<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2014">
  <Title>Soft Syntactic Constraints for Word Alignment through Discriminative Training</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Given a parallel sentence pair, or bitext, bilingual word alignment finds word-to-word connections across languages. Originally introduced as a byproduct of training statistical translation models in (Brown et al., 1993), word alignment has become the first step in training most statistical translation systems, and alignments are useful to a host of other tasks. The dominant IBM alignment models (Och and Ney, 2003) use minimal linguistic intuitions: sentences are treated as flat strings. These carefully designed generative models are difficult to extend, and have resisted the incorporation of intuitively useful features, such as morphology.</Paragraph>
    <Paragraph position="1"> There have been many attempts to incorporate syntax into alignment; we will not present a complete list here. Some methods parse two flat strings at once using a bitext grammar (Wu, 1997). Others parse one of the two strings before alignment begins, and align the resulting tree to the remaining string (Yamada and Knight, 2001). The statistical models associated with syntactic aligners tend to be very different from their IBM counterparts.</Paragraph>
    <Paragraph position="2"> They model operations that are meaningful at a syntax level, like re-ordering children, but ignore features that have proven useful in IBM models, such as the preference to align words with similar positions, and the HMM preference for links to appear near one another (Vogel et al., 1996).</Paragraph>
    <Paragraph position="3"> Recently, discriminative learning technology for structured output spaces has enabled several discriminative word alignment solutions (Liu et al., 2005; Moore, 2005; Taskar et al., 2005). Discriminative learning allows easy incorporation of any feature one might have access to during the alignment search. Because the features are handled so easily, discriminative methods use features that are not tied directly to the search: the search and the model become decoupled.</Paragraph>
    <Paragraph position="4"> In this work, we view synchronous parsing only as a vehicle to expose syntactic features to a discriminative model. This allows us to include the constraints that would usually be imposed by a tree-to-string alignment method as a feature in our model, creating a powerful soft constraint. We add our syntactic features to an already strong flat-string discriminative solution, and we show that they provide new information resulting in improved alignments.</Paragraph>
  </Section>
class="xml-element"></Paper>