<?xml version="1.0" standalone="yes"?>
<Paper uid="E93-1007">
  <Title>Data-Oriented Methods for Grapheme-to-Phoneme Conversion</Title>
  <Section position="8" start_page="51" end_page="51" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> In computational linguistics, one of the common-sense beliefs is that the performance of a system solving a linguistic problem improves with the amount of linguistic knowledge and sophistication incorporated into it. We have shown that at least for one linguistic task, this is not the case. The linguistically least informed method (compression of a training set into a table, complemented with a rudimentary form of probabilistic reasoning) performed better on unseen input than a linguistically sophisticated, state-of-the-art knowledge-based system. We have reason to believe that this also applies to other linguistic categorisation problems where superficial input features and local context solve most ambiguities (we have positive results on stress assignment, \[Gillis et al., 1992; Daehmans et al., 1993\], and part of speech tagging).</Paragraph>
    <Paragraph position="1"> The data-oriented algorithms described are simple and domain-independent, and introduce a new kind of reusability into computational linguistics: reusability of the acquisition method (on different data sets) rather than reusability of (hand-coded) knowledge in different applications or formalisms.</Paragraph>
    <Paragraph position="2"> The former type of reusability seems to be easier to achieve than the latter.</Paragraph>
  </Section>
class="xml-element"></Paper>