<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0506">
  <Title>Pre-processing Closed Captions for Machine Translation</Title>
  <Section position="9" start_page="44" end_page="44" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> It is apparent that the peculiarity of closed captions, both in terms of transcribed speech characteristic and constraints due to the input format, require an ad hoc treatment, considerably different from the approaches suitable for written documents. Yet the knowledge about a language (or the bilingual knowledge about a language-pair) is largely invariant across different applications domains and should therefore be portable from one application domain to another. The architecture we have proposed strives to combine the need for domain independent linguistic resources and linguistically principled methods with the need for robust MT systems tuned to real world, noisy and idiosyncratic input, as encountered when embedding MT in real woi:ld devices.</Paragraph>
    <Paragraph position="1"> In terms of adequacy, a standard evaluation and a comparison among different MT systems frtom different domains is hard, as the adequacy of a system depends on its application (Church and Hovy, 1993). This is even truer with-closed captions, where the use of translation output is heavily influenced by operational constraints (time constraints, the presence of images, sound, etc.). In some cases such constraints may place a heavier burden on a system (e.g. the time constraint), in some other cases Judge 1 Judge 2 Both agreed  they can make an imperfect translation acceptable (e.g. the presence of images and sounds). We did not attempt an assessment in absolute terms, which we believe should take into account the operational environment and involve real-world users. More modestly, we aimed at showing that our pre-processing techniques provide an improvement in performance.</Paragraph>
    <Paragraph position="2"> Our work on closed captions also shows that the challenges coming from this domain, even in terms on low-level issues of input format, can lead to interesting developments of new linguistic techniques. We believe that our solutions to specific problems (namely, proper name recognition and segmentation) in the closed caption domain bear relevance to a wider context, and offer techniques that can be usefully employed in a wider range of applications.</Paragraph>
  </Section>
class="xml-element"></Paper>