<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1042">
  <Title>A Dynamic Bayesian Framework to Model Context and Memory in Edit Distance Learning: An Application to Pronunciation Classification</Title>
  <Section position="7" start_page="344" end_page="344" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> We have shown how the problem of learning edit distance costs from data can be modeled quite naturally using Dynamic Bayesian Networks even though the problem lacks the temporal or order constraints that other problems such as speech recognition exhibit. This gives us confidence that other important problems such as machine translation can benefit from a Graphical Models perspective. Machine translation presents a fresh set of challenges because of the large combinatorial space of possible alignments between the source string and the target.</Paragraph>
    <Paragraph position="1"> There are several extensions to this work that we intend to implement or have already obtained preliminary results on. One is simple and block transposition. Another natural extension is modeling edit distance of multiple strings.</Paragraph>
    <Paragraph position="2"> It is also evident from the large number of dependency structures that were explored that our learning algorithm would benefit from a structure learning procedure. Maximum likelihood optimization might, however, not be appropriate in this case, as exemplified by the failure of some models to discriminate between different pronunciations. Discriminative methods have been used with significant success in training HMMs. Edit distance learning could benefit from similar methods.</Paragraph>
  </Section>
class="xml-element"></Paper>