<?xml version="1.0" standalone="yes"?>
<Paper uid="N04-4009">
  <Title>Competitive Self-Trained Pronoun Interpretation</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The last several years have seen a number of feature-based systems for pronoun interpretation in which the feature weights are determined via manual experimentation or supervised learning (see Mitkov (2002) for a useful survey). Reliable estimation of the weights in both paradigms requires a substantial manually-annotated corpus of examples. In this short paper we describe a system for (third-person) pronoun interpretation that is self-trained from raw data, that is, using no annotated training data whatsoever. The result outperforms a Hobbsian baseline algorithm and is only marginally inferior (2.3%) to an essentially identical, state-of-the-art supervised model trained from a manually-annotated coreference corpus. This result leaves open the possibility that systems self-trained on very large datasets with more finely-grained features could eventually outperform supervised models that rely on manually-annotated datasets.</Paragraph>
    <Paragraph position="1"> The remainder of the paper is organized as follows. We first briefly describe the supervised system (described in more detail in Kehler et al. (2004)) to which we will compare the self-trained system. Both systems use the same learning algorithm and feature set; they differ with respect to whether the data they /Department of Linguistics.</Paragraph>
    <Paragraph position="2"> yDepartment of Computer Science and Engineering.</Paragraph>
    <Paragraph position="3"> are trained on is annotated by a human or the algorithm itself. We then describe our Hobbsian baseline algorithm, and present the results of all three systems. null</Paragraph>
  </Section>
class="xml-element"></Paper>