<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2014">
  <Title>Soft Syntactic Constraints for Word Alignment through Discriminative Training</Title>
  <Section position="9" start_page="111" end_page="111" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have presented a discriminative, syntactic word alignment method. Discriminative training is conducted using a highly modular SVM for structured output, which allows code reuse between the syntactic aligner and a maximum matching baseline. An ITG parser is used for the alignment search, exposing two syntactic features: the use of inverted productions, and the use of spans that would not be available in a tree-to-string system. This second feature creates a soft phrasal cohesion constraint. Discriminative training allows us to maintain all of the features that are useful to the maximum matching baseline in addition to the new syntactic features. We have shown that these features produce a 22% relative reduction in error rate with respect to a strong flat-string model.</Paragraph>
  </Section>
class="xml-element"></Paper>