<?xml version="1.0" standalone="yes"?>
<Paper uid="A88-1007">
  <Title>IMPROVED PORTABILITY AND PARSING THROUGH INTERACTIVE ACQUISITION OF SEMANTIC INFORMATION t</Title>
  <Section position="10" start_page="54" end_page="54" type="concl">
    <SectionTitle>
SS
</SectionTitle>
    <Paragraph position="0"> We are also examining what kind of knowledge the user must draw upon in order to answer the system's questions. Users' answers are usually based on a combination of commonsense knowledge (e.g., losses cannot install things) and domain-specific information. In certain cases, however, the user can be called upon to make fine linguistic distinctions. For example, in the sentence Sac disengaged immediately after alarm, does the adverb immediately modify the verb dimengaged, or the prepositional phrase after alarm? Most users, and even trained linguists familiar with the domain, find it difficult to provide definitive answers to such questions, because there is often no definitively correct answer. In this case, the adverbial attachment would seem to be genuinely ambiguous. It would be helpful to recognise patterns which a user cannot be reasonably expected to pass judgment on, and not generate queries about these, perhaps allowing them to succeed by default.</Paragraph>
    <Paragraph position="1"> 8.2. Measuring the System's Learning As more sentences are parsed and more patterns are classified, we can expect the system to grow &amp;quot;smarter&amp;quot; in the sense that it will ask the user increasingly fewer questions. Eventually, the system should reach a state of reasonably complete domain knowledge, at which time few unknown patterns would be encountered, and the user would almost never be queried. We do not know how many sentences SPQR would have to examine before attaining such a plateau, but an estimate would be in the range of 500 to 1000 \[Grishman1986\]. We plan to measure the decrease in the frequency of queries to the user as a function of the number of sentences parsed and the number of patterns collected in order to evaluate the system's learning. This will enable us to determine the feasibility of using this technique to bootstrap into a new domain.</Paragraph>
    <Section position="1" start_page="54" end_page="54" type="sub_section">
      <SectionTitle>
8.3. Preference-Based Parsing
</SectionTitle>
      <Paragraph position="0"> A long-term goal is to implement a parsing algorithm based on preference rather than on the current success/failure paradigm. This would allow the system to use statistical information on the frequency of observed patterns as one factor in weighting. Frequently occurring patterns would be assigned greater weight than unknown patterns, and bad patterns would detract from the overall weighting. This would allow the system to make intelligent &amp;quot;guesses&amp;quot; about parsing without constantly querying the user.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>