<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2019">
  <Title>Early Deletion of Fillers In Processing Conversational Speech</Title>
  <Section position="5" start_page="75" end_page="75" type="concl">
    <SectionTitle>
4 Conclusion
</SectionTitle>
    <Paragraph position="0"> While early deletion of fillers has limited benefit for in-domain parsing of speech data, it can play an important role in bleaching speech data for more accurate text-based processing. Alternative methods of integrating detected filler information, such as parse reranking (Kahn et al., 2005), also merit investigation. It will also be important to evaluate the interaction with ASR error and sentence boundary detection error. In terms of bleaching, we saw that even with oracle detection of disfluency, our text-trained model still significantly under-performed the in-domain model, indicating additional methods for bleaching are still needed. We also plan to evaluating the benefit of disfluency modeling in bleaching speech data for text-based machine translation.</Paragraph>
  </Section>
class="xml-element"></Paper>