<?xml version="1.0" standalone="yes"?>
<Paper uid="A88-1005">
  <Title>TWO SIMPLE PREDICTION ALGORITHMS TO FACILITATE TEXT PRODUCTION</Title>
  <Section position="8" start_page="39" end_page="39" type="concl">
    <SectionTitle>
CONCLUSION
</SectionTitle>
    <Paragraph position="0"> In summary, evidence exists that for a system built around a single user's language, a prediction scheme that simply anticipated fifty or so words would on average be correct about half the time.</Paragraph>
    <Paragraph position="1"> Limiting such a system to only the top 20 most frequent words would give a success rate of about 30 per cent. However, not all of the high frequency words are distributed evenly by sentence position. A system that offers the top 20 most frequently occurring words for each position of a sentence was successful about 40 per cent of the time on the next 97 sentences.</Paragraph>
    <Paragraph position="2"> Allowing a user to reject the first set of words by giving the first letter of the desired word and offering the 20 most frequent words beginning with that letter resulted in success for the combined first and second menus 82 per cent of the time.</Paragraph>
    <Paragraph position="3"> After a training body of 1750 sentences (14,669 words), with a vocabulary of 1512 words, it was still the case that about six per cent of the desired words were unknown to the system.</Paragraph>
    <Paragraph position="4"> An alternative algorithm for the first offering of 20 words, based primarily on the right hand contexts of the high frequency words, is successful on the first guess 50 per cent of the time.</Paragraph>
  </Section>
class="xml-element"></Paper>