<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2050">
  <Title>Comparing the roles of textual, acoustic and spoken-language features on spontaneous-conversation summarization</Title>
  <Section position="4" start_page="197" end_page="197" type="intro">
    <SectionTitle>
2 Utterance-extraction-based
</SectionTitle>
    <Paragraph position="0"> summarization Still at its early stage, current research on speech summarization targets a less ambitious goal: conducting extractive, single-document, generic, and surface-level-feature-based summarization. The pieces to be extracted could correspond to words (Koumpis, 2002; Hori and Furui, 2003). The extracts could be utterances, too. Utterance selection is useful. First, it could be a preliminary stage applied before word extraction, as proposed by Kikuchi et al. (2003) in their two-stage summarizer. Second, with utterance-level extracts, one can play the corresponding audio to users, as with the speech-to-speech summarizer discussed in Furui et al. (2003). The advantage of outputting audio segments rather than transcripts is that it avoids the impact of WERs caused by automatic speech recognition (ASR). We will focus on utterance-level extraction, which at present appears to be the only way to ensure comprehensibility and naturalness if the summaries are to be delivered as excerpts of audio themselves.</Paragraph>
    <Paragraph position="1"> Previous work on spontaneous conversations mainly focuses on using textual features. Gurevych &amp; Strube (2004) develop a shallow knowledge-based approach. The noun portion of WordNet is used as a knowledge source. The noun senses were manually disambiguated rather than automatically.</Paragraph>
    <Paragraph position="2"> Zechner (2001) applies maximum marginal relevance (MMR) to select utterances for spontaneous conversation transcripts.</Paragraph>
  </Section>
class="xml-element"></Paper>