<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3219">
  <Title>Monolingual Machine Translation for Paraphrase Generation</Title>
  <Section position="3" start_page="1" end_page="1" type="intro">
    <SectionTitle>
2 Related work
</SectionTitle>
    <Paragraph position="0"> Until recently, efforts in paraphrase were not strongly focused on generation and relied primarily on narrow data sources. One data source has been multiple translations of classic literary works (Barzilay &amp; McKeown 2001; Ibrahim 2002; Ibrahim et al. 2003). Pang et al. (2003) obtain parallel mono-lingual texts from a set of 100 multiply-translated news articles. While translation-based approaches to obtaining data do address the problem of how to identify two strings as meaning the same thing, they are limited in scalability owing to the difficulty (and expense) of obtaining large quantities of multiply-translated source documents.</Paragraph>
    <Paragraph position="1"> Other researchers have sought to identify patterns in large unannotated monolingual corpora.</Paragraph>
    <Paragraph position="2"> Lin &amp; Pantel (2002) derive inference rules by parsing text fragments and extracting semantically similar paths. Shinyama et al. (2002) identify dependency paths in two collections of newspaper articles. In each case, however, the information extracted is limited to a small set of patterns.</Paragraph>
    <Paragraph position="3"> Barzilay &amp; Lee (2003) exploit the meta-information implicit in dual collections of news- null Barzilay &amp; McKeown (2001), for example, reject the idea owing to the noisy, comparable nature of their data. wire articles, but focus on learning sentence-level patterns that provide a basis for generation. Multisequence alignment (MSA) is used to identify sentences that share formal (and presumably semantic) properties. This yields a set of clusters, each characterized by a word lattice that captures n-gram-based structural similarities between sentences. Lattices are in turn mapped to templates that can be used to produce novel transforms of input sentences. Their methodology provides striking results within a limited domain characterized by a high frequency of stereotypical sentence types. However, as we show below, the approach may be of limited generality, even within the training domain.</Paragraph>
  </Section>
class="xml-element"></Paper>