<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-2302">
  <Title>Another Evaluation of Anaphora Resolution Algorithms and a Comparison with GETARUNS' Knowledge Rich Approach</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The problem of anaphora resolution (hence AR) looms more and more as a prominent one in unrestricted text processing due to the need to recover semantically consistent information in most current NLP applications. This problem does not lend itself easily to a statistical approach so that rule-based approaches seem the only viable solution.</Paragraph>
    <Paragraph position="1"> We present a new evaluation of three state-of-the-art algorithms for anaphora resolution - GuiTAR, JavaRAP, MARS - on the basis of a portion of Susan Corpus (derived from Brown Corpus) a much richer testbed than the ones previously used for evaluation, and in any case a much more comparable source with such texts as newspaper articles and stories. Texts used previously ranged from scientific manuals to descriptive scientific texts and were generally poor on pronouns and rich on nominal descriptions. Two of the algorithms GuiTAR and JavaRAP - use Charniak's parser output, which contributes to the homogeneity of the type of knowledge passed to the resolution procedure. MARS, on the contrary, uses a more sophisticated input, the one provided by Connexor FDG-parser. The algorithms will then be compared to our system, GETARUNS, which incorporated an AR algorithm at the end of a pipeline of interconnected modules that instantiate standard architectures for NLP. The version of the algorithm presented here is a newly elaborated one, and is devoted to unrestricted text processing. It is an upgraded version from the one discussed in Delmonte (1999;2002a;2002b) and tries to incorporate as much as possible of the more sophisticated version implemented in the complete GETARUN (see Delmonte 1990;1991;1992;1994; 2003;2004).</Paragraph>
    <Paragraph position="2"> The paper is organized as follows: in section 2 below we briefly discuss architectures and criteria for AR of the three algorithms evaluated. In section 3 we present our system. Section 4 is dedicated to a compared evaluation and a general discussion.</Paragraph>
  </Section>
class="xml-element"></Paper>