<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1020">
  <Title>Machine Learning for Coreference Resolution: From Local Classification to Global Ranking</Title>
  <Section position="3" start_page="157" end_page="157" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> As mentioned before, our approach differs from the standard approach primarily by (1) explicitly learning a ranker and (2) optimizing for clustering-level accuracy. In this section we will focus on discussing related work along these two dimensions.</Paragraph>
    <Paragraph position="1"> Ranking candidate partitions. Although we are not aware of any previous attempt on training a available computing resources.</Paragraph>
    <Paragraph position="2"> ranking model using global features of an NP partition, there is some related work on partition ranking where the score of a partition is computed via a heuristic function of the probabilities of its NP pairs being coreferent.2 For instance, Harabagiu et al. (2001) introduce a greedy algorithm for finding the highest-scored partition by performing a beam search in the space of possible partitions. At each step of this search process, candidate partitions are ranked based on their heuristically computed scores.</Paragraph>
    <Paragraph position="3"> Optimizing for clustering-level accuracy. Ng and Cardie (2002a) attempt to optimize their rule-based coreference classifier for clustering-level accuracy, essentially by finding a subset of the learned rules that performs the best on held-out data with respect to the target coreference scoring program.</Paragraph>
    <Paragraph position="4"> Strube and M&amp;quot;uller (2003) propose a similar idea, but aim instead at finding a subset of the available features with which the resulting coreference classifier yields the best clustering-level accuracy on held-out data. To our knowledge, our work is the first attempt to optimize a ranker for clustering-level accuracy.</Paragraph>
  </Section>
class="xml-element"></Paper>