<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-6007">
  <Title>Syntactic Identification of Attribution in the RST Treebank</Title>
  <Section position="3" start_page="57" end_page="59" type="metho">
    <SectionTitle>
3 Identifying Attributions
</SectionTitle>
    <Paragraph position="0"> We define three forms of Attribution relations:  a2 Basic: A verb is followed by a sentential complement position a2 Backwards: The sentential complement precedes the verb. In these cases, a trace ap- null pears as complement to the verb, and is coindexed with the sentential complement a2 According-To: the phrase &amp;quot;according to&amp;quot; occurs</Paragraph>
    <Section position="1" start_page="58" end_page="58" type="sub_section">
      <SectionTitle>
3.1 Basic Attributions
</SectionTitle>
      <Paragraph position="0"> In this form, a sentential object immediately follows a verb.</Paragraph>
      <Paragraph position="1"> Consider the example  (1) Now, the firm says it's at a turning point. In PTB, the sentence is annotated as in : (2)</Paragraph>
      <Paragraph position="3"> Sentential objects are annotated as SBAR regardless of the presence of complementizers. Thus, the subroutine searches the corpus for structures matching the template (3), which matches verb phrases in which a verb is followed by an SBAR.</Paragraph>
      <Paragraph position="4">  (3) (VP ... (V.. ...) (SBAR ...) ... )  The SBAR must follow immediately after the verb, which may be the last verb in a verbal cluster. This represents a simplification, since adverbials may occur between the verb and its SBAR complement. Our implementation correctly identifies 1497 occurrences, and incorrectly identifies 215 occurrences of attributions, corresponding to a contribution to the total recall of 0.615 with a precision of 0.874.</Paragraph>
    </Section>
    <Section position="2" start_page="58" end_page="58" type="sub_section">
      <SectionTitle>
3.2 Backwards Attributions
</SectionTitle>
      <Paragraph position="0"> Where a sentential object does not immediately follow its corresponding verb, it is represented as a trace which is coindexed with the S. In the following example, the sentential complement precedes the sentence: (4) &amp;quot;I believe that any good lawyer should be able to figure out and understand patent law&amp;quot;,a0 Judge Mayer says a1a2a0 The example is represented as follows in PTB:</Paragraph>
      <Paragraph position="2"> The sentential object of &amp;quot;says&amp;quot; is represented by the trace ((S (-NONE- *T*-6) )))), which is coindexed with the outer sentence ((S-6)).</Paragraph>
      <Paragraph position="3"> The procedure searches for sentences of the types S, S/SBAR, and VP/S-TPC which are linked to a trace in the surrounding sentence.</Paragraph>
      <Paragraph position="4"> Thus, it covers cases of topicalization and sentence inversion which are the most frequent reasons for sentential objects not occurring immediately after the verb.</Paragraph>
      <Paragraph position="5"> The subroutine covering sentential objects linked by traces make 700 correct and 4 incorrect predictions, corresponding to a recall contribution of 0.287 with a precision of 0.994.</Paragraph>
    </Section>
    <Section position="3" start_page="58" end_page="59" type="sub_section">
      <SectionTitle>
3.3 According-To Attributions
</SectionTitle>
      <Paragraph position="0"> Also categorized as attributions are &amp;quot;according to&amp;quot; expressions. These are identified with a separate subroutine which simply identifies occur- null rences of the two words &amp;quot;according&amp;quot; and &amp;quot;to&amp;quot; in sequence.</Paragraph>
      <Paragraph position="1"> Example:  (6) Now, according to a Kidder World story about Mr. Megargel, all the firm has to do is &amp;quot;position ourselves more in the deal  The subroutine identifies 87 &amp;quot;according to&amp;quot; expressions correctly, and 1 incorrectly.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="59" end_page="59" type="metho">
    <SectionTitle>
4 Discussion of Results
</SectionTitle>
    <Paragraph position="0"> Our system for recognizing Attributions is a quite direct implementation of the description of Attribution given in the RST Tagging Manual, relying on simple structural characteristics. In developing the system, we examined data in the Training portion of the RST Treebank. To ensure that our implementation was not tuned to any idiosyncrasies of the data we examined, we performed two tests of our system, on the Test portion of the RST Treebank as well as the Training portion. We avoided any examination of data in the Test portion of the Treebank.</Paragraph>
    <Paragraph position="1"> Given the general nature of the syntactic characteristics of our system, it is not surprising that the results on the Training and Test portions of the Treebank our quite similar. We present the overall results on both portions of the Treebank, followed by more detailed results, giving the contributions of the main subparts of the system.</Paragraph>
    <Section position="1" start_page="59" end_page="59" type="sub_section">
      <SectionTitle>
4.1 Overall Results
</SectionTitle>
      <Paragraph position="0"> The following figure summarizes the results of executing the procedure on the two portions of the</Paragraph>
    </Section>
    <Section position="2" start_page="59" end_page="59" type="sub_section">
      <SectionTitle>
4.2 Subparts of the System
</SectionTitle>
      <Paragraph position="0"> Next, we present the contribution of each of the three subparts of the system.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>