<?xml version="1.0" standalone="yes"?>
<Paper uid="H91-1070">
  <Title>Interactive Problem Solving and Dialogue in the ATIS Domain 1</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> A key goal of spoken language systems is to provide support for interactive problem solving. To be effective in this role, a system should permit the user to build up a solution incrementally, and therefore to make implicit and explicit reference to information from earlier parts of the system/user dialogue. In addition, in certain scenarios, e.g., making flight reservations, it is natural to have the system play an active role in the dialogue. It is in fact our belief that spoken language systems will never reach a practical level of performance unless special attention is paid to issues of discourse and dialogue. To this end, we spent considerable effort over the last several months augmenting our system's discourse capabilities, and developing a much more active role for the computer during the dialogue with a user.</Paragraph>
    <Paragraph position="1"> We can view interactive problem solving systems on a continuum of system/user control. At one extreme would be a system that answers user's questions without at any point  asking questions or making suggestions. At the other extreme would be a menu system that forces the user into a very limited range of choices. This paper will describe an attempt at MIT to explore a middle ground in this continuum, in the context of allowing users to make flight reservations.</Paragraph>
    <Paragraph position="2"> There are a number of advantages to an approach which explores a mixed initiative system for flight reservations. First, it provides the user with a more focused way to interact: making a reservation provides a natural notion of task completion. Second , a system that can take the initiative on occasion not only can provide a more flexible and natural interaction (e.g., asking for clarification, eliciting missing information) but also can reduce the search space for spoken input, thus potentially improving performance.</Paragraph>
    <Paragraph position="3"> There are also a number of potential disadvantages to such an approach. One problem with making reservations is the difficulty in eliciting clean data that does not contain extraneous information ('names of travelers, VISA card numbers, etc.) A problem that arises when the system takes the initiative is that it must have substantial problem solving expertise; it crosses the line into an expert system, and there{ore may extend beyond the scope of the Spoken Language Systems effort as currently defined. Third, such systems are much more difficult to evaluate automatically. Finally, by making the system more natural, we run the risk of raising user expectations too high.</Paragraph>
    <Paragraph position="4"> This paper describes the present status of the discourse and dialogue models within the MIT AWlS system. After describing the models, we will illustrate some of the system's capabilities by way of an example. We then describe our preliminary attempts at collecting data in a booking mode, for which we hive included a complete dialogue elicited from one of our subjects. Finally, potential implications for improvemeats in speech recognition are discussed.</Paragraph>
  </Section>
class="xml-element"></Paper>