<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2003">
  <Title>A Probabilistic Genre-Independent Model of Pronominalization</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Generating adequate referring expressions is an active research topic in Natural Language Generation.</Paragraph>
    <Paragraph position="1"> Adequate referring expressions are those that enable the user to quickly and unambiguously identify the discourse entity that the expression co-specifies with. In this paper, we concentrate on an important aspect of that question, which has received less attention than the question of anaphora resolution in discourse interpretation, i.e., when is it feasible to pronominalize? Our aim is to identify the central factors that influence pronominalization across genres. Section 2 motivates and presents the factors that were investigated in this study: distance from last mention, parallelism, ambiguity, syntactic function, agreement, sortal class, syntactic function of the antecedent and form of the antecedent. Our analyses are based on a corpus of twelve texts from four different genres with a total of more than 24,000 words and 7126 referring expressions (Section 3). The results of the statistical analyses are summarized in Section 4. There are strong statistical associations between each of the factors and pronominalization. Only when we combine them into a probabilistic model we can identify those factors whose contribution is really important, i.e. distance from last mention, agreement, and to a certain degree form of the antecedent. Since these factors can be annotated telatively cheaply, we conclude that it is possible to develop reasonable statistical pronominalization algorithms. null</Paragraph>
  </Section>
class="xml-element"></Paper>