<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3122">
  <Title>Language Models and Reranking for Machine Translation</Title>
  <Section position="6" start_page="152" end_page="152" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> By analyzing the results, we observe that a very powerful component of our system is the MERT component of Phramer. It provided a very high baseline for the devtest2006 sets (WPT05 test sets).</Paragraph>
    <Paragraph position="1"> The additional language models seem to consistently improve the results, although the increase is not very significant on FR-EN and ES-EN subtasks.</Paragraph>
    <Paragraph position="2"> The cause might be the specifics of the data involved  in this shared task - mostly European Parliament proceedings, which is different than the domain of both Treebank and English Gigaword - newswire.</Paragraph>
    <Paragraph position="3"> The enhanced LMs compete with the default LM (which is also part of the model) that is trained on European Parliament data.</Paragraph>
    <Paragraph position="4"> The word splitting heuristics offers also a small improvement for the performance on DE-EN subtask. null Voting seems to slightly improve the results in some cases (ES-EN subtask). We believe that the voting implementation reduces l weights overfitting, by combining the output of multiple local maxima of the development set. The size of the development set used to generate l1 and l2 (1000 sentences) compensates the tendency of the unsmoothed MERT algorithm to overfit (Och, 2003) by providing a high ratio between number of variables and number of parameters to be estimated.</Paragraph>
  </Section>
class="xml-element"></Paper>