<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1059">
  <Title>Language Model Adaptation for Statistical Machine Translation with Structured Query Models</Title>
  <Section position="9" start_page="11" end_page="11" type="concl">
    <SectionTitle>
TM
</SectionTitle>
    <Paragraph position="0"> Q gives now the best translation results. Adding word order information to the queries obviously helps to reduce the noise in the retrieved data by selecting sentences, which are closer to the good translations, The best results using the adapted language models are NIST score 8.12 for using the 2000 most similar sentences, whereas Bleu score goes up to 0.2068 when using 4000 sentences for language model adaptation.</Paragraph>
    <Section position="1" start_page="11" end_page="11" type="sub_section">
      <SectionTitle>
4.5 Example
</SectionTitle>
      <Paragraph position="0"> Table-3 shows translation examples for the 17 th Chinese sentence in the test set. We applied the baseline system (Base), the bag-of-word query model (Hyp1), and the structured query model (Hyp2) using AFE corpus.</Paragraph>
      <Paragraph position="1"> Ref The police has already blockade the scene of the explosion.</Paragraph>
      <Paragraph position="2"> Base At present, the police had cordoned off the explosion.</Paragraph>
      <Paragraph position="3"> Hyp1 At present, police have sealed off the explosion.</Paragraph>
      <Paragraph position="4"> Hyp2 Currently, police have blockade on the scene of the explosion.</Paragraph>
    </Section>
    <Section position="2" start_page="11" end_page="11" type="sub_section">
      <SectionTitle>
Table-3 Translation examples
4.6 Oracle Experiment
</SectionTitle>
      <Paragraph position="0"> Finally, we run an oracle experiments to see how much improvement could be achieved if we only selected better data for the specific language models. We converted the four available reference translations into structured query models and retrieved the top 4000 relevant sentences from AFE corpus for each source sentence. Using these language models, interpolated with the background language model gave a NIST score of 8.67, and a Bleu score of 0.2228. This result indicates that there is room for further improvements using this language model adaptation technique.</Paragraph>
      <Paragraph position="1"> The oracle experiment suggests that better initial translations lead to better language models and thereby better 2 nd iteration translations. This lead to the question if we can iterate the retrieval process several times to get further improvement, or if the observed improvement results form using for (good) translations, which have more diversity than the translations in an n-best list.</Paragraph>
      <Paragraph position="2"> On the other side the oracle experiment also shows that the optimally expected improvement is limited by the translation model and decoding algorithm used in the current SMT system.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>