<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1662">
  <Title>Sentence Ordering with Manifold-based Classification in Multi-Document Summarization</Title>
  <Section position="4" start_page="526" end_page="526" type="intro">
    <SectionTitle>
2. Overview
</SectionTitle>
    <Paragraph position="0"> Fig. 1 gives the overall structure of the proposed method, which includes three modules: construction of sentence networks, sentence  The first step is to build a sentence neighborhood network with weights on edges, which can serve as the basis for a Markov random walk (Tishby et al., 2000). The neighborhood is based on similarity between sentences, and weights on edges can be seen as transition probabilities for the random walk. From this network, we can derive new representations for sentences.</Paragraph>
    <Paragraph position="1"> The second step is to make a classification of sentences, with each summary sentence as a class label. Since only one labeled example exists for each class, we use a semi-supervised method based on a Markov random walk to reveal the manifold structure for the classification.</Paragraph>
    <Paragraph position="2"> The third step is to order summary sentences according to the original positions of their partners in the same class. During this process, the next selection of a sentence is based on the whole history of selection, i.e., the association of the sentence with all those already selected.</Paragraph>
  </Section>
class="xml-element"></Paper>