<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0304">
  <Title>Statistical Translation Alignment with Compositionality Constraints</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Since the pioneering work of the IBM machine translation team almost 15 years ago (Brown et al., 1990), statistical methods have proven to be valuable tools in approaching the automation of translation. Word alignments (WA) play a central role in the statistical modeling process, and reliable WA techniques are crucial in acquiring the parameters of the models (Och and Ney, 2000).</Paragraph>
    <Paragraph position="1"> Yet, the very nature of these alignments, as defined in the IBM modeling approach (Brown et al., 1993), lead to descriptions of the correspondences between source-language (SL) and target-language (TL) words of a translation that are often unsatisfactory, at least from a human perspective.</Paragraph>
    <Paragraph position="2"> One notion that is typically evacuated in the statistical modeling process is that of compositionality: a fundamental assumption in statistical machine translation is that, ultimately, all the words of a SL segment S contribute to produce all the words of its TL translation T, at least to some degree. While this makes perfect sense from a stochastic point of view, it contrasts with the hypothesis at the basis of most (if not all) other MT approaches, as well as with our natural intuitions about translation: that individual portions of the SL text produce individual TL portions autonomously, and that the final translation T is obtained by somehow piecing together these TL portions.</Paragraph>
    <Paragraph position="3"> In what follows, we show how re-integrating compositionality into the statistical translation word alignment process leads to better alignments. We first take a closer look at the &amp;quot;standard&amp;quot; statistical WA techniques in section 2, and then propose a way of imposing a compositionality constraint on these techniques in section 3. In section 4, we discuss various implementation issues, and finally present the experimental results of this approach on the WPT-03 shared task on WA in section 5.</Paragraph>
  </Section>
class="xml-element"></Paper>