<?xml version="1.0" standalone="yes"?>
<Paper uid="J00-2004">
  <Title>Models of Translational Equivalence among Words</Title>
  <Section position="10" start_page="90000" end_page="90000" type="concl">
    <SectionTitle>
7. Conclusion
</SectionTitle>
    <Paragraph position="0"> There are many ways to model translational equivalence and many ways to estimate translation models. &amp;quot;The mathematics of statistical machine translation&amp;quot; proposed by Brown et al. (1993b) are just one kind of mathematics for one kind of statistical trans- null Melamed Models of Translational Equivalence lation. In this article, I have proposed and evaluated new kinds of translation model biases, alternative parameter estimation strategies, and techniques for exploiting pre-existing knowledge that may be available about particular languages and language pairs. On a variety of evaluation metrics, each infusion of knowledge about the problem domain resulted in better translation models.</Paragraph>
    <Paragraph position="1"> Each innovation presented here opens the way for more research. Model biases can be mixed and matched with each other, with previously published biases like the word order correlation bias, and with other biases yet to be invented. The competitive linking algorithm can be generalized in various ways. New kinds of preexisting knowledge can be exploited to improve accuracy for particular language pairs or even just for particular bitexts. It is difficult to say where the greatest advances will come from. Yet, one thing is clear from our current vantage point: Research on empirical methods for modeling translational equivalence has not run out of steam, as some have claimed, but has only just begun.</Paragraph>
  </Section>
class="xml-element"></Paper>