<?xml version="1.0" standalone="yes"?>
<Paper uid="E93-1007">
  <Title>Data-Oriented Methods for Grapheme-to-Phoneme Conversion</Title>
  <Section position="6" start_page="49" end_page="50" type="evalu">
    <SectionTitle>
3.2.1 Results
</SectionTitle>
    <Paragraph position="0"> When comparing the phonemisation accuracy of the linguistic knowledge-based approach in MORPA-CUM-MORPHON to the results on the same data by the table method, we see that the table scores significantly higher.</Paragraph>
    <Paragraph position="1">  In the knowledge-based approach, errors of morphological analysis (spurious ambiguity or no analysis) account for a considerable amount of incorrect phoneme output (even after removal by \[Nunn and Van Heuven, 1993\] of proper names and other difficult cases from the test set). A new data-oriented version of MORPA (\[Heemskerk, 1993\]) assigns a priority ordering to the set of morphological decom-Sin a different set of experiments, we successfully applied the IBL approach and two other data-oriented algorithms, analogical modeling and backprop, to the stress assignment problem (see \[Gillis et al., 1992\], \[Daelemans et al., 1993\], but we have not yet tried to combine the two tasks.</Paragraph>
    <Paragraph position="2">  positions, based on a probabilistic grammar derived from a corpus of examples of correct decompositions. This new approach raises the overall performance of MORPA-CUM-MORPHON tO 88.7%, which remains slightly worse than the table method.</Paragraph>
    <Paragraph position="3"> On the basis of an analysis of confusion matrices (misclassifications per grapheme), we find that the same types of errors are made by both systems, mainly on vowels (especially on the transcription of grapheme &lt;e&gt;), but less frequently by the table method. E.g. an intended /~/ was assigned category/~/ 112 times by MOttPA-CUM-MORPHON, and only 23 times by the table method. Another difference is that while confusions by the table method are symmetric, confusions in MORPA-CUM-MORPHON seem to be directed (e.g. an intended /~/is often misclassified as/E/, but almost never the other way round).</Paragraph>
  </Section>
class="xml-element"></Paper>