<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0304">
  <Title>Statistical Translation Alignment with Compositionality Constraints</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> In this article, we showed how a compositionality constraint could be imposed when computing word alignments with IBM Models-2. Our experiments on the WPT-03 shared task on WA demonstrated how this improves the quality of resulting alignments, when compared to standard Viterbi alignments. Our results also highlight  the benefit of using both forward and reverse translation models for this task.</Paragraph>
    <Paragraph position="1"> One of the weaknesses of the proposed method is the inability to produce many-to-many alignments. To allow for such alignments, it would be necessary to establish a &amp;quot;stopping condition&amp;quot; on the recursion process, so as to prevent partitioning pairs of segments that display &amp;quot;noncompositional&amp;quot; phenomena in both SL and TL languages. We have begun experimenting with various such mechanisms. One of these is to stop the recursion as soon as the pair of segments under consideration contains less than two &amp;quot;sure&amp;quot; alignments, i.e. connections predicted by both the forward and reverse models. Another possibility is to establish a threshold on the probability &amp;quot;drop&amp;quot; incurred by the optimal split on any given pair of segments. So far, these experiments are inconclusive.</Paragraph>
    <Paragraph position="2"> Another problem is with &amp;quot;null&amp;quot; alignments, which the program is also unable to account for. Currently, omissions and insertions in translation find themselves incorporated into aligned segments. A simple way to deal with this problem would be to exclude from the final alignment links that are not predicted by either the forward or reverse Viterbi alignments. But early experiments with this approach are unconvincing, and more elaborate filtering mechanisms will probably be necessary.</Paragraph>
    <Paragraph position="3"> Finally, IBM Model 2 is certainly not the state of the art in statistical translation modeling. Thenagain, the methods proposed here are not dependent on the underlying translation model, and similar WA methods could be based on more elaborate models, such as Models 3-5, or the HMM-based models proposed by Och et al. (1999) for example. On the other hand, our compositional alignment method could be used during the training process of higher-level models. Whether this would lead to better estimates of the models' parameters remains to be seen, but it is certainly a direction worth exploring.</Paragraph>
  </Section>
class="xml-element"></Paper>