<?xml version="1.0" standalone="yes"?>
<Paper uid="W01-1610">
  <Title>Labeling Corrections and Aware Sites in Spoken Dialogue Systems</Title>
  <Section position="5" start_page="1" end_page="1" type="concl">
    <SectionTitle>
4 Discussion
</SectionTitle>
    <Paragraph position="0"> Thispaperhasdealt withusercorrections and aware sites of system errors in the TOOT spoken dialogue system. We have described our corpus, and havegiven details on our procedure to label corrections and aware sites. Then, we have shown that corrections and aware sites exhibit some prosodic and other propertieswhich set themapart from`normal' utterances. It appears that some correction types, such as simple repeats, are more likely to be correctly recognized than other types, such as paraphrases. We have also presented evidence that system dialogue strategy aects users' choice of correction type, suggesting that strategy-specic methods of detecting or coaching users on corrections may be useful. Aware sites tend to be shorter than other utterances, and are also more dicult to recognize correctly for the ASR system.</Paragraph>
    <Paragraph position="1"> In addition to the descriptive study presented in this paper, wehave also tried to automatically predict corrections and aware sites using the machine learning program RIP-PER (Cohen, 1996). These experiments show that corrections and aware sites can be classied as such automatically, with a considerable degree of accuracy (Litman et al., 2001;; Hirschberg et al., 2001). Such classication, we believe, will be especially useful in error-handling for SDS. If aware sites are detectable, they can function as backward-looking error-signaling devices, making it clear to the system that something has gone wrong in the preceding context, so that, for example, the system can reprompt for information. In this way, they are similar to what others have termed `go-back' signals (Krahmer et al., 1999). Aware sites can also be used as forward-looking signals, indicating upcoming corrections or moredrastic changesinuserbehavior, such as complete restarts of the task. Given that, in current systems, both corrections and restarts often lead to recognition error (Swerts et al., 2000), aware sites may be useful in preparing systems to deal with such problems. An accurate detection of turnsthat are corrections maytrigger the useof specially trained ASR models to better recognize corrections, or can be used to change dialogue strategy (e.g. from user or mixed initiativeto system initiative after errors).</Paragraph>
  </Section>
class="xml-element"></Paper>