<?xml version="1.0" standalone="yes"?>
<Paper uid="N03-1003">
  <Title>Learning to Paraphrase: An Unsupervised Approach Using Multiple-Sequence Alignment</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 Related work
</SectionTitle>
    <Paragraph position="0"> Previous work on automated paraphrasing has considered different levels of paraphrase granularity.</Paragraph>
    <Paragraph position="1"> Learning synonyms via distributional similarity has been well-studied (Pereira et al., 1993; Grefenstette, 1994; Lin, 1998). Jacquemin (1999) and Barzilay and McKeown (2001) identify phrase-level paraphrases, while Lin and Pantel (2001) and Shinyama et al. (2002) acquire structural paraphrases encoded as templates. These latter are the most closely related to the sentence-level paraphrases we desire, and so we focus in this section on template-induction approaches.</Paragraph>
    <Paragraph position="2"> Lin and Pantel (2001) extract inference rules, which are related to paraphrases (for example, X wrote Y implies X is the author of Y), to improve question answering. They assume that paths in dependency trees that take similar arguments (leaves) are close in meaning. However, only two-argument templates are considered. Shinyama et al. (2002) also use dependency-tree information to extract templates of a limited form (in their case, determined by the underlying information extraction application). Like us (and unlike Lin and Pantel, who employ a single large corpus), they use articles written about the same event in different newspapers as data.</Paragraph>
    <Paragraph position="3"> Our approach shares two characteristics with the two methods just described: pattern comparison by analysis of the patterns' respective arguments, and use of non-parallel corpora as a data source. However, extraction methods are not easily extended to generation methods.</Paragraph>
    <Paragraph position="4"> One problem is that their templates often only match small fragments of a sentence. While this is appropriate for other applications, deciding whether to use a given template to generate a paraphrase requires information about the surrounding context provided by the entire sentence. null</Paragraph>
  </Section>
class="xml-element"></Paper>