<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-1039">
  <Title>Multi-Document Summarization of Evaluative Text</Title>
  <Section position="2" start_page="0" end_page="305" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Many organizations are faced with the challenge ofsummarizing large corpora oftextdata. Oneimportant application is evaluative text, i.e. any document expressing an evaluation of an entity as either positive or negative. For example, many websites collect large quantities of online customer reviews of consumer electronics. Summaries of this literature could be of great strategic value to product designers, planners and manufacturers. There are other equally important commercial applications, suchasthesummarization oftravel logs, and non-commercial applications, such as the summarization of candidate reviews.</Paragraph>
    <Paragraph position="1"> The general problem we consider in this paper is how to effectively summarize a large corpora of evaluative text about a single entity (e.g., a product). In contrast, most previous work on multi-document summarization has focused on factual text (e.g., news (McKeown et al., 2002), biographies (Zhou et al., 2004)). For factual documents, the goal of a summarizer is to select the most important facts and present them in a sensible ordering while avoiding repetition. Previous work has shown that this can be effectively achieved by carefully extracting and ordering the most informative sentences from the original documents in a domain-independent way. Notice however that when the source documents are assumed to contain inconsistent information (e.g., conflicting reports of a natural disaster (White et al., 2002)), a different approach is needed. The summarizer needs first to extract the information from the documents, then process such information to identify overlaps and inconsistencies between the different sources and finally produce a summary that points out and explain those inconsistencies.</Paragraph>
    <Paragraph position="2"> A corpus of evaluative text typically contains a large number of possibly inconsistent 'facts' (i.e.</Paragraph>
    <Paragraph position="3"> opinions), as opinions on the same entity feature may be uniform or varied. Thus, summarizing a corpus of evaluative text is much more similar to summarizing conflicting reports than a consistent set of factual documents. When there are diverse opinions on the same issue, the different perspectives need to be included in the summary.</Paragraph>
    <Paragraph position="4"> Based on this observation, we argue that any strategy to effectively summarize evaluative text about a single entity should rely on a preliminary phase of information extraction from the target corpus. In particular, the summarizer should at least know for each document: what features of the entity were evaluated, the polarity of the evaluations and their strengths.</Paragraph>
    <Paragraph position="5"> Inthis paper, weexplore this hypothesis by considering two alternative approaches. First, we developed a sentence-extraction based summarizer that uses the information extracted from the corpus to select and rank sentences from the corpus.</Paragraph>
    <Paragraph position="6"> We implemented this system, called MEAD*, by  adapting MEAD (Radev et al., 2003), an open-source frameworkformulti-document summarization. Second, we developed a summarizer that produces summaries primarily by generating language from the information extracted from the corpus. We implemented this system, called the Summarizer of Evaluative Arguments (SEA), by adapting the Generator of Evaluative Arguments (GEA) (Carenini and Moore, expected 2006) a framework for generating user tailored evaluative arguments.</Paragraph>
    <Paragraph position="7"> Wehave performed anempirical formative evaluation of MEAD* and SEA in a user study. In this evaluation, we also tested the effectiveness of human generated summaries (HGS) as a topline and of summaries generated by MEAD without access to the extracted information as a baseline.</Paragraph>
    <Paragraph position="8"> The results indicate that SEA and MEAD* quantitatively perform equally well above MEAD and below HGS. Qualitatively, we find that they perform well for different but complementary reasons. While SEA appears to provide a more general overview of the source text, MEAD*seems to provide a more varied language and detail about customer opinions.</Paragraph>
  </Section>
class="xml-element"></Paper>