<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0821">
  <Title>Improved Language Modeling for Statistical Machine Translation</Title>
  <Section position="7" start_page="127" end_page="127" type="evalu">
    <SectionTitle>
6 Results
</SectionTitle>
    <Paragraph position="0"> The results from the first decoding pass on the development set are shown in Table 1. The second column in Table 1 lists the oracle BLEU scores for the N-best lists, i.e. the scores obtained by always selecting the hypothesis known to have the highest individual BLEU score. We see that considerable improvements can in principle be obtained by a better second-pass selection of hypotheses. The language model rescoring results are shown in Table 2, for both types of second-pass language models individually, and for their combination. In both cases we obtain small improvements in BLEU score, with the 4-gram providing larger gains than the 3-gram FLM.</Paragraph>
    <Paragraph position="1"> Since their combination only yielded negligible additional improvements, only 4-grams were used for processing the final evaluation sets. The evaluation results are shown in Table 3.</Paragraph>
  </Section>
class="xml-element"></Paper>