<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1607">
  <Title>Phrasetable Smoothing for Statistical Machine Translation</Title>
  <Section position="8" start_page="58" end_page="59" type="concl">
    <SectionTitle>
6 Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> We tested different phrasetable smoothing techniques in two different translation settings: European language pairs with relatively small corpora, and Chinese to English translation with large corpora. The smoothing techniques fall into two  categories: black-box methods that work only on phrase-pair counts; and glass-box methods that decompose phrase probabilities into lexical probabilities. In our implementation, black-box techniques use linear interpolation to combine relative frequency estimates with smoothing distributions, while glass-box techniques are combined in log-linear fashion with either relative-frequencies or black-box estimates.</Paragraph>
    <Paragraph position="1"> All smoothing techniques tested gave statistically significant gains over pure relative-frequency estimates. In the small-corpus setting, the best technique is a loglinear combination of Kneser-Ney count smoothing with Zens-Ney glass-box smoothing; this yields an average gain of 1.6 BLEU points over relative frequencies. In the large-corpus setting, the best technique is a log-linear combination of relative-frequency estimates with Zens-Ney smoothing, with a gain of 1.1 BLEU points. Of the two glass-box smoothing methods tested, Zens-Ney appears to have a slight advantage over Koehn-Och-Marcu. Of the black-box methods tested, Kneser-Ney is clearly better for small corpora, but is equivalent to Good-Turing for larger corpora.</Paragraph>
    <Paragraph position="2"> The paper describes several smoothing alternatives which we intend to test in future work:</Paragraph>
  </Section>
class="xml-element"></Paper>