<?xml version="1.0" standalone="yes"?>
<Paper uid="A88-1005">
  <Title>TWO SIMPLE PREDICTION ALGORITHMS TO FACILITATE TEXT PRODUCTION</Title>
  <Section position="7" start_page="38" end_page="39" type="relat">
    <SectionTitle>
RELATED WORK
</SectionTitle>
    <Paragraph position="0"> That a small number of words account for a large proportion of the total verbiage in conversation has been known for some time \[Kucera and Francis, 1967\].</Paragraph>
    <Paragraph position="1"> The idea of using the first several letters typed by a handicapped individual to anticipate the next desired word has been used in numerous systems (e.g., \[Giblet and Childress, 1982\], \[Picketing et al., 1984\]). The Gibler and Childress system is typical in that it uses a fewthousand-word vocabulary drawn from the general public, plus a few hundred words specific to the user of the system. The user must type the first two letters before the system provides a menu of words beginning with the letter pair. If the desired word was not on the menu, the user had to spell the word out. It was felt that one letter was not informative enough to warrant a menu. Furthermore, Gilbler and Childress showed that increasing the system vocabulary degraded the performance of their system and they recommended limitation of the vocabulary for human factors reasons.</Paragraph>
    <Paragraph position="2"> By contrast, our system costs the user no more effort in terms of selecting the first two letters - if indeed they have needed to go that far; 80 per cent of the time, they haven't needed to provide two letters. Further, there is no question that for our system, allowing the vocabulary to grow is of benefit both to system performance and to user satisfaction. null Galliers \[1987\] describes a different approach for physically handicapped  persons conversant in the Bliss communications system. Communication with Bliss involves a high degree of interpretation by the &amp;quot;listener', and Galliers reports an impressive 75 per cent success rate in automating such interpretation. The Galliers system is single-subject, as ours is, and it does use past history to facilitate interpretation. It was, however, limited to a very small domain for the experiment described.</Paragraph>
    <Paragraph position="3"> One statistic cited by this last paper was that the same text produced from the Bliss communication, had it been produced by typing into a word processing system, would have required three times as many key-press operations. Our own ratio of key-press operations to characters produced was 45 per cent for the sentence position algorithm. That is, on average it took 45 presses of a mouse button to produce 100 characters. Part of the reason for such a high ratio has to do with punctuation, capitalization, and special screens such as the number screen, which requires not only the same number of presses of the button as there are digits, for example, but additional presses of the button to summon the screen and quit the menu. But primarily the ratio seems to derive from the fact that many of the words in any text are short - &amp;quot;a', &amp;quot;to', &amp;quot;the', &amp;quot;of', &amp;quot;in', and &amp;quot;on&amp;quot; being examples from this very paragraph. If the first menu does not contain a desired two-letter word, one has to spell the first letter and then make a selection from the second menu - requiring two presses of a button.</Paragraph>
    <Paragraph position="4"> By contrast, Bliss users commonly use a telegraphic style of communication and omit function words altogether.</Paragraph>
  </Section>
class="xml-element"></Paper>