<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1622">
  <Title>Semantic Role Labeling via Instance-Based Learning</Title>
  <Section position="10" start_page="186" end_page="187" type="concl">
    <SectionTitle>
6 Summary and Remarks
</SectionTitle>
    <Paragraph position="0"> This paper has shown that basic syntactic information is useful for Semantic role labeling using instance-based learning techniques. Specifically, the following have been demonstrated: 1. It is possible to achieve acceptable F1 scores with considerably faster execution times (compared to Gildea &amp; Jurasky, 2002) for the Semantic role labeling problem using the Priority Maximum Likelihood instance-based learning algorithm and the  (PARA) as a preprocessing step, without any training given a state-of-the-art parser such as Charniak's parser. The overall performance on WSJ 23 dataset is 71.02 in F1 score. Performance drops to 60.55 for the Brown corpus, but this appears to be similar to performance drops experienced by other systems reported in CoNLL-2005.</Paragraph>
    <Paragraph position="1"> 2. F1 performance is better for PML than for kNN, where the computational complexity for PML is O( m * log n ) as opposed to O( m * n ) for kNN, where m denotes the number of features and n denotes the number of training instances.</Paragraph>
    <Paragraph position="2"> 3. Execution time for the instance-based learning presented here is about four times faster for SRL than the comparable approach used by Palmer, (2005). That is, PARA plays an important role reducing the overhead during classification when using instance-based learning.</Paragraph>
    <Paragraph position="3"> 4. Using PARA, and other modifications such as the preposition feature and Actor heuristic, improves the accuracy of both kNN and PML in comparison to similar approaches.</Paragraph>
  </Section>
class="xml-element"></Paper>