<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0834">
  <Title>Word Graphs for Statistical Machine Translation</Title>
  <Section position="8" start_page="196" end_page="197" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have described word graphs for statistical machine translation. The generation of word graphs during the search process has been described in detail. We have shown detailed statistics of the individual steps of the translation process and have given insight in the experimental complexity of each step. We have described an ef cient and optimal  score as a function of the word graph density.</Paragraph>
    <Paragraph position="1"> pruning method for word graphs. Using these technique, we have generated compact word graphs for two Chinese English tasks. For the IWSLT task, the graph error rate drops from about 50% for the single-best hypotheses to 17% of the word graph. Even for the NIST task, with its very large vocabulary and long sentences, we were able to reduce the graph error rate signi cantly from about 64% down to 23%.</Paragraph>
  </Section>
class="xml-element"></Paper>