<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2023">
  <Title>Summarizing Speech Without Text Using Hidden Markov Models</Title>
  <Section position="3" start_page="0" end_page="89" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> Most speech summarization systems (Christensen et. al., 2004; Hori et. al., 2002; Zechner, 2001) use lexical features derived from human or Automatic Speech Recognition (ASR) transcripts as features to select words or sentences to be included in a summary. However, human transcripts are not generally available for spoken documents, and ASR transcripts are errorful. So, lexical features have practical limits as a means of choosing important segments for summarization. Other research efforts have focussed on text-independent approaches to extractive summarization (Ohtake et. al., 2003), which rely upon acoustic/prosodic cues. However, none of these efforts allow for the context-dependence of extractive summarization, such that the inclusion of  one word or sentence in a summary depends upon prior selection decisions. While HMMs are used in many language processing tasks, they have not been employed frequently in summarization. A significant exception is the work of Conroy and O'Leary (2001), which employs an HMM model with pivoted QR decomposition for text summarization. However, the structure of their model is constrained by identifying a fixed number of 'lead' sentences to be extracted for a summary. In the work we present below, we introduce a new HMM approach to extractive summarization which addresses some of the deficiencies of work done to date.</Paragraph>
  </Section>
class="xml-element"></Paper>