<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-1019">
  <Title>A Comparison of Syntactically Motivated Word Alignment Spaces</Title>
  <Section position="6" start_page="151" end_page="151" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have presented two new alignment spaces based on adependency treeprovided forone ofthe sentences in a sentence pair. We have given grammars to conduct a perfect search of these spaces using an ITG parser. The grammars derive exactly one structure for each alignment.</Paragraph>
    <Paragraph position="1"> We have shown that syntactic constraints alone can have a very positive effect on alignment error rate. With a learned objective function, ITG constraints reduce maximum matching's error rate by 10%, while D-ITG constraints produce a 31% reduction. This gap in error rate demonstrates that a dependency tree over the English sentence can be a very powerful tool when making alignment decisions. We have also shown that while dependency constraints might limit alignment expressiveness too much for some tasks, enforcing ITG constraints results in almost no reduction in achievable recall.</Paragraph>
  </Section>
class="xml-element"></Paper>