<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2014">
  <Title>Soft Syntactic Constraints for Word Alignment through Discriminative Training</Title>
  <Section position="8" start_page="111" end_page="111" type="relat">
    <SectionTitle>
6 Related Work
</SectionTitle>
    <Paragraph position="0"> Several other aligners have used discriminative training. Our work borrows heavily from (Taskar et al., 2005), which uses a max-margin approach with a weighted maximum matching aligner.</Paragraph>
    <Paragraph position="1"> (Moore, 2005) uses an averaged perceptron for training with a customized beam search. (Liu et al., 2005) uses a log-linear model with a greedy search. To our knowledge, ours is the first alignment approach to use this highly modular structured SVM, and the first discriminative method to use an ITG for the base aligner.</Paragraph>
    <Paragraph position="2"> (Gildea, 2003) presents another aligner with a soft syntactic constraint. This work adds a cloning operation to the tree-to-string generative model in (Yamada and Knight, 2001). This allows subtrees to move during translation. As the model is generative, it is much more difficult to incorporate a wide variety of features as we do here. In (Zhang and Gildea, 2004), this model was tested on the same annotated French-English sentence pairs that we divided into training and test sets for our experiments; it achieved an AER of 0.15.</Paragraph>
  </Section>
class="xml-element"></Paper>