<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1640">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics Partially Supervised Coreference Resolution for Opinion Summarization through Structured Rule Learning</Title>
  <Section position="4" start_page="336" end_page="337" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Work relevant to our problem can be split into three main areas - sentiment analysis, traditional noun phrase coreference resolution, and supervised and weakly supervised clustering. Related work in the former two areas is summarized briefly below. Supervised and weakly supervised clustering approaches are discussed in Section 4.</Paragraph>
    <Paragraph position="1"> Sentiment analysis. Much of the relevant research in sentiment analysis addresses sentiment classification, a text categorization task of extracting opinion at the coarse-grained document level. The goal in sentiment classification is to assign to [Source Zacarias Moussaoui] [[?] complained] at length today about [Target his own lawyer], telling a federal court jury that [Target he] was [[?] more interested in achieving fame than saving Moussaoui's life].</Paragraph>
    <Paragraph position="2"> Mr. Moussaoui said he was appearing on the witness stand to tell the truth. And one part of the truth, [Source he] said, is that [Target sending him to prison for life] would be &amp;quot;[[?] a greater punishment] than being sentenced to death.&amp;quot; &amp;quot;[[?] [Target You] have put your interest ahead of [Source my] life],&amp;quot; [Source Mr. Moussaoui] told his court-appointed lawyer Gerald T. Zerkin.</Paragraph>
    <Paragraph position="3"> ...</Paragraph>
    <Paragraph position="4"> But, [Source Mr. Zerkin] pressed [Target Mr. Moussaoui], was it [[?] not true] that he told his lawyers earlier not to involve any Muslims in the defense, not to present any evidence that might persuade the jurors to spare his life? ...</Paragraph>
    <Paragraph position="5"> [Source Zerkin] seemed to be trying to show the jurors that while [Target the defendant] is generally [+ an honest individual], his conduct shows [Target he] is [[?] not stable mentally], and thus [[?] undeserving] of [Target the ultimate punishment].</Paragraph>
    <Paragraph position="6">  eted; opinion expressions are shown in italics and bracketed with associated polarity, either positive (+) or negative (-). The underlined phrase will be explained later in the paper.</Paragraph>
    <Paragraph position="7"> a document either positive (&amp;quot;thumbs up&amp;quot;) or negative (&amp;quot;thumbs down&amp;quot;) polarity (e.g. Das and Chen (2001), Pang et al. (2002), Turney (2002), Dave et al. (2003)). Other research has concentrated on analyzing fine-grained opinions at, or below, the sentence level. Recent work, for example, indicates that systems can be trained to recognize opinions and their polarity, strength, and sources to a reasonable degree of accuracy (e.g. Dave et al. (2003), Riloff and Wiebe (2003), Bethard et al. (2004), Wilson et al. (2004), Yu and Hatzivassiloglou (2003), Choi et al. (2005), Kim and Hovy (2005), Wiebe and Riloff (2005)). Our work extends research on fine-grained opinion extraction by augmenting the opinions with additional information that allows the creation of concise opinion summaries. In contrast to the opinion extracts produced by Pang and Lee (2004), our summaries are not text extracts, but rather explicitly identify and  characterize the relations between opinions and their sources.</Paragraph>
    <Paragraph position="8"> Coreference resolution. Coreference resolution is a relatively well studied NLP problem (e.g.</Paragraph>
    <Paragraph position="9"> Morton (2000), Ng and Cardie (2002), Iida et al.</Paragraph>
    <Paragraph position="10"> (2003), McCallum and Wellner (2003)). Coreference resolution is defined as the problem of deciding which noun phrases in the text (mentions) refer to the same real world entities (are coreferent). Generally, successful approaches to coreference resolution have relied on supervised classification followed by clustering. For supervised classification these approaches learn a pairwise function to predict whether a pair of noun phrases is coreferent. Subsequently, when making coreference resolution decisions on unseen documents, the learnt pairwise NP coreference classifier is run, followed by a clustering step to produce the final clusters (coreference chains) of coreferent NPs. For both training and testing, coreference resolution algorithms rely on feature vectors for pairs of noun phrases that encode linguistic information about the NPs and their local context. Our general approach to source coreference resolution is inspired by the state-of-the-art performance of one such approach to coreference resolution, which relies on a rule learner and single-link clustering as described in Ng and Cardie (2002).</Paragraph>
  </Section>
class="xml-element"></Paper>