<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2101">
  <Title>Optimization Finnish- French- German- Procedure English English English</Title>
  <Section position="10" start_page="793" end_page="793" type="concl">
    <SectionTitle>
8 Conclusions
</SectionTitle>
    <Paragraph position="0"> Despite the challenging shape of the error surface, we have seen that it is practical to optimize task-specific error measures rather than optimizing likelihood--it produces lower-error systems. Different methods can be used to attempt this global, non-convex optimization. We showed that for MT, and sometimes for dependency parsing, an annealed minimum risk approach to optimization performs significantly better than a previous line-search method that does not smooth the error surface. It never does significantly worse.</Paragraph>
    <Paragraph position="1"> With such improved methods for minimizing error,wecanhopetomakebetteruseoftask-specific null training criteria in NLP.</Paragraph>
  </Section>
class="xml-element"></Paper>