<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1605">
  <Title>Generating and selecting grammatical paraphrases</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> As is well known, natural language has a very high paraphrastic power so that the same core meaning can be expressed in many different ways [Gross, 1975; Mel'Vcuk, 1988]. Yet not all paraphrases are appropriate for all contexts. So for instance, a sentence and its converse (1a) express the same core meaning and so can be considered paraphrases of each other.</Paragraph>
    <Paragraph position="1"> Yet as example (1b) illustrates, they are not interchangeable in the context of a control verb: (1) a. John borrowed a book from Mary.</Paragraph>
    <Paragraph position="2"> [?]Mary lent a book to John b. Peter persuaded John to borrow a book from Mary.</Paragraph>
    <Paragraph position="3"> negationslash[?]Peter persuaded Mary to lend a book to John Similarly, a canonical and a cleft sentence (2a) communicate the same core meaning yet a contrastive context (2b) only admits the cleft version.</Paragraph>
    <Paragraph position="4"> (2) a. John looks at Mary.</Paragraph>
    <Paragraph position="5"> [?]It is Mary that John looks at b. * It is not Sarah, John looks at Mary.</Paragraph>
    <Paragraph position="6"> It is not Sarah, it is Mary that John looks at More generally, the anaphoric potential (that is, the discourse status of the entities being talked about) of the preceding discourse, its structure, the presence of an embedding verb or of a given subordinating or coordinating conjunction are all factors which may restrict the use of paraphrases. To preserve completeness, it is therefore important that a generator be able to produce paraphrases in a systematic fashion. On the other hand, it is also well known that surface realisation (the task of producing the set of sentences associated by a grammar with a given semantic representation) is NP-complete [Brew, 1992].</Paragraph>
    <Paragraph position="7"> In this paper, we present a TAG based surface realiser which supports both the generation and the selection of grammatical paraphrases (section 2 and 3). To deal with the resulting combinatorics, we introduce a number of new optimisations (section 4). We then show how one of these optimisations can be used to support the selection of contextually appropriate paraphrases (section 5). Finally, we relate our approach to similar proposals and show that it compares favourably in terms of efficiency (section 6 and 7).</Paragraph>
  </Section>
class="xml-element"></Paper>