<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1065">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics Improved Discriminative Bilingual Word Alignment</Title>
  <Section position="11" start_page="518" end_page="519" type="concl">
    <SectionTitle>
9 Conclusions
</SectionTitle>
    <Paragraph position="0"> For Canadian Hansards data, the test-set AER of 4.7% for our stage 2 model is one of the lowest yet reported for an aligner that makes no use of the expensive IBM models, and our test-set AER of 3.7% for the stage 2 model in combination with the HMM log odds and Model 4 intersection features is the lowest yet reported for any aligner.4 Perhaps if any general conclusion is to be drawn from our results, it is that in creating a discrim3At this writing we have not yet had time to try this with SVM training.</Paragraph>
    <Paragraph position="1"> 4However, the difference between our result and the 3.8% of Lacoste-Julien et al. is almost certainly not significant.  inative word alignment model, the model structure and features matter the most, with the discriminative training method of secondary importance. While we obtained a small improvements by varying the training method, few of the differences were statistically significant. Having better features was much more important.</Paragraph>
  </Section>
class="xml-element"></Paper>