<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-1009">
  <Title>Hybrid Text Summarization: Combining External Relevance Measures with Structural Analysis</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> Structural sentence extraction systems including Summarist and PALSUMM that create summaries by choosing sentences or parts of sentences corresponding to nodes at a given level of depth of a 6 In other applications of the algorithms described here, where the purpose is not that of retrieving a full summary of a document but rather that of building the necessary minimal context for interpreting a certain selected discourse constituent, percolation is only limited to the immediate surrounding context, where certain relations (usually ad-hoc binaries) constitute a barrier to further percolation towards upwards constituent.</Paragraph>
    <Paragraph position="1">  1. For seeding V, each leaf node l is assigned an a priori score V(l). 2. Repeat for each node c0 with children c1...cn, and relation type R until no values change: 2.1 Percolate or maintain highest score: V(c0) := maxi (V(ci)) , 0=i=n 2.2 Percolate highest score downwards into non-subordinated nodes:  if R is subordination and ci is the head of n: V(ci) := V(c0) if V(ci) &lt; V(c0) if R is coordination or n-aries: for all i=n, V(ci) := V(c0) if V(ci) &lt; V(c0),  (V=T) are percolated according to this algorithm, resulting in values S(n) and T(n) for nodes n. tree structured representation of the structure of the text produce excellent summaries that preserve the style and &amp;quot;flavor&amp;quot; of the original text. However, the summaries constructed may be longer than needed, including information that could be omitted without serious loss of informativity7. The excessive length results from the top-down nature of standard structural extraction algorithms which start by choosing the top context and then includes every possible sub-context down to a certain level.</Paragraph>
    <Paragraph position="2"> In this paper, we have proposed hybrid algorithms which capitalize on the strengths of these methods while compensating for their limitations by proposing additional manipulations on the base trees. In our view, the value of the summarization methods described here, is the ability to compress a summary further without substantia l loss of informativity. For summaries, especially those designed for display on various sized devices, the work presented here constitutes an advance in the state of the art.</Paragraph>
  </Section>
class="xml-element"></Paper>