File Information

File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/intro/98/w98-1304_intro.xml

Size: 2,028 bytes

Last Modified: 2025-10-06 14:06:52

<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1304">
  <Title>I</Title>
  <Section position="4" start_page="38" end_page="38" type="intro">
    <SectionTitle>
3 Parsing with Traversal Strings
</SectionTitle>
    <Paragraph position="0"> Figure 2 shows the system design. When presented with a sentence, the first component finds the most likely set of part of speech tags (not shown) and traversal strings matching the words in the sentence. After this a second component assembles a tree from the traversal strings.</Paragraph>
    <Paragraph position="1"> l predict s p /P ~ \[ combine l am singing.., traversal Iamsm.~...J strings into 1.. Iamsmghlg...</Paragraph>
    <Paragraph position="2"> -\[ strings 1 tree !  The prediction of traversal strings presents us with a problem, since traversal strings of arbitrary length are too numerous to be predicted accurately. To see our answer to this problem it is instructive to look at the analysis of a longer sentence, please see figure 3. In particular, notice the shaded areas that show how traversal strings of neighboring words are often equal or partially equal. The most common relation that is seen is a 'shift', where the vertex at position n for word wi becomes vertex n + 1 for word wi+l. At the bottom of the traversal string one vertex is added, and nothing else changes.</Paragraph>
    <Paragraph position="3"> This illustrates the next step we will take. Even after cutting off traversal strings at a fixed maximum length, it is still possible to reconstruct the tree. The dotted line in figure 3 shows how traversal strings are cut off at a maximum length of 5 vertices. Having part of the traversal string still leaves it possible to see that a particular word is likely to be in the same context as his neighbor. More generally, we look at what subtrees are likely to share part of their context with other, neighboring subtrees. We will show how doing this iteratively makes it possible to restore the tree with a high degree of accuracy.</Paragraph>
  </Section>
class="xml-element"></Paper>
Download Original XML