<?xml version="1.0" standalone="yes"?>
<Paper uid="C00-2122">
  <Title>Learning to Select a Good Translation</Title>
  <Section position="8" start_page="847" end_page="848" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have described certain difficulties that arise fl'om the attempt to integrate multiple alternative translation paths and to choose their optimal coml)ination into one 'l)est' translatiou.</Paragraph>
    <Paragraph position="1"> Using confidence values that originate fl'om different translation modules as our basic selection criterion, we have introduced a learning method which enables us to select in maximal accord with decisions taken by human annotators. Along the way, we have also tackled some problematic aspects of translation evaluation as such, described some additional sources of information that are used by our selection module, and briefly sketched the way in which it; supports quality of service specifications. The extent to which this module succeeds in creating higher quality compound translations is of course highly dependent on the appropriate assignment of confidence values, pertbrmed by the translation modules themselves. As a rough criterion tbr evaluating our success, we compared the selection module's output to the best resuits achieved by a single translation path. Recent Verbmobil evaluation results demonstrate an improvement of 27.8% achieved by the selection module, measured by the number of dialogue turns that were marked 'good' by ammtators who were presented with live alternative translations tbr each turn, namely, those delivered by the four single paths, and the coml)ound translation delivered by the selection module.</Paragraph>
  </Section>
class="xml-element"></Paper>