driven tools would become effective in filling in the knowledge base.</Paragraph>
    <Paragraph position="5"> Such an approach would offer the assurance of good coverage provided by a text-based system while requiring a smaller text sample than a purely text-based procedure.</Paragraph>
  </Section>
class="xml-element"></Paper>