<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2040">
  <Title>A Finite State and Data-Oriented Method for Grapheme to Phoneme Conversion</Title>
  <Section position="6" start_page="308" end_page="308" type="concl">
    <SectionTitle>
5 Concluding remarks
</SectionTitle>
    <Paragraph position="0"> We have presented a method for grapheme to phoneme conversion, which combines a hand-crafted finite state transducer with rules induced by a transformation-based learning. An advantage of this method is that it is able to achieve a high level of accuracy using relatively small training sets. Busser (1998), for instance, uses a memory-based learning strategy to achieve 90.1% word accuracy on the same task, but used 90% of the CELEX data (over 300K words) as training set and a (character/phoneme) window size of 9. Hoste et al. (2000a) achieve a word accuracy of 95.7% and a phoneme accuracy of 99.5% on the same task, using a combination of machine learning techniques, as well as additional data obtained from a second dictionary.</Paragraph>
    <Paragraph position="1"> Given the result of Roche and Schabes (1997), an obvious next step is to compile the induced rules into an actual transducer, and to compose this with the hand-crafted transducer. It should be noted, however, that the number of induced rules is quite large in some of the experiments, so that the compilation procedure may require some attention.</Paragraph>
  </Section>
class="xml-element"></Paper>