<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0814">
  <Title>ISI's Participation in the Romanian-English Alignment Task</Title>
  <Section position="10" start_page="93" end_page="93" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> * Considering multiple stemming possibilities for each word seems important.</Paragraph>
    <Paragraph position="1"> * Alternating between increasing likelihood and decreasing error rate is a useful training approach which can be used for many problems.</Paragraph>
    <Paragraph position="2"> * Our model and training method improve upon a strong baseline for producing 1-to-many alignments. null * Our model and training method can be used with the &amp;quot;intersection&amp;quot; heuristic to produce higher quality 1-to-1 alignments * Models which can directly model many-to-many alignments and do not require heuristic symmetrization are needed to produce higher quality many-to-many alignments. Our training method can be used to train them.</Paragraph>
  </Section>
class="xml-element"></Paper>