<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1304">
  <Title>I</Title>
  <Section position="10" start_page="75" end_page="75" type="evalu">
    <SectionTitle>
9 Parsing
</SectionTitle>
    <Paragraph position="0"> We have now dealt with all parts of the parsing process. Whenever a new word is seen, a few tags are selected according to (13). After this a set of about 300 (depending on the confidence in the tags) traversal strings is selected according to (14) or (15). The forward probability of these candidates is calculated (16) and this is used to further reduce the candidates to 15 tag-traversal string pairs. This set is saved with their forward probabilities, and when the end-of-sentence signal is received the best series is given by the Viterbi algorithm. A tree is then produced according to the algorithm described in section 4.</Paragraph>
    <Paragraph position="1"> Until now we h~ve set the maximum traversal string length to 5 but now we can show how variation in the maximum length affects the result. The experiments we present here were carried out with data from the Wall Street Journal Treebank. Parameters were estimated with the first 22 sections (over 40,000 sentences), section 24 was used for smoothing (interpolation)  and section 23 (2,416 sentences) was used exclusively for testing. Figure 8 shows the labeled accuracy and recall for various maximum lengths that result from this data.</Paragraph>
    <Paragraph position="2"> This shows that the optimal length is about 4, 5 or perhaps 6. This picture is slightly favoring the shorter lengths, since # and ~ are fixed while the longer lengths have more candidates to choose from. But on the other handl keeping p and ~7 fixed corresponds to giving the algorithm a certain time and letting it do its best in the given time* The longer lengths also have a disadvantage in that they lead to larger tables, using more memory.</Paragraph>
    <Paragraph position="3"> The differences between 4, 5 and 6 axe minor, and the performance degrades seriously at 3 or 7. This shows that a maximum of 5 is a sensible choice. The first colnm, of table 2 gives detailed information about the final performance. It is also possible to restrict the parser to lower level structures, t~iclng only those parts which are the most safe, namely low-level brackets that do not depend on long distance dependencies. We carried this out by removing brackets covering more than three words and some particular nonterminals that often result in errors, such as SBAR~ These results axe indicated in the ~Shallow Parsing z col-re,</Paragraph>
  </Section>
class="xml-element"></Paper>