<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1633">
  <Title>BESTCUT: A Graph Algorithm for Coreference Resolution</Title>
  <Section position="3" start_page="0" end_page="275" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Recent coreference resolution algorithms tackle the problem of identifying coreferent mentions of the same entity in text as a two step procedure: (1) a classification phase that decides whether pairs of noun phrases corefer or not; and (2) a clusterization phase that groups together all mentions that refer to the same entity. An entity is an object or a set of objects in the real world, while a mention is a textual reference to an entity1. Most of the previous coreference resolution methods have similar classification phases, implemented either as decision trees (Soon et al., 2001) or as maximum entropy classifiers (Luo et al., 2004). Moreover, these methods employ similar feature sets.</Paragraph>
    <Paragraph position="1"> The clusterization phase is different across current approaches. For example, there are several linking decisions for clusterization. (Soon et al., 2001) advocate the link-first decision, which links a mention to its closest candidate referent, while (Ng and Cardie, 2002) consider instead the link-best decision, which links a mention to its most confident 1This definition was introduced in (NIST, 2003).</Paragraph>
    <Paragraph position="2"> candidate referent. Both these clustering decisions are locally optimized. In contrast, globally optimized clustering decisions were reported in (Luo et al., 2004) and (DaumeIII and Marcu, 2005a), where all clustering possibilities are considered by searching on a Bell tree representation or by using the Learning as Search Optimization (LaSO) framework (DaumeIII and Marcu, 2005b) respectively, but the first search is partial and driven by heuristics and the second one only looks back in text. We argue that a more adequate clusterization phase for coreference resolution can be obtained by using a graph representation.</Paragraph>
    <Paragraph position="3"> In this paper we describe a novel representation of the coreference space as an undirected edge-weighted graph in which the nodes represent all the mentions from a text, whereas the edges between nodes constitute the confidence values derived from the coreference classification phase. In order to detect the entities referred in the text, we need to partition the graph such that all nodes in each subgraph refer to the same entity.</Paragraph>
    <Paragraph position="4"> We have devised a graph partitioning method for coreference resolution, called BESTCUT, which is inspired from the well-known graph-partitioning algorithm Min-Cut (Stoer and Wagner, 1994).</Paragraph>
    <Paragraph position="5"> BESTCUT has a different way of computing the cut weight than Min-Cut and a different way of stopping the cut2. Moreover, we have slightly modified the Min-Cut procedures. BESTCUT replaces the bottom-up search in a tree representation (as it was performed in (Luo et al., 2004)) with the top-down problem of obtaining the best partitioning of a graph. We start by assuming that all mentions refer to a single entity; the graph cut splits the mentions into subgraphs and the split2Whenever a graph is split in two subgraphs, as defined in (Cormen et al., 2001), a cut of the graph is produced.</Paragraph>
    <Paragraph position="6">  ting continues until each subgraph corresponds to one of the entities. The cut stopping decision has been implemented as an SVM-based classification (Cortes and Vapnik, 1995).</Paragraph>
    <Paragraph position="7"> The classification and clusterization phases assume that all mentions are detected. In order to evaluate our coreference resolution method, we have (1) implemented a mention detection procedure that has the novelty of employing information derived from the word senses of common nouns as well as selected lexico-syntactic information; and (2) used a maximum entropy model for coreference classification. The experiments conducted on MUC and ACE data indicate state-of-the-art results when compared with the methods reported in (Ng and Cardie, 2002) and (Luo et al., 2004).</Paragraph>
    <Paragraph position="8"> The remainder of the paper is organized as follows. In Section 2 we describe the coreference resolution method that uses the BESTCUT clusterization; Section 3 describes the approach we have implemented for detecting mentions in texts; Section 4 reports on the experimental results; Section 5 discusses related work; finally, Section 6 summarizes the conclusions.</Paragraph>
  </Section>
class="xml-element"></Paper>