<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1006">
  <Title>Use of Deep Linguistic Features for the Recognition and Labeling of Semantic Arguments</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 The PropBank and the Labeling of
</SectionTitle>
    <Paragraph position="0"> Semantic Roles The PropBank (Kingsbury et al., 2002) annotates the PTB with dependency structures (or 'predicateargument' structures), using sense tags for each word and local semantic labels for each argument and adjunct. Argument labels are numbered and used consistently across syntactic alternations for the same verb meaning, as shown in Figure 1. Adjuncts are given special tags such as TMP (for temporal), or LOC (for locatives) derived from the original annotation of the Penn Treebank. In addition to the annotated corpus, PropBank provides a lexicon which lists, for each meaning of each annotated verb, its roleset, i.e., the possible arguments in the predicate and their labels. As an example, the entry for the verb kick, is given in Figure 2. The notion of &amp;quot;meaning&amp;quot; used is fairly coarse-grained, typically motivated from differing syntactic behavior. Since each verb meaning corresponds to exactly one roleset, these terms are often used interchangeably. The roleset also includes a &amp;quot;descriptor&amp;quot; field which is intended for use during annotation and as documentation, but which does not have any theoretical standing. Each entry also includes examples. Currently there are frames for about 1600 verbs in the corpus, with a total of 2402 rolesets.</Paragraph>
    <Paragraph position="1"> Since we did not yet have access to a corpus annotated with rolesets, we concentrate in this paper on predicting the role labels for the arguments. It is only once we have both that we can interpret the relation between predicate and argument at a very fine level (for example, truck in he kicked the truck withhay as the destination of the loading action). We will turn to the problem of assigning rolesets to predicates once the data is available. We note though that preliminary investigations have shown that for about 65% of predicates (tokens) in the WSJ, there is only one roleset. In a further 7% of predicates (tokens), the set of semantic labels on the arguments of that predicate completely disambiguates the roleset.</Paragraph>
  </Section>
class="xml-element"></Paper>