<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0433">
  <Title>Kowloon</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> This paper has presented an overview of our entry to the CoNLL-2003 shared task. As individual component models, we constructed strong AdaBoost.MH models, SVM models, and Stacked TBL models, and provided them with detailed features on the data.</Paragraph>
    <Paragraph position="1"> We then demonstrated several stacking and voting models that proved capable of improving performance further. This was non-trivial since the individual component models were all quite strong to begin with. Because of this the vast majority of classifier combination models we tested actually turned out to degrade performance, or showed zero improvement. The models presented here worked well because they were each motivated by detailed analyses.</Paragraph>
    <Paragraph position="2"> We did investigate a number of ways in which gazetteers could be incorporated. The gazetteer supplied for the shared task was found not to improve performance significantly, because our models were already adequately powerful to correctly identify most of the named entities supplied by the gazetteer. However, minimal effort to augment the gazetteers did result in a performance boost. Moreover, performance was further improved by the inclusion of a common word lexicon not containing any named entities.</Paragraph>
    <Paragraph position="3"> Inspection revealed that some errors found in the output of the system stemmed from either erroneous sentence boundaries in the test data, or difficult-to-avoid inconsistencies in the the gold standard annotations. For example, in the following:  1. . . . [ Panamanian ]MISC boxing legend . . .</Paragraph>
    <Paragraph position="4"> 2. . . . [ U.S. ]LOC collegiate national champion . . .  both &amp;quot;Panamanian&amp;quot; and &amp;quot;U.S.&amp;quot; are used as modifiers, but one is annotated as a MISC-type NE while the other is considered a LOC-type.</Paragraph>
    <Paragraph position="5"> The stacked voted stacked model obtained an improvement of 4.83 F-Measure points on the English development set over our best model from the CoNLL-2002 shared task which we took as our baseline, resulting in a substantial 19.7% error rate reduction. The system achieves this respectable performance using very little in the way of outside resources--only a part-of-speech tagger and some common wordlists--which can be obtained easily for almost any major language. Most features we used can also be used for uninflected and non-Indo-European languages such as Chinese, where the prefixes and suffixes can be replaced by decomposing the words at the character level. This is in keeping with the the language-independent spirit of the shared task.</Paragraph>
  </Section>
class="xml-element"></Paper>