<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-5012">
  <Title>a1 Information and Communication Technologies</Title>
  <Section position="2" start_page="0" end_page="88" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Human authored summaries are more than just a list of extracted sentences. Often the summary sentence is a paraphrase of a sentence in the source text, or else a combination of phrases and words from important sentences that have been pieced together to form a new sentence. These sentences, referred to as Non-Verbatim Sentences, can replace extracted text to improve readability and coherence in the summary.</Paragraph>
    <Paragraph position="1"> Consider the example in Figure 1 which presents an alignment between a human authored summary sentence and a source sentence. The Summary Sentence: Every province in the country, except one, endured sporadic fighting, looting or armed banditry in 2003.</Paragraph>
    <Section position="1" start_page="0" end_page="88" type="sub_section">
      <SectionTitle>
Source Sentence:
</SectionTitle>
      <Paragraph position="0"> However, as the year unfolded, every province has been subjected to fighting, looting or armed banditry, with the exception of just one province (Kirundo, in northern Burundi).</Paragraph>
      <Paragraph position="1">  text is taken from a corpus of Humanitarian Aid Proposals1 produced by the United Nations for the purpose of convincing donors to support a relief effort.</Paragraph>
      <Paragraph position="2"> Theexample illustrates that sentence extraction alone cannot account forthebreadth of humanauthored summary sentences. This is supported by evidence presented in (Jing and McKeown, 1999) and (Daum'e III and Marcu, 2004).</Paragraph>
      <Paragraph position="3"> Moving towards the goal of abstract-like automatic summary generation challenges us to consider mechanisms for generating non-verbatim sentences. Such a mechanism can usefully be considered as automatically generating a paraphrase.2 We treat the problem as one in which a new and previously unseen summary sentence is to be automatically produced given some closely related sentences extracted from a source text.</Paragraph>
      <Paragraph position="4"> Following on from (Witbrock and Mittal, 1999), we use and extend the Viterbi algorithm (Forney, 1973) for the purposes of generating non-verbatim sentences. This approach treats  sentence generation as a search problem. Given a set of words (taken from some set of sentences to paraphrase), we search for the most likely sequence given some language model. Intuitively, we want the generated string to be grammatical and to accurately reflect the content of the source text.</Paragraph>
      <Paragraph position="5"> Within the Viterbi search process, each timewe append a word tothe partially generated sentence, we consider how well it attaches to a dependency structure. The focus of this paper is to evaluate whether or not a series of iterative considerations of dependency structure results in a grammatical generated sentence. Previous preliminary evaluations (Wan et al., 2005) indicate that the generated sequences contain less fragmented text as measured by an off-the-shelf dependency parser; more fragments would indicate a grammatically problematic sentence.</Paragraph>
      <Paragraph position="6"> However, while encouraging, such an evaluation says little about what the actual sentence looks like. For example, such generated text might only be useful if it contains complete clauses. Thus, in this paper, we use the precision and recall metric to measure how many generated verb arguments, as extracted from dependency relations, are correct.</Paragraph>
      <Paragraph position="7"> Theremainder of this paper is structured as follows. Section 2 provides an overview introducing our approach. In Section 3, we briefly illustrate our algorithm with examples. A brief survey of related work is presented in Section 4. Wepresent our grammaticality experiments in Section 5. We conclude with further work in Section 6.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>