<?xml version="1.0" standalone="yes"?>
<Paper uid="P98-2246">
  <Title>Neural Network Recognition of Spelling Errors</Title>
  <Section position="4" start_page="1491" end_page="1491" type="concl">
    <SectionTitle>
3 Position-maintaining and position-
</SectionTitle>
    <Paragraph position="0"> altering errors The results for the four types of error can be used to create two groupings: position-maintaining and position-altering errors. The position-maintaining errors are substitution and reversal errors, which do not cause other letters to shift to different positions. The position-altering errors (insertions and deletions), however, do cause such a shift.</Paragraph>
    <Paragraph position="1"> The scores demonstrate that for FLLB representations, position-altering errors cause greater difficulty than position-maintaining errors. The traditional left-aligned FLLB performed dramatically worse on position-altering errors (scores of 99197 and 1251152) than on position-maintaining errors (1.711.8 and 4.815.9). Both the split and bi-directional FLLBs display muchimproved performance on the position-altering errors. The bi-directional FLLB, however, still has substantially more difficulty with deletion errors than does the split FLLB. The split FLLB thus demonstrated the best overall performance of the three FLLB representations.</Paragraph>
    <Paragraph position="2"> Along the local/distributed variable, the local representations consistently equal or surpass the performance of the distributed representations.</Paragraph>
    <Paragraph position="3"> The advantage, however, is relatively minor, unlike the clear distinctions between FLLB type.</Paragraph>
    <Paragraph position="4"> Conclusion This paper has found that word and letter representations can have a significant effect on ANN recognition of spelling errors. It has specifically found that: * Methods of word representation call have substantial and measureable effects on ANN performance.</Paragraph>
    <Paragraph position="5"> * Position-altering (insertion and deletion) and position-maintaining errors (substitution and reversal) have different effects on ANN recognition of spelling errors.</Paragraph>
    <Paragraph position="6"> * An FLLB may, in addition to a traditional left-aligned representation, be organized ill split and bi-directional structures. These new FLLBs result in improved performance on position-altering errors, with tile split representation offering the best performance.</Paragraph>
    <Paragraph position="7"> Research in progress includes development of other ANN word representation methods and testing with data from other languages.</Paragraph>
  </Section>
class="xml-element"></Paper>