<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0835">
  <Title>A Recursive Statistical Translation Model[?]</Title>
  <Section position="10" start_page="206" end_page="206" type="evalu">
    <SectionTitle>
10 Results and discussion
</SectionTitle>
    <Paragraph position="0"> The results over the test sets can be seen in Table 5.</Paragraph>
    <Paragraph position="1"> It can be seen that, except for French, the influence of the initial alignment is very small. Also, the best results are obtained for Spanish and French, which are more similar to English that German or Finnish.</Paragraph>
    <Paragraph position="2"> There are still many open questions that deserve more experimentation. The first is the influence of the split of the original corpora. Although the similarity of results seem to indicate that it has little influence, this has to be tested. Two more relevant aspects are whether the weighting schema is the best for the decoder. In particular, it is surprising that the normalization of counts had so little effect.</Paragraph>
    <Paragraph position="3"> Finally, the average number of words per template is below two, which probably is too low. It is interesting to find alternate ways of obtaining the templates, for instance using internal nodes up to a given height or covering portions of the sentences up to a predefined number of words.</Paragraph>
  </Section>
class="xml-element"></Paper>