<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1079">
  <Title>Generating Overview Summaries of Ongoing Email Thread Discussions</Title>
  <Section position="4" start_page="0" end_page="0" type="relat">
    <SectionTitle>
3 Related Work
</SectionTitle>
    <Paragraph position="0"> To date, email thread summarization has not been explored in any great depth within the Natural Language Processing (NLP) research community.</Paragraph>
    <Paragraph position="1"> Research on thread summarization has included some work on using dialogue structure for email summarization. Nenkova et al. (2003) advocate the use of overview sentences similar to ours.</Paragraph>
    <Paragraph position="2"> They extract sentences based on the presence of subject line key words. However, should the subject line not reflect the content of the thread, our method has the potential to extract the true discussion issue since it based on the responses of other participants.</Paragraph>
    <Paragraph position="3"> Lam et al. (2002) use the context of the preceding thread to provide background information for email summaries. However, they note that even after appropriate preprocessing of email text, simply concatenating preceding context can lead to long summaries. In contrast, instead of extracting email texts verbatim, we extract single sentences from particular emails in the thread. As a result, our summaries tend to be much shorter.</Paragraph>
    <Paragraph position="4"> Murakoshi et al. (1999) describe an approach which extracts question-answer pairs from an email thread. Extraction is based on the use of pattern-based information extraction methods. The summary thus provides the question-answer pair intact, thereby improving the coherence. Question-answer summaries would presumably be suited to discussions which support an information provision task, a complementary task to the one we examine.</Paragraph>
    <Paragraph position="5"> Rambow et al. (2004) apply sentence extraction techniques to the thread to construct a generic summary. Though not specifically using dialogue structure, one feature used marks if a sentence is a question or not.</Paragraph>
    <Paragraph position="6"> Work has also been done on more accurately constructing the dialogue structure. Newman and Blitzer (2003) focus on clustering related newsgroup messages into dialogue segments. The segments are then linked using email header information to form a hierarchical structure. Their summary is simply the first sentence from each segment. We envisage dialogue structure summaries showing an overview of topics would be combined with approaches such as ours which provide summaries of segments.</Paragraph>
    <Paragraph position="7"> We also note the existing work that explores the summarization of speech transcripts. Speech is a very different mode of communication. An overview of the differences between asynchronous and synchronous modes of communication is provided by Clark (1991) and Simpson-Young et al. (2000). Alexandersson et al. (2000) note that in speech there is a tendency not to repeat shared conversation context. They use the preceding dialogue structure, modeled using Dialogue Representation Theory, to provide additional ellipsed information. It is unclear how such an approach might apply to an email corpus which has the potential to cover a broader set of domains.</Paragraph>
    <Paragraph position="8"> More recently, Zechner and Lavie (2001) identify question-answer dialogue segments in order to extract the pair as a whole.</Paragraph>
    <Paragraph position="9"> Hillard et al. (2003) have also produced a system which generates summaries of speech discussions supporting a decision-making process. Their work differs from ours in that they focus on categorizing the polarity of responses in order to summarize consensus.</Paragraph>
  </Section>
class="xml-element"></Paper>