<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0801">
  <Title>Identifying Word Correspondences in Parallel Texts. In Proceedings of the Speech and Natural</Title>
  <Section position="10" start_page="7" end_page="7" type="concl">
    <SectionTitle>
9 Conclusions
</SectionTitle>
    <Paragraph position="0"> The conventional wisdom in the statistical MT community has been that &amp;quot;heuristic&amp;quot; alignment methods based on word association statistics could not be competitive with methods that have a &amp;quot;well-founded mathematical theory that underlies their parameter estimation&amp;quot; (Och and Ney, 2003, p. 37). Our results seem to suggest that this is not the case.</Paragraph>
    <Paragraph position="1"> While we would not claim to have demonstated that association-based methods are superior to the established approach, they certainly now appear to be worth investigating further.</Paragraph>
    <Paragraph position="2"> Moreover, our alignment method is faster than standard models to train; potentially much faster if it were re-implemented in a language like C++. Efficiency issues, especially in training, are often dismissed as unimportant, but one should consider simply the number of experiments that it is possible to do in the course of system development. In our case, for example, it was impractical to try to try to optimize all the options and parameters of the Giza++ models in a reasonable amount of time, given the computational resources at our disposal.</Paragraph>
    <Paragraph position="3"> While the wealth of details regarding various passes through the data in our best methods might seem to undercut our claim of simplicity, it is important to realize that each of our methods makes a fixed number of passes, and each of those passes involves a simple procedure of computing LLR scores, collecting co-occurrence counts to estimate link probabilities, or performing competitive linking; plus one best first search for minimally nonmonotonic alignments. All these procedures are simple to understand and straightforward to implement, in contrast to some of the difficult mathematical and computational issues with the standard models.</Paragraph>
  </Section>
class="xml-element"></Paper>