<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2023">
  <Title>Summarizing Speech Without Text Using Hidden Markov Models</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The goal of single document text or speech summarization is to identify information from a text or spoken document that summarizes, or conveys the essence of a document. EXTRACTIVE SUM-MARIZATION identifies portions of the original document and concatenates these segments to form a summary. How these segments are selected is thus critical to the summarization adequacy.</Paragraph>
    <Paragraph position="1"> Many classifier-based methods have been examined for extractive summarization of text and of speech (Maskey and Hirschberg, 2005; Christensen et. al., 2004; Kupiec et. al., 1995). These approaches attempt to classify segments as to whether they should or should not be included in a summary.</Paragraph>
    <Paragraph position="2"> However, the classifiers used in these methods implicitly assume that the posterior probability for the inclusion of a sentence in the summary is only dependent on the observations for that sentence, and is not affected by previous decisions. Some of these (Kupiec et. al., 1995; Maskey and Hirschberg, 2005) also assume that the features themselves are independent. Such an independence assumption simplifies the training procedure of the models, but it does not appear to model the factors human beings appear to use in generating summaries. In particular, human summarizers seem to take previous decisions into account when deciding if a sentence in the source document should be in the document's summary.</Paragraph>
    <Paragraph position="3"> In this paper, we examine a Hidden Markov Model (HMM) approach to the selection of segments to be included in a summary that we believe better models the interaction between extracted segments and their features, for the domain of Broadcast News (BN). In Section 2 we describe related work on the use of HMMs in summarization. We present our own approach in Section 3 and discuss our results in Section 3.1. We conclude in Section 5 and discuss future research.</Paragraph>
  </Section>
class="xml-element"></Paper>