<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2034">
  <Title>To Parse or Not to Parse: Relation-Driven Text Skimming</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 The Text Extraction Task and the
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Skimrning Problem
</SectionTitle>
      <Paragraph position="0"> Skilrmling is of little use if the processing task demands a complete analysis. The skimming method presented here is aimed at the task of information extraction from text, where the program looks for relal;ively superficial facts that appear in texts with conslrained content. Since the program is looking only for certain key information, it should spend most of its time analyzing sections thal; contain that information. The more extraneous information there is in a text, the more skimming improves performance over full parsing.</Paragraph>
      <Paragraph position="1"> The skirnmirig problem assunles that a text prc~ c~-;sing program works from set of predefined conceptual roles that represent some of the information from a t.ext. These conceptual roles forln a &amp;quot;temo plate&amp;quot; that the program fills in while scanning a text, The words &amp;quot;relevant&amp;quot;, &amp;quot;important&amp;quot;, and &amp;quot;extraneo~s&amp;quot; used throughout this paper describe the rela, tionship of portions of text to this template-filling task.</Paragraph>
      <Paragraph position="2"> For example, the following is a l.ypical i)ow Jones m'w story, along with the template structure representing the ilaformat.ion extracted by SCIS()R: I\]~put Text: th.uaswick Corp Up; Aclive amid Continued</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Takeover Talk
</SectionTitle>
      <Paragraph position="0"> New ~)rk -D J- Traders and market sources say shares of Brunswick Corp., the world's largest Inanufaeturer of recreatioaal boats, are trading actively for the second consecutive week amid growing speculal, ion that someone is accumulating a post{ion in the company. Brunswick is up 5-8 a.t 20 on NYSE-composite voluime (sic) of 1,I 60,400 shares, compared with an average daily vohtme of 435,20(I shares. The stock rose 11-2 yesterday on more than 1.1 million shares. In the past two we(:ks, l~runswick shares have traded abow~&amp;quot; average daily vohiinc on all bug one day, inching up from a low of 16 3-4 on March 14. The, most often rttmored suitor for Brunswick is Miniieapolis investor and boat company owner Irwin Ja.eobs, whose name surfaced about a. year ago when Brunswick shares made a similar move eli nnfounded speculation. One trader tells Dew Jones Professional hivestor Report a New Jersey-based &amp;quot;tape reading&amp;quot; service today named the stock as the target of a S30-a-share bid fl'om Jacobs. The service can't be reached for confirmation. A secretary to ,Iacobs said late yesterda.y he'll be out of t, is office until Friday. Jacobs has become a popular rumore.d shark since walking away with cash from his failed bid earlier this month for Shaklee Corp.. lie's also one el' several rumored suitors for NWA In(:.</Paragraph>
      <Paragraph position="1">  \[rite system conht obtain the same minimal information if the text read as follows: Brunswick is up 5-8 at 20....rumored._.the target of a $30-a-share bid from Jacobs.</Paragraph>
      <Paragraph position="2"> The objective of skimrning is to read the text as if it were closer to this condensed form. The problem for text skimming is thus (1) to identify sections of text that will contribute to information extraction, and (2) to limit processing in those sections. The introduction outlined three different types of pertbrmance improvements that come from skimming. 'l'he next section describes the relation-driven method and how it achieves these improvements.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>