<?xml version="1.0" standalone="yes"?>
<Paper uid="H94-1019">
  <Title>AUTOMATIC EVALUATION OF COMPUTER GENERATED TEXT: A PROGRESS REPORT ON THE TEXTEVAL PROJECT</Title>
  <Section position="10" start_page="112" end_page="112" type="concl">
    <SectionTitle>
CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> There is some evidence that ratings based on multinomials are capable of capturing some human intuitions about translation quality.</Paragraph>
    <Paragraph position="1"> Although this varies from text to text, it does appear that the part of speech information which can be obtained by automatic tagging represents a promising way of collapsing across equivalence classes of words. By ,contrast, the results of multidimensional scaling using word counts suggest that there is too much irrelevant information in these counts to allow an automatic system to make much use of them in rating translations.</Paragraph>
    <Paragraph position="2"> Both the direct approach and that using multidimensional scaling show some success, although each failed on the translation for which the other succeeded.</Paragraph>
  </Section>
class="xml-element"></Paper>