<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0703">
  <Title>Directions For Multi-Party Human-Computer Interaction Research</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Most current work on spoken human-computer interaction (HCI) involves dialog systems. In recent years, spoken dialog systems with system initiative have become more commonplace in commercial telephony applications, and there have been important advances in mixed initiative and multi-modal research systems. Telephone-based systems have made it possible to collect large amounts of human-computer interaction data, which has benefited empirical research as well as methods based on automatic training. In addition, evaluation frameworks have improved beyond the single utterance accuracy measures used a decade ago to dialog-level subjective and quantitative measures (Walker et al., 1998).</Paragraph>
    <Paragraph position="1"> As dialog systems have advanced, a new area of research has also been developing in automatic recognition and analysis of multi-party human-human spoken interactions, such as meetings, talk shows, courtroom proceedings, and industrial settings (Cohen et al., 2002). Multi-party interactions pose challenges for speech recognition and speaker tracking because of frequent talker overlap (Shriberg et al., 2001), noise and room reverberation, but they also introduce new challenges for discourse modeling. Until recently, empirical research was only possible using single-speaker and dialog corpora, but now there are many hours of data being collected in multi-talker environments (Morgan et al, 2001; Schultz et al, 2001).</Paragraph>
    <Paragraph position="2"> While many challenges remain in dialog systems from error handling and user modeling to response generation - technology has advanced to the point where one can also envision tackling the combined problem of multi-party human-computer interaction. A key motivation for research in such a domain is supporting human-human collaboration. We envision a scenario where a computer plays a role as a conversational agent, much as in a dialog system, except that it interacts with multiple collaborating humans. The human participants may be at distributed locations, perhaps with small subgroups at each location, possibly with different platforms for input and output. For example, one might imagine a group of people in a facility with high-end computers interacting with workers in the field with lightweight communication clients, using the computer assistant to help gather vital information or help plan a transportation route. A key difference from previous work in such scenarios is the idea of computer initiative. The computer as a participant also significantly changes the focus of research relative to that involved in transcription and analysis of meetings, from work aimed at indexing and summarization to a focus on interaction.</Paragraph>
    <Paragraph position="3"> Besides the application-oriented motivation for research on multi-party human-computer interaction, the scenario provides a useful technology pull. In current dialog systems, there is a disincentive to explore user initiative, simply because much better accuracy can be achieved by &amp;quot;controlling&amp;quot; the dialog. However, it would be impractical for a system to try to constrain the inputs from a group of users. Secondly, current dialog systems generally assume a fixed platform, and hence the response generation can be greatly simplified. With varying platforms and participants with different needs, a more complex output rendering strategy will be needed, which will also have implications for future dialog systems as well. In the follow section, we expand on these issues and many more research questions that arise in the context of multi-party HCI research.</Paragraph>
  </Section>
class="xml-element"></Paper>