<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-1029">
  <Title>Compiling French-Japanese Terminologies from the Web</Title>
  <Section position="8" start_page="2" end_page="2" type="concl">
    <SectionTitle>
5 Conclusion and future work
</SectionTitle>
    <Paragraph position="0"> We have proposed a method for compiling bilingual terminologies of compositionally translated MWTs. As opposed to previous work, we use the web rather than comparable corpora as a source of bilingual data. Our main insight is to constrain source and target candidate MWTs to only those strongly related to the seed. This allows us to achieve term alignment with high precision. We showed that coverage reaches satisfactory levels by using thesauri and bootstrapping.</Paragraph>
    <Paragraph position="1"> Due to the difference in objectives and in corpora, it is very hard to compare results: our method produces a rather small set of highly accurate alignments, whereas extraction from comparable corpora generates much more candidates, but with an inferior precision. These two approaches have very different applications. Our method does however eliminate the requirement of comparable corpora, which means that we can use seeds from any domain, provided we have reasonably rich dictionaries and thesauri.</Paragraph>
    <Paragraph position="2"> Let us not forget that this article describes only a first attempt at compiling French-Japanese terminology, and that various sources of improvement have been left untapped. In particular, our alignment suffers from the fact that we do not discriminate between different candidate translations. This could be achieved by using any of the more sophisticated selection methods proposed in the literature. Currently, corpus features are used solely for the collection of related terms.</Paragraph>
    <Paragraph position="3"> These could also be utilized in the translation selection, which Baldwin and Tanaka have shown to be quite effective. We could also make use of bilingual dictionary features as they did.</Paragraph>
    <Paragraph position="4"> Lexical context is another resource we have not exploited. Context vectors have successfully been applied in translation selection by Fung as well as Daille and Morin.</Paragraph>
    <Paragraph position="5"> On a different level, we could also apply the bootstrapping to expand the French set of related terms. Finally, we are investigating the possibil- null ity of resolving the alignments in the opposite direction: from Japanese to French. Surely the constructional variability of French MWTs would present some difficulties, but we are confident that this could be tackled using translation templates, as proposed by Baldwin and Tanaka.</Paragraph>
  </Section>
class="xml-element"></Paper>