<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3118">
  <Title>PORTAGE: with Smoothed Phrase Tables and Segment Choice Models</Title>
  <Section position="5" start_page="135" end_page="135" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> Benchmarking with same language model and parameters as WPT05 reproduces the results with a tiny improvement. The larger language model used  in2006forEnglishyieldsabouthalfaBLEU.Good-Turing phrase table smoothing yields roughly half a BLEU point. Kneser-Ney phrase table smoothing yields between a third and half a BLEU point more than Good-Turing. Decision tree based distortion yields a small improvement for the devtest set when rescoring was not used but failed to show improvement on the test set.</Paragraph>
    <Paragraph position="1"> In summary, the results from phrase-table smoothing are extremely encouraging. On the other hand, the feature-rich decision tree distortion modelling requires additional work before it provides a good pay-back. Fortunately we have some encouraging avenues under investigation. Clearly there is more work needed for both of these areas.</Paragraph>
  </Section>
class="xml-element"></Paper>