<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1106">
  <Title>References</Title>
  <Section position="4" start_page="52" end_page="54" type="metho">
    <SectionTitle>
4 Examples
</SectionTitle>
    <Paragraph position="0"> In this section, we show several examples of case frames that were generated automatically by our system. Figure 5 shows a simple case frame triggered by active forms of the verb &amp;quot;ambushed&amp;quot;. The subject is extracted as a perpetrator and has a selectional restriction of TERRORIST. The direct object is extracted as a target and has a selectional restriction of VEHICLE. Note that the case frame does not contain a victim slot, even though it is theoretically possible to ambush people. During training, the &amp;quot;ambushed &lt;direct-obj&gt;&amp;quot; pattern extracted 13 people, 11 of whom were recognized as MILITARYPEOPLE. Since our domain roles only list civilians and government officials as legitimate terrorism victims 3, a victim slot was not created. This example shows how the case frames are tailored for the domain empirically.</Paragraph>
    <Paragraph position="1">  Figure 6 shows a case frame triggered by active forms of &amp;quot;blew_up&amp;quot; .4 This case frame extracts information from an entire sentence into a single structure. The subject (perpetrator), direct object (target), and a prepositional phrase (in location) will all be extracted together.</Paragraph>
    <Paragraph position="2">  The case frame in Figure 7 illustrates how a semantic category can show up in multiple places.</Paragraph>
    <Paragraph position="3"> This case frame will handle phrases like &amp;quot;the guerrillas detonated a bomb&amp;quot;, as well as &amp;quot;the bomb detonated&amp;quot;. Both constructions are very common in the training corpus so the system added slots for both possibilities. It would be easy for a human to overlook some of these variations when creating case frames by hand.</Paragraph>
    <Paragraph position="4"> The case frame in Figure 8 is activated by the noun &amp;quot;attack&amp;quot; and includes slots for a variety of prepositional phrases. The same preposition can recognize different types of information (e.g., &amp;quot;on&amp;quot; can recognize targets, victims, locations, and dates). And the same role can be filled by different prepositions  A disadvantage of this automated method is that inappropriate slots sometimes end up in the case frames. For example, Figure 9 shows a case frame that is activated by passive forms of the verb &amp;quot;killed&amp;quot;. Some of the slots are correct: the sub-ject is assigned to the victim slot and objects of the preposition &amp;quot;by&amp;quot; are assigned to the perpetrator and instrument slots. However, the remaining slots do not make sense. The location slot is the result ofpolysemy; many person names are also location names, such as &amp;quot;Flores'. The date slot was produced by incorrect parses of date expressions. The perpetrator (subject) and victim (pp (by)) slots were caused by incorrect role assignments. The list of domain roles assumes that terrorists are always perpetrators and civilians are always victims, but of course this is not true. Terrorists can be killed and civilians can be killers.</Paragraph>
    <Paragraph position="5">  The previous example illustrates some of the problems that can occur when generating case frames automatically. Currently, we are assuming that each semantic category will be uniquely associated with  a conceptual role, which may be an unrealistic assumption for some domains. One avenue for future work is to develop more sophisticated methods for mapping semantic preferences to conceptual roles.</Paragraph>
    <Paragraph position="6"> One could also have a human review the case frames and manually remove inappropriate slots. For now, we chose to avoid additional human interaction and used the case frames exactly as they were generated.</Paragraph>
  </Section>
class="xml-element"></Paper>