<?xml version="1.0" standalone="yes"?>
<Paper uid="M92-1019">
  <Title>SRI INTERNATIONAL FASTUS SYSTEM MUC-4 TEST RESULTS AND ANALYSIS</Title>
  <Section position="7" start_page="144" end_page="146" type="concl">
    <SectionTitle>
CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> FASTUS was more successful than we ever dreamed when the idea was originally conceived . In retrospect ,  we attribute its success to the fact that its processing is extremely well suited to the demands of the task . The system's phase-3 works successfully because the input from phase-2 is already reliably processed . Phase two does only the linguistic processing that can be done reliably and fast, ignoring all the problems of makin g attachment decisions, and the ambiguity introduced by coordination and appositives . This input is adequate for phase-3 because the domain pragmatics are sufficiently constrained that given this initial chunking, th e relevant information can be reliably detected and extracted .</Paragraph>
    <Paragraph position="1"> One source of frustration with the development of this system is that we never had the opportunity t o produce a decent developer's interface . We believe that phase-2 is almost completely domain independent , with all the domain specific knowledge embedded in the phase-3 automata . We feel that with some carefu l thought devoted to such an interface, we could produce a general text processing system that could b e brought up to our current level of performance on a MUC-like or TIPSTER-like task in even less than th e three and a half weeks of effort that we required .</Paragraph>
    <Paragraph position="2"> Another discovery of this experience is that a MUG-like task is much easier than anyone ever thought .</Paragraph>
    <Paragraph position="3"> Although the full linguistic complexity of the MUC texts is very high, with long sentences and interestin g discourse structure problems, the relative simplicity of the information-extraction task allows much of thi s linguistic complexity to by bypassed - indeed much more than we had originally believed was possible . The key to the whole problem, as we see it from our FASTUS experience, is to do exactly the right amount of syntax, so that pragmatics can take over its share of the load . For the MUC task, we think FASTUS display s exactly the right mixture .</Paragraph>
    <Paragraph position="4"> Finally, we point out that while FASTUS is an elegant engineering achievement,, the whole host o f linguistic problems that were bypassed are still out there, and will have to be addressed eventually fo r more complex tasks . and to achieve ever higher performance on simpler tasks . The nature of competitive evaluations is that they force everyone to deal with the easiest problems first . However, the hard problems cannot be ignored forever, and scientific progress requires that they be addressed .</Paragraph>
  </Section>
class="xml-element"></Paper>