<?xml version="1.0" standalone="yes"?>
<Paper uid="P99-1071">
  <Title>Information Fusion in the Context of Multi-Document Summarization</Title>
  <Section position="7" start_page="555" end_page="555" type="concl">
    <SectionTitle>
6 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> In this paper, we presented an implemented algorithm for multi-document summarization which moves beyond the sentence extraction paradigm. Assuming a set of similar sentences as input extracted from multiple documents on the same event (McKeown et al., 1999; Eskin et al., 1999), our system identifies common phrases across sentences and uses language generation to reformulate them as a coherent summary.</Paragraph>
    <Paragraph position="1"> The use of generation to merge similar information is a new approach that significantly improves the quality of the resulting summaries, reducing repetition and increasing fluency.</Paragraph>
    <Paragraph position="2"> The system we have developed serves as a point of departure for research in a variety of directions. First is the need to use learning techniques to identify paraphrasing patterns in corpus data. As a first pass, we found paraphrasing rules manually. This initial set might allow us to automatically identify more rules and increase the performance of our comparison algorithm.</Paragraph>
    <Paragraph position="3"> From the generation side, our main goal is to make the generated summary more concise, primarily by combining clauses together. We will be investigating what factors influence the combination process and how they can be computed from input articles. Part of combination will involve increasing coherence of the generated text through the use of connectives, anaphora or lexical relations (Jing, 1999).</Paragraph>
    <Paragraph position="4"> One interesting problem for future work is the question of how much context to include from a sentence from which an intersected phrase is drawn. Currently, we include no context, but in some cases context is crucial even though it is not a part of the intersection. This is the case, for example, when the context negates, or denies, the embedded sub-clause which matches a sub-clause in another negating context. In such cases, the resulting summary is actually false. This occurs just once in our test cases, but it is a serious error. Our work will characterize the types of contextual information that should be retained and will develop algorithms for the case of negation, among others.</Paragraph>
  </Section>
class="xml-element"></Paper>