<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0623">
  <Title>A Joint Model for Semantic Role Labeling</Title>
  <Section position="8" start_page="175" end_page="175" type="concl">
    <SectionTitle>
6 Experiments and Results
</SectionTitle>
    <Paragraph position="0"> For our final results we used a joint model with a = 1.5 (local model weight), b = 1 (parse tree log-probability weight) , N = 15 (candidate labelings from the local model to consider) , and k = 5 (number of alternative parses). The whole training set for the CoNLL-2005 task was used to train the models. It takes about 2 hours to train a local identification model, 40 minutes to train a local classification model, and 7 hours to train a joint re-ranking model.2 In Table 1, we present our final development and test results using this model. The percentage of perfectly labeled propositions for the three sets is 55.11% (development), 56.52% (test), and 37.06% (Brown test). The improvement achieved by the joint model relative to the local model is about 2 points absolute in F-Measure, similar to the improvement when gold-standard syntactic parses are used (Toutanova et al., 2005). The relative error reduction is much lower for automatic parses, possibly due to a lower upper bound on performance. It is clear from the drop in performance from the WSJ to Brown test set that our learned model's features do not generalize very well to related domains.</Paragraph>
  </Section>
class="xml-element"></Paper>