<?xml version="1.0" standalone="yes"?>
<Paper uid="M91-1013">
  <Title>THE NEWSPAPER STATED THAT THE ALLEGED RIFT BETWEEN THE MILITARY OFFICERS</Title>
  <Section position="2" start_page="0" end_page="95" type="intro">
    <SectionTitle>
SET FILLS ONLY 40 62 18 1
</SectionTitle>
    <Paragraph position="0"> Although the template specifications call for us not to produce templates for non-specific event descriptions, ou r base run did not attempt to filter out such descriptions, in part because the criteria are hard to specify, but mainl y because at our low recall there was a fair chance of missing date or perpetrator information and thus eliminatin g templates incorrectly. However, we made two optional runs to attempt to eliminate these templates . For both runs , we eliminated all military and terrorist targets (we had found from our training runs that these targets were frequently being incorrectly combined with civilian targets and therefore emitted as parts of templates) . For the firs t run, we then eliminated all templates with neither perpetrator nor targets (typically arising from nominalization s such as &amp;quot;the recent attack&amp;quot;, when there was no preceding attack being referred to) ; the results were:  Since this is a process of template elimination, the changes are most noticeable on the &amp;quot;all templates&amp;quot; line, wher e the precision rises from 36% to 45% . The effect can also be seen in template overgeneration (not shown in the above tables), which falls from 48% on the max-tradeoff run to 30% on the final run .</Paragraph>
  </Section>
class="xml-element"></Paper>