<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1026">
  <Title>Learning the Structure of Task-driven Human-Human Dialogs</Title>
  <Section position="4" start_page="0" end_page="201" type="intro">
    <SectionTitle>
2 Current Methodology for Building
</SectionTitle>
    <Paragraph position="0"> Dialog systems Current approaches to building dialog systems involve several manual steps and careful crafting of different modules for a particular domain or application. The process starts with a small scale Wizard-of-Oz data collection where subjects talk to a machine driven by a human 'behind the curtains'. A user experience (UE) engineer analyzes the collected dialogs, subject matter expert interviews, user testimonials and other evidences (e.g. customer care history records). This heterogeneous set of information helps the UE engineer to design some system functionalities, mainly: the  semantic scope (e.g. call-types in the case of call routing systems), the LG model, and the DM strategy. A larger automated data collection follows, and the collected data is transcribed and labeled by expert labelers following the UE engineer recommendations. Finally, the transcribed and labeled data is used to train both the ASR and the SLU.</Paragraph>
    <Paragraph position="1"> This approach has proven itself in many commercial dialog systems. However, the initial UE requirements phase is an expensive and error-prone process because it involves non-trivial design decisions that can only be evaluated after system deployment. Moreover, scalability is compromised by the time, cost and high level of UE know-how needed to reach a consistent design.</Paragraph>
    <Paragraph position="2"> The process of building speech-enabled automated contact center services has been formalized and cast into a scalable commercial environment in which dialog components developed for different applications are reused and adapted (Gilbert et al., 2005). However, we still believe that exploiting dialog data to train/adapt or complement hand-crafted components will be vital for robust and adaptable spoken dialog systems.</Paragraph>
  </Section>
class="xml-element"></Paper>