<?xml version="1.0" standalone="yes"?>
<Paper uid="C00-1075">
  <Title>Application of Analogical Modelling to Example Based Machine Translation</Title>
  <Section position="7" start_page="520" end_page="521" type="concl">
    <SectionTitle>
5. Conclusion &amp; Future Work
</SectionTitle>
    <Paragraph position="0"> We have presented a self-modelling, incremental analogical algorithm for extracting translatiou patterns fi'om existing bilingual corpora as well as a method for efficient storage and representation of extracted relations between various units of text.</Paragraph>
    <Paragraph position="1"> Not surprisingly, the quality of the results depends on the available information in terms of quantity as  well as quality and depth. Lack of any kind of linguistic information will consequently result in translation rules based only owl &amp;quot;shallow&amp;quot; evidence. Similarly, information of low quality will generate erroneous rules. However, this is a basic presupposition of any EBMT system: &amp;quot;what you give is what you get...&amp;quot;.</Paragraph>
    <Paragraph position="2"> Tile proposed fi'amework was initially evaluated based only owl string form information. However, tile model can easily take into account &amp;quot;deeper&amp;quot; linguistic knowledge during the learning phase, thus improving tile quality of the final results.</Paragraph>
    <Paragraph position="3"> Evaluation of learning performance in this case is the main object of ctn'rent work.</Paragraph>
    <Paragraph position="4"> Another, interesting issue is how the current fi'amework can constrain acceptable multi-word variables in order to reduce computational complexity. In present, accepting or rqiecting candidate variables extracted from the sentence matching process, is based on a simple heuristic of length in content words. This type of approach would presumably require some kind of clue on what could be an acceptable translation unit pattern (Juola, P. 1994), (Furuse, O., lida, 14. 1996).</Paragraph>
    <Paragraph position="5"> Finally, future work will mainly fbcus on how the system can invoke all existing information in order to generate new translations, mainly aiming at automatic and (senti-) automatic methods for &amp;quot;recursive&amp;quot; as well as &amp;quot;parallel&amp;quot; utilization of multiple translation rules towards optimal &amp;quot;coverage&amp;quot; of new incoming sentences.</Paragraph>
  </Section>
class="xml-element"></Paper>