<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0424">
  <Title>Language Independent NER using a Maximum Entropy Tagger</Title>
  <Section position="6" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
5 Results
</SectionTitle>
    <Paragraph position="0"> The baseline development results for English using the supertagger features only are given in Table 3. The full system results for the English development data are given in Table 7. Clearly the additional features have a significant impact on both precision and recall scores across all entities. We have found that the word type features are particularly useful, as is the memory feature. The performance of the final system drops by 1.97% if these features are removed. The performance of the system if the gazetteer features are removed is given in Table 4. The sizes of our gazetteers are given in Table 6. We have experimented with removing the other contextual predicates but each time performance was reduced, except for the next-next unigram tag feature which was switched off for all final experiments.</Paragraph>
    <Paragraph position="1"> The results for the Dutch test data are given in Table 5.</Paragraph>
    <Paragraph position="2"> These improve upon the scores of the best performing  the development data. The results for the German development and test sets are given in Table 7. For the German NER we removed the lowercase more frequent than uppercase feature. Apart from this change, the system was identical. We did not add any extra gazetteer information for German.</Paragraph>
  </Section>
class="xml-element"></Paper>