<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-1067">
  <Title>Toward General-Purpose Learning for Information Extraction</Title>
  <Section position="4" start_page="404" end_page="405" type="metho">
    <SectionTitle>
3 Case Study
</SectionTitle>
    <Paragraph position="0"> SRV's default feature set, designed for informal domains where parsing is difficult, includes no features more sophisticated than those immediately computable from a cursory inspection of tokens. The experiments described here were an exercise in the design of features to capture syntactic and lexical information.</Paragraph>
    <Section position="1" start_page="404" end_page="404" type="sub_section">
      <SectionTitle>
3.1 Domain
</SectionTitle>
      <Paragraph position="0"> As part of these experiments we defined an information extraction problem using a publicly available corpus. 600 articles were sampled from the &amp;quot;acquisition&amp;quot; set in the Reuters corpus (Lewis, 1992) and tagged to identify instances of nine fields. Fields include those for the official names of the parties to an acquisition (acquired, purchaser, seller), as well as their short names (acqabr, purchabr, sellerabr), the location of the purchased company or resource (acqloc), the price paid (dlramt), and any short phrases summarizing the progress of negotiations (status).</Paragraph>
      <Paragraph position="1"> The fields vary widely in length and frequency of occurrence, both of which have a significant impact on the difficulty they present for learners. null</Paragraph>
    </Section>
    <Section position="2" start_page="404" end_page="405" type="sub_section">
      <SectionTitle>
3.2 Feature Set Design
</SectionTitle>
      <Paragraph position="0"> We augmented SRV's default feature set with features derived using two publicly available  NLP tools, the link grammar parser and Wordnet. null The link grammar parser takes a sentence as input and returns a complete parse in which terms are connected in typed binary relations (&amp;quot;links&amp;quot;) which represent syntactic relationships (Sleator and Temperley, 1993). We mapped these links to relational features: A token on the right side of a link of type X has a corresponding relational feature called left_)/ that maps to the token on the left side of the link. In addition, several non-relational features, such as part of speech, are derived from parser output. Figure 1 shows part of a link grammar parse and its translation into features.</Paragraph>
      <Paragraph position="1"> Our object in using Wordnet (Miller, 1995) is to enable 5RV to recognize that the phrases, &amp;quot;A bought B,&amp;quot; and, &amp;quot;X acquired Y,&amp;quot; are instantiations of the same underlying pattern. Although &amp;quot;bought&amp;quot; and &amp;quot;acquired&amp;quot; do not belong to the same &amp;quot;synset&amp;quot; in Wordnet, they are nevertheless closely related in Wordnet by means of the &amp;quot;hypernym&amp;quot; (or &amp;quot;is-a') relation. To exploit such semantic relationships we created a single token feature, called wn_word. In contrast with features already outlined, which are mostly boolean, this feature is set-valued. For nouns and verbs, its value is a set of identifiers representing all synsets in the hypernym path to the root of the hypernym tree in which a word occurs. For adjectives and adverbs, these synset identifiers were drawn from the cluster of closely related synsets. In the case of multiple Word-net senses, we used the most common sense of a word, according to Wordnet, to construct this set.</Paragraph>
    </Section>
    <Section position="3" start_page="405" end_page="405" type="sub_section">
      <SectionTitle>
3.3 Competing Learners
</SectionTitle>
      <Paragraph position="0"> \Y=e compare the performance of 5RV with that of two simple learning approaches, which make predictions based on raw term statistics. Rote (see (Freitag, 1998)), memorizes field instances seen during training and only makes predictions when the same fragments are encountered in novel documents. Bayes is a statistical approach based on the &amp;quot;Naive Bayes&amp;quot; algorithm (Mitchell, 1997). Our implementation is described in (Freitag, 1997). Note that although these learners are &amp;quot;simple,&amp;quot; they are not necessarily ineffective. We have experimented with them in several domains and have been surprised by their level of performance in some cases.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>