<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3110">
  <Title>N-Gram Posterior Probabilities for Statistical Machine Translation</Title>
  <Section position="9" start_page="74" end_page="75" type="concl">
    <SectionTitle>
8 Conclusions
</SectionTitle>
    <Paragraph position="0"> We introduced n-gram and sentence length posterior probabilities and demonstrated their usefulness for rescoring purposes. We performed systematic experiments on the Chinese-English NIST task and showed significant improvements of the translation quality. The improvements were consistent among several evaluation sets.</Paragraph>
    <Paragraph position="1"> An interesting property of the introduced methods is that they do not require additional knowledge sources. Thus the given knowledge sources are better exploited. Our intuition is that the posterior models prefer hypotheses with n-grams that are common in the N-best list.</Paragraph>
    <Paragraph position="2"> The achieved results are promising. Despite that, there are several ways to improve the approach.</Paragraph>
    <Paragraph position="3">  Baseline At present, there is no organization claimed the attack. Rescored At present, there is no organization claimed responsibility for the attack. Reference So far, no organization whatsoever has claimed responsibility for the attack. Baseline FIFA to severely punish football fraud Rescored The International Football Federation (FIFA) will severely punish football's deception Reference FIFA will severely punish all cheating acts in the football field Baseline In more than three months of unrest, a total of more than 60 dead and 2000 injured. Rescored In more than three months of unrest, a total of more than 60 people were killed and more than 2000 injured.</Paragraph>
    <Paragraph position="4"> Reference During the unrest that lasted more than three months, a total of more than 60 people died and over 2,000 were wounded.</Paragraph>
    <Paragraph position="5"> For the decision rule in Equation 3, the model scaling factors lM1 can be multiplied with a constant factor without changing the result. This global factor would affect the proposed posterior probabilities.</Paragraph>
    <Paragraph position="6"> So far, we have not tuned this parameter, but a proper adjustment might result in further improvements.</Paragraph>
    <Paragraph position="7"> Currently, the posterior probabilities are computed on an N-best list. Using word graphs instead should result in more reliable estimates, as the number of hypotheses in a word graph is some orders of a magnitude larger than in an N-best list.</Paragraph>
  </Section>
class="xml-element"></Paper>