<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1096">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics An End-to-End Discriminative Approach to Machine Translation</Title>
  <Section position="11" start_page="767" end_page="767" type="concl">
    <SectionTitle>
9 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have presented a novel end-to-end discriminative system for machine translation. We studied update strategies, an important issue in on-line discriminative training for MT, and conclude that making many smaller (conservative) updates is better than making few large (aggressive) updates. We also investigated the effect of adding many expressive features, which yielded a 0.8 increase in BLEU score over monotonic Pharaoh.</Paragraph>
    <Paragraph position="1"> Acknowledgments We would like to thank our reviewers for their comments. This work was suponly used DEV to optimize the number of training iterations. 13This result is significant with p-value 0.0585 based on approximate randomization (Riezler and Maxwell, 2005).</Paragraph>
    <Paragraph position="2"> ported by a FQRNT fellowship to second author and a Microsoft Research New Faculty Fellowship to the third author.</Paragraph>
  </Section>
class="xml-element"></Paper>