<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1662">
  <Title>Sentence Ordering with Manifold-based Classification in Multi-Document Summarization</Title>
  <Section position="9" start_page="532" end_page="532" type="concl">
    <SectionTitle>
6. Conclusion and Future Work
</SectionTitle>
    <Paragraph position="0"> In this paper, we propose a sentence ordering method for multi-document summarization based on semi-supervised classification and historical ordering. For sentence classification, the semi-supervised classification groups sentences based on their global distribution, rather than on local comparisons. Thus, even with a small amount of labeled data (just 1 labeled example in our case) we nevertheless ensure good performance for sentence classification.</Paragraph>
    <Paragraph position="1"> For sentence ordering, we propose a kind of history-based ordering strategy, which determines the next selection based on the whole selection history, rather than the most recent single selection in probabilistic ordering, which could result in topic bias, or in-out difference in MO, which could result in topic disruption.</Paragraph>
    <Paragraph position="2"> In this work, we mainly use sentence-level information, including sentence similarity and sentence order, etc. In future, we may explore the role of term-level or word-level features, e.g., proper nouns, in the ordering of summary sentences. To make summaries more coherent and readable, we may also need to discover how to detect and control topic movement automatic summaries. One specific task is how to generate co-reference among sentences in summaries. In addition, we will also try other semi-supervised classification methods, and other evaluation metrics, etc.</Paragraph>
  </Section>
class="xml-element"></Paper>