<?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="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
8 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have presented various alternative approaches to predicting PropBank role labels using forms of linguistic information that are deeper than the PTB's surface-syntax labels. These features may either be directly derived from a TAG, such as Supertag path, or indirectly via aspects of supertags, such  Task: determine Recall Precision F base + arg 0.65 0.75 0.70 base + bnd 0.48 0.55 0.51 base + bnd + arg 0.48 0.55 0.51 as deep syntactic features like Drole. These are found to produce substantial improvements in accuracy. We believe that such improvement is due to these features better capturing the syntactic information that is relevant for the task of semantic labeling. Also, these features represent syntactic categories about which there is a broad consensus in the literature. Therefore, we believe that our results are portable to other frameworks and differently annotated corpora such as dependency corpora.</Paragraph>
    <Paragraph position="1"> We also show that predicting labels from a &amp;quot;lightweight&amp;quot; parser that generates deep syntactic features performs comparably to using a full parser that generates only surface syntactic features. Improvements along this line may be attained by use of a full TAG parser, such as Chiang (2000) for example. null</Paragraph>
  </Section>
class="xml-element"></Paper>