<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-0302">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics Toward Opinion Summarization: Linking the Sources</Title>
  <Section position="3" start_page="0" end_page="9" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Sentiment analysis is concerned with the extraction and representation of attitudes, evaluations, opinions, and sentiment from text. The area of sentiment analysis has been the subject of much recent research interest driven by two primary motivations. First, there is a desire to provide applications that can extract, represent, and allow the exploration of opinions in the commercial, government, and political domains. Second, effective sentiment analysis might be used to enhance and improve existing NLP applications such as information extraction, question answering, summarization, and clustering (e.g. Riloff et al. (2005), Stoyanov et al. (2005)).</Paragraph>
    <Paragraph position="1"> Several research efforts (e.g. Riloff and Wiebe (2003), Bethard et al. (2004), Wilson et al. (2004), Yu and Hatzivassiloglou (2003), Wiebe and Riloff (2005)) have shown that sentiment information can be extracted at the sentence, clause, or individualopinionexpressionlevel(fine-grainedopin- null ion information). However, little has been done to develop methods for combining fine-grained opinion information to form a summary representation in which expressions of opinions from the same source/target1 are grouped together, multiple opinions from a source toward the same target are accumulated into an aggregated opinion, and cumulative statistics are computed for each source/target. A simple opinion summary2 is shown in Figure 1. Being able to create opinion summaries is important both for stand-alone applications of sentiment analysis as well as for the potentialusesofsentimentanalysisaspartofother NLP applications.</Paragraph>
    <Paragraph position="2"> In this work we address the dearth of approaches for summarizing opinion information.</Paragraph>
    <Paragraph position="3"> In particular, we focus on the problem of source coreference resolution, i.e. deciding which source mentions are associated with opinions that belong to the same real-world entity. In the example from Figure 1 performing source coreference resolution amounts to determining that Stanishev, he, and he refer to the same real-world entities. Given the associated opinion expressions and their polarity, this source coreference information is the critical  knowledgeneededtoproducethesummaryofFigure 1 (although the two target mentions, Bulgaria and our country, would also need to be identified as coreferent).</Paragraph>
    <Paragraph position="4"> Our work is concerned with fine-grained expressions of opinions and assumes that a system can rely on the results of effective opinion and source extractors such as those described in Riloff and Wiebe (2003), Bethard et al. (2004), Wiebe andRiloff(2005)andChoietal.(2005). Presented with sources of opinions, we approach the problem of source coreference resolution as the closely  &amp;quot; [Target Delaying of Bulgaria's accession to the EU] would be a serious mistake&amp;quot; [Source Bulgarian Prime Minister Sergey Stanishev] said in an interview for the German daily Suddeutsche Zeitung. &amp;quot;[Target Our country] serves as a model and encourages countries from the region to follow despite the difficulties&amp;quot;, [Source he] added.</Paragraph>
    <Paragraph position="5"> [Target Bulgaria] is criticized by [Source the EU] because of slow reforms in the judiciary branch, the newspaper notes. Stanishev was elected prime minister in 2005. Since then, [Source he] has been a prominent supporter of [Target his country's accession to the EU].</Paragraph>
    <Paragraph position="6">  (above) and a summary of the opinions (below).</Paragraph>
    <Paragraph position="7"> In the text, sources and targets of opinions are marked and opinion expressions are shown in italic. In the summary graph, + stands for positive opinion and - for negative.</Paragraph>
    <Paragraph position="8"> related task of noun phrase coreference resolution. However, source coreference resolution differsfromtraditional nounphrase(NP)coreference resolution in two important aspects discussed in Section4. Nevertheless, asafirstattemptatsource coreference resolution, we employ a state-of-the-art machine learning approach to NP coreference resolution developed by Ng and Cardie (2002).</Paragraph>
    <Paragraph position="9"> Using a corpus of manually annotated opinions, we perform an extensive evaluation and obtain strong initial results for the task of source coreference resolution.</Paragraph>
  </Section>
class="xml-element"></Paper>