<?xml version="1.0" standalone="yes"?>
<Paper uid="H01-1033">
  <Title>Improved Cross-Language Retrieval using Backoff Translation</Title>
  <Section position="5" start_page="3" end_page="3" type="evalu">
    <SectionTitle>
4. RESULTS
</SectionTitle>
    <Paragraph position="0"> Table 1 summarizes our results. Increasing thresholds seem to be helpful with the STRAND tralex, although the differences were not found to be statistically significant by a paired two-tailed t-test with p&lt;0:05. Merging the tralexes provided no improvement over using the WebDict tralex alone, but our backoff strategy produced a statistically significant 12% improvement in mean average precision (at p&lt;0:01) over the next best tralex (WebDict alone).</Paragraph>
    <Paragraph position="1"> As Figure 1 shows, the improvement is remarkably consistent, with only four of the 34 topics adverselyaffected and only one topic showing a substantial negative impact.</Paragraph>
    <Paragraph position="2"> Breaking down the backoff results by stage (Table 2), we find that the majority of query-to-document hits are obtained in the first stage, i.e. matches of the term's surface form in the document to a translation of the surface form in the dictionary. However, the back-off process improves by-token coverage of terms in documents by 8%, and gives a 3% relative improvement in retrieval results; it also contributed additional translations to the top-2 set in approximately 30% of the cases, leading to the statistically significant 12% relative improvement in mean averageprecision as compared to the baseline using WebDict alone with 4-stage backoff.</Paragraph>
  </Section>
class="xml-element"></Paper>