<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0821">
  <Title>Improved Language Modeling for Statistical Machine Translation</Title>
  <Section position="8" start_page="127" end_page="127" type="concl">
    <SectionTitle>
7 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have demonstrated improvements in BLEU score by utilizing more complex language models in the rescoring pass of a two-pass SMT system.</Paragraph>
    <Paragraph position="1"> We noticed that FLMs performed worse than word-based 4-gram models. However, only trigram FLM were used in the present experiments; larger improvements might be obtained by 4-gram FLMs.</Paragraph>
    <Paragraph position="2"> The weights assigned to the second-pass language models during weight optimization were larger than those assigned to the first-pass language model, suggesting that both the word-based model and the FLM provide more useful scores than the baseline language model. Finally, we observed that the overall improvement represents only a small portion of the possible increase in BLEU score as indicated by the oracle results, suggesting that better language models do not have a significant effect on the overall system performance unless the translation model is improved as well.</Paragraph>
  </Section>
class="xml-element"></Paper>