<?xml version="1.0" standalone="yes"?>
<Paper uid="M91-1015">
  <Title>SRI INTERNATIONAL'S TACITUS SYSTEM : MUC-3 TEST RESULTS AND ANALYSI S</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
RESULTS
</SectionTitle>
    <Paragraph position="0"> This site report is intended as a companion piece to the System Summary appearing in this volume an d is best read in conjunction with it. In particular, it refers to the various modules of the system which ar e described in that paper .</Paragraph>
    <Paragraph position="1"> Here only the overall results will be summarized . A more detailed, component-by-component analysis of the results is contained in the System Summary .</Paragraph>
    <Paragraph position="2"> Our results for the TST2 corpus were as follows :  Our precision was the highest of any of the sites . Our recall was somewhere in the middle . It is as yet unclear whether high recall-high precision systems will evolve more rapidly from low recall-high precisio n systems or high recall-low precision systems .</Paragraph>
    <Paragraph position="3"> The significant drop in recall we experienced from Matched Templates Only to Matched/Missing is a n indication that we were failing on messages with a large number of template entries . Much of this is probabl y due to failures in handling lists of names, and could be improved by specialized handling of this phenomenon . We also ran our system, configured identically to the TST2 run, on the first 100 messages of the development set. The results were as follows:  Here recall was considerably better, as would be expected since the messages were used for development . While there are a number of parameter settings possible in our system, we decided upon optimal values , and those values were used. An explanation of the parameters and how we decided what was optimal is to o detailed and system-particular for this report . None of the decisions was made on the basis of total recal l and precision on a test set . All the decisions were made on a much more local basis .</Paragraph>
  </Section>
class="xml-element"></Paper>