<?xml version="1.0" standalone="yes"?>
<Paper uid="N04-1014">
  <Title>Training Tree Transducers</Title>
  <Section position="11" start_page="0" end_page="0" type="concl">
    <SectionTitle>
10 Related Work
</SectionTitle>
    <Paragraph position="0"> Tree substitution grammars or TSG (Schabes, 1990) are equivalent to regular tree grammars. xR transducers are similar to (weighted) Synchronous TSG, except that xR can copy input trees (and transform the copies differently), but does not model deleted input subtrees.</Paragraph>
    <Paragraph position="1"> (Eisner, 2003) discusses training for Synchronous TSG.</Paragraph>
    <Paragraph position="2"> Our training algorithm is a generalization of forward-backward EM training for finite-state (string) transducers, which is in turn a generalization of the original forward-backward algorithm for Hidden Markov Models.</Paragraph>
  </Section>
class="xml-element"></Paper>