<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1097">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics Semi-Supervised Training for Statistical Word Alignment</Title>
  <Section position="8" start_page="774" end_page="775" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> We presented a semi-supervised algorithm based on IBM Model 4, with modeling and search extensions, which produces alignments of improved F-measure over unsupervised Model 4 training.</Paragraph>
    <Paragraph position="1"> We used these alignments to produce translations of higher quality.</Paragraph>
    <Paragraph position="2">  The semi-supervised learning literature generally addresses augmenting supervised learning tasks with unlabeled data (Seeger, 2000). In contrast, we augmented an unsupervised learning task with labeled data. We hope that Minimum Error / Maximum Likelihood training using the EMD algorithm can be used for a wide diversity of tasks where there is not enough labeled data to allow supervised estimation of an initial model of reasonable quality.</Paragraph>
  </Section>
class="xml-element"></Paper>