<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1110">
  <Title>Semantic Similarity Applied to Spoken Dialogue Summarization</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Concluding Remarks
</SectionTitle>
    <Paragraph position="0"> We introduced a new approach to spoken dialogue summarization. Our approach combines statistical, i.e. corpus-based, and knowledge-based techniques.</Paragraph>
    <Paragraph position="1"> It utilizes the knowledge encoded in the noun part of WordNet and applies a set of semantic similarity metrics to dialogue summarization. All semantic similarity-based summarization methods outperform RANDOM, LEAD and TF*IDF baseline systems. In the following, we discuss some remaining challenges and future research.</Paragraph>
    <Paragraph position="2"> More sophisticated data pre-processing. We plan 4Roughly speaking, the differences are most evident for compression rates between 20% and 30%.</Paragraph>
    <Paragraph position="3"> to incorporate the pre-processing components used by Zechner (2002) and evaluate their contribution to our task. Including an anaphora resolution component would also result in better Recall.</Paragraph>
    <Paragraph position="4"> Automatic word sense disambiguation. Switchboard conversational speech is highly ambiguous. Automatic disambiguation of noun senses to Word-Net concepts is important in order to integrate our approach into real-life summarization systems.</Paragraph>
    <Paragraph position="5"> Investigating other types of information in parallel.</Paragraph>
    <Paragraph position="6"> A clear desideratum will be assessing the overall coherence of the discourse, speaker info, turn type, information about non-nouns.</Paragraph>
    <Paragraph position="7"> Application to text and speech summarization. Our approach can be applied to written-to-be-spoken speech and text summarization. It will be interesting to investigate whether conceptual structures of texts (the input to our system) are comparable to the conceptual structures found in dialogues.</Paragraph>
    <Paragraph position="8"> Readability, coherence, and usability of the summaries produced. A close examination of summaries based on human comprehension will be interesting. It may be necessary to introduce filtering or other post-processing techniques improving the quality of summaries.</Paragraph>
    <Paragraph position="9"> Even without very sophisticated pre-processing of the dialogue data, our algorithm yields promising results. It was evaluated on the Switchboard data, which is a challenging evaluation corpus. Our vision is to adopt the summarization approach presented here in a system used for the automatic production of meeting minutes.</Paragraph>
  </Section>
class="xml-element"></Paper>