<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1622">
  <Title>Semantic Role Labeling via Instance-Based Learning</Title>
  <Section position="11" start_page="187" end_page="187" type="ackno">
    <SectionTitle>
5. The best system developed for this paper
</SectionTitle>
    <Paragraph position="0"> (PML &amp; PARA) is still outperformed by some of the best systems from CoNLL-2005 when it comes to accuracy, but it is much simpler and is many orders-ofmagnitude faster at delivering acceptable performance.</Paragraph>
    <Paragraph position="1"> With the latest revised and optimized PML, the performance on WSJ 23 is 71.22 in F1, and the speed is 0.623 second per sentence with 3.0G CPU and 1 G RAM. Koomen et al. (2006), with more than 25 features, achieved the best results reported in CoNLL2005 on WSJ 24; but PML's performance (using PARA as a preprocessor, and seven features) achieves an F1 measure 5.10 less than Kooman's system (74.76) on WSJ 24 utilising Charniak-1 parses, and 4.07 less when using Kooman's test result (WSJ 23) as knownboundary input. In this experiment, with the Actor heuristic, PML delivers better accuracy for A0 (89.96%) than Kooman's (88.22%), but the recall (83.53%) is 4.35 % lower than Kooman's (87.88%). There are some spaces to improve PML such as low accuracy on AM-MOD, and AM-NEG, and duplicate core roles, and forth.</Paragraph>
    <Paragraph position="2"> Future work will investigate using more features, new heuristics and/or other ML approaches to improve the performance of instance-based learning algorithms at the SRL task.</Paragraph>
  </Section>
class="xml-element"></Paper>