<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-1015">
  <Title>Word Alignment via Quadratic Assignment</Title>
  <Section position="6" start_page="118" end_page="118" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have shown that the discriminative approach to word alignment can be extended to allow flexible fertility modeling and to capture first-order interactions between alignments of consecutive words.</Paragraph>
    <Paragraph position="1"> These extensions significantly enhance the expressive power of the discriminative approach; in particular, they make it possible to capture phenomena of monotonicity, local inversion and contiguous fertility trends--phenomena that are highly informative for alignment. They do so while remaining computationally efficient in practice both for prediction and for parameter estimation.</Paragraph>
    <Paragraph position="2"> Our best model achieves a relative AER reduction of 25% over the basic matching formulation, beating intersected IBM Model 4 without the use of any compute-intensive features. Including Model 4 predictions as features, we achieve a further relative AER reduction of 32% over intersected Model 4 alignments. By also including predictions of another model, we drive AER down to 3.8. We are currently investigating whether the improvement in AER results in better translation BLEU score. Allowing higher fertility and optimizing a recall biased cost function provide a significant increase in recall relative to the intersected IBM model 4 (from 88.1% to 94.4%), with only a small degradation in precision. We view this as a particularly promising aspect of our work, given that phrase-based systems such as Pharaoh (Koehn et al., 2003) perform better with higher recall alignments.</Paragraph>
  </Section>
class="xml-element"></Paper>