<?xml version="1.0" standalone="yes"?>
<Paper uid="P99-1027">
  <Title>Should we Translate the Documents or the Queries in Cross-language Information Retrieval?</Title>
  <Section position="7" start_page="212" end_page="212" type="concl">
    <SectionTitle>
5 Conclusions and Future
</SectionTitle>
    <Paragraph position="0"> Work We have performed experiments to compare query and document translation-based CLIR systems using statistical translation models that are trained identically for both translation directions. Our study is the largest comparative study of document translation and query translation of which we are aware; furthermore we have contrasted query and document translation systems on both directions within a language pair. We find no clear advantage for either the query translation system or the document translation system; instead French=eeEnglish translation appears advantageous over English~French translation, in spite of identical procedures used in constructing both. However a hybrid system incorporating both directions of translation outperforms either. Furthermore, by incorporating human query translations rather than machine translations, we show that the hybrid system continues to outperform query translation. We have based our conclusions by comparing TREC-style average precisions of retrieval with a two-pass IR system; the same conclusions follow if we instead compare precisions at rank 20 or average precisions from first pass (Okapi) scores. Thus we conclude that even in the limit of extremely high quality query translation, it will remain advantageous to incorporate both document and query translation into a CLIR system. Future work will involve investigating translation direction differences in retrieval performance for other language pairs, and for statistical translation systems trained from comparable, rather than parallel corpora.</Paragraph>
  </Section>
class="xml-element"></Paper>