<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0431">
  <Title>Meta-Learning Orthographic and Contextual Models for Language Independent Named Entity Recognition</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Discussion
</SectionTitle>
    <Paragraph position="0"> The methodology employed was significantly more successful at identifying location and person entities (see table 1). The recalls for these values for English are especially high considering that precision is typically the higher value in named entity recognition. Although the lower value for miscellaneous entities was expected, due to the relatively smaller number of items and idiosyncrasies of the category membership, the significantly low values for organisations was surprising. There are three possible reasons for this: organisations are more likely than people or places to take their names from the contemporary lexicon, and are therefore less likely to contain orthographic structures able to be exploited by n-gram modelling; in the training set, organisations were relatively over represented in the errors made in the normalising of case information, most likely due to the previous reason; and organisations may be represented metonymically, creating ambiguity about the entity class.</Paragraph>
    <Paragraph position="1"> As the difference that meta-learning made to German was very large, but to English very small (see Results), it is reasonable to assume that the individual English classifiers were much more homogeneous, indicating both that the attribute space for the individual classifiers for English were very successful, but only certain classifiers or combinations of them were beneficial for German. The flexibility of the strategy as a whole was successful when generalising across languages</Paragraph>
  </Section>
class="xml-element"></Paper>