<?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="8" start_page="53" end_page="53" type="concl">
    <SectionTitle>
5. Conclusions
</SectionTitle>
    <Paragraph position="0"> The error rate of both Charniak's and Connexor's as reported in the literature, is approximately the same, 20%; this notwithstanding, MARS has a slightly reduced coverage when compared with JavaRAP, 96%. GuiTAR has the worst coverage, 85%. As to accuracy, none of the three algorithms overruns 50%: JavaRAP has the best score 49.9%. However GETARUNS has 63% correct score, with 90% coverage.</Paragraph>
    <Paragraph position="1"> There are at least three reasons why our system has a better performance: one is the presence of a richer functional and semantic information as explained above, which comes with augmented head-dependent structures. Second reason is the decision to split the referential process into two and treat utterance level pronominal expressions separately from discourse level ones. Third reason is the way in which discourse level anaphora resolution is  organized: our version of the Centering algorithm hinges on a record of a list of best antecedents weighted on the basis of their behaviour in History List and on their intrinsic semantic properties. These three properties of our AR algorithm can be dubbed the Knowledge Rich approach.</Paragraph>
    <Paragraph position="2"> F-measures approximates very closely what we obtained in a previous experiment: however, as a whole it is an insufficient score to insure adequate confidence in semantic substitution of anaphoric items by the head of the antecedent. Improvements need to come from parsing and the lexical component.</Paragraph>
  </Section>
class="xml-element"></Paper>