<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3121">
  <Title>Phramer - An Open Source Statistical Phrase-Based Translator</Title>
  <Section position="6" start_page="148" end_page="148" type="concl">
    <SectionTitle>
3.5 Conclusions
</SectionTitle>
    <Paragraph position="0"> The enhancements that we proposed provide small improvements on the devtest2006 files. As expected, when we used the NN-ADJ inversion the ratio lDlLM increased from 0.545 to 0.675. The LM is the only model that opposes the tendency of the distortion model towards monotone phrase order.</Paragraph>
    <Paragraph position="1"> Phramer delivers a very good baseline system.</Paragraph>
    <Paragraph position="2"> Using only the baseline system, we obtain +0.68 on DE-EN, +0.43 on FR-EN and -0.18 on ES-EN difference in BLEU score compared to WPT05's best system (Koehn and Monz, 2005). This fact is caused by the MERT module. This module is capable of estimating parameters over a large development corpus in a reasonable time, thus it is able to generate highly relevant parameters.</Paragraph>
  </Section>
class="xml-element"></Paper>