<?xml version="1.0" standalone="yes"?>
<Paper uid="A94-1016">
  <Title>Three Heads are Better than One</Title>
  <Section position="6" start_page="99" end_page="99" type="concl">
    <SectionTitle>
5 CURRENT AND FUTURE
WORK
</SectionTitle>
    <Paragraph position="0"> Ultimately, a multi-engine system depends on the quality of each particular engine. We expect the performance of KBMT and EBMT to grow. We plan to use a standard regression mechanism to modify the scoring system based on feedback from having humans select the best covers for test texts.</Paragraph>
    <Paragraph position="1"> The current system is human-aided. We have begun an experiment with a fully-automated mode, with the understanding that the quality will drop.</Paragraph>
    <Paragraph position="2"> The most important effect of this change is that accurate quality scores become much more important, since the first choice becomes the only choice. Besides improving the KBMT and EBMT scoring mechanisms, we need to provide finer distinctions for the lexical transfer engine's output. As the databases for this are quite large (all together, over 400,000 entries), adding scores to individual entries is, in the short run, prohibitive. We have not as yet discovered any feasible automatic technique for generating such scores. Instead, we are planning to use an English language model on the output, in a manner similar to that done by speech and statistical translation systems (Brown et al., 1990). Statistically generating such a model is feasible, since it does not rely on knowing correspondences between source</Paragraph>
  </Section>
class="xml-element"></Paper>