<?xml version="1.0" standalone="yes"?>
<Paper uid="P02-1058">
  <Title>From Single to Multi-document Summarization: A Prototype System and its Evaluation</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> We described a multi-document summarization system, NeATS, and its evaluation in DUC-2001. We were encouraged by the content and readability of the results. As a prototype system, NeATS deliberately used simple methods guided by a few principles: * Extracting important concepts based on reliable statistics.</Paragraph>
    <Paragraph position="1"> * Filtering sentences by their positions and stigma words.</Paragraph>
    <Paragraph position="2"> * Reducing redundancy using MMR.</Paragraph>
    <Paragraph position="3"> * Presenting summary sentences in their chronological order with time annotations. These simple principles worked effectively. However, the simplicity of the system also lends itself to further improvements. We would like to apply some compression techniques or use linguistic units smaller than sentences to improve our retention score. The fact that NeATS performed as well as the human in pseudo precision but did less well in retention indicates its summaries might include good but duplicated information. Working with sub-sentence units should help.</Paragraph>
    <Paragraph position="4"> To improve NeATS's capability in content selection, we have started to parse sentences containing key unigram, bigram, and trigram concepts to identify their relations within their concept clusters.</Paragraph>
    <Paragraph position="5"> To enhance cohesion and coherence, we are looking into incorporating discourse processing techniques (Marcu 1999) or Radev and McKeown's (1998) summary operators.</Paragraph>
    <Paragraph position="6"> We are analyzing the DUC evaluation scores in the hope of suggesting improved and more stable metrics.</Paragraph>
  </Section>
class="xml-element"></Paper>