<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1007">
  <Title>Maximum Entropy Models for FrameNet Classification</Title>
  <Section position="4" start_page="5" end_page="5" type="metho">
    <SectionTitle>
. The
</SectionTitle>
    <Paragraph position="0"> accuracy of the ME semantic classifier on the automatically identified frame elements is 81.5%, not a statistically significant difference from its performance on hand labeled elements, but a statistically significant difference from the classifier of  Figure 5 shows the results of varying the training set size on identification performance. For each data set, thresholds were chosen to maximize F-Score.</Paragraph>
    <Paragraph position="1">  It is clear from the results above that the performance of the ME model for frame element classification is robust to the use of automatically identified frame element boundaries. Further, the ME  G&amp;J's results for the combined task were generated with a threshold applied to the FE classifier (Dan Gildea, personal communication). This is why their precision/recall scores are dissimilar to their accuracy scores, as reported in section 3. Because the ME classifier does not employ a threshold, comparisons must be based on F-score.</Paragraph>
    <Paragraph position="2"> framework yields better results on the frame element identification task than the simple linear interpolation model of Gildea and Jurafsky. This result is not surprising given the discussion in Section 3.</Paragraph>
    <Paragraph position="3"> What is striking, however, is the drastic overall reduction in performance on the combined identification and classification task. The bottleneck here is the identification of frame element boundaries. Unlike with classification though, Figure 5 indicates that a plateau in the learning curve has been reached, and thus, more data will not yield as dramatic an improvement for the given feature set and model.</Paragraph>
  </Section>
class="xml-element"></Paper>