<?xml version="1.0" standalone="yes"?>
<Paper uid="H89-2017">
  <Title>DATA COLLECTION AND ANALYSIS IN THE AIR TRAVEL PLANNING DOMAIN</Title>
  <Section position="3" start_page="0" end_page="119" type="intro">
    <SectionTitle>
INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> Spoken language systems must, obviously, deal with spontaneous speech. However, most research to date has dealt primarily with read speech, because read speech is much easier to collect in a controlled manner. There are, however, substantial differences between read speech and spontaneous speech.</Paragraph>
    <Paragraph position="1"> Differences include the many phenomena that are less likely to occur in read speech (pauses, speech and grammatical false starts, filler words, non-standard grammar), as well as important phonological phenomena, such as the frequency of deletions (Bernstein and Baldwin, 1985). On the other hand, it is possible that both the speech and the language of human-machine interactions in a restricted domain will be more constrained and more predictable than those occurring in human-human spontaneous interactions. The goal of the preliminary work presented here is to collect and analyze spontaneous, goal-directed speech and language in the interest of designing and evaluating eventual spoken language systems.</Paragraph>
    <Paragraph position="2"> Perhaps the greatest variable affecting performance in current and future systems is the human involved in the human-machine interface. It is therefore important to assess systems over many different subjects. We have chosen the domain of air travel planning because it provides a natural problem-solving domain familiar to many people (120 SRI employees per day on average use spoken interactions to solve travel planning problems). This has greatly facilitated the task of collecting data. Further, the domain can be constrained as desired for initial development (as we have done by allowing only one-way travel between two cities), or expanded naturally to include a great deal of complex problem-solving for future SLSs (inclusion of data on connections, classes of seats, and restrictions on fares, availability of fares, hotels, car rentals, expert system reasoning, etc.). In addition, the air travel planning domain has the advantage of large, real databases in the public domain.</Paragraph>
    <Paragraph position="3"> We initially studied human-human interactions, to gain insight into how interactive problem solving is currently used in this domain. We noted that database queries were rare, and that more typically the traveler expresses a few constraints, and then the agent takes the lead and asks questions. We wondered how adaptable subjects would be in a simulated machine interaction: would their travel planning task be more difficult if they were forced to use only database queries? We, simulated an SLS in two conditions: one that permitted the expression of constraints but that were not strictly database queries CI  need to be there before 3 pm&amp;quot;), and one which accepted only database queries (responding &amp;quot;cannot handle that request&amp;quot; to any other type of utterance). The system responds, in both conditions, with graphics placed on the user's screen (shared information, schedule tables, fare tables, etc.).</Paragraph>
    <Paragraph position="4"> The goal of this initial won is to assess human-human problems solving in the air travel domain, and to assess possible differences between human-human and human-machine interactions. It is clear that people are very adaptable, far more so than our current technology. It is not so clear how adaptable they will be and on what dimensions in human-machine interactions. What aspects of the interaction will require a technological solution and what aspects can be handled via a human factors solution? If, for example, it is desirable to handle only database queries, how difficult is it for humans to adapt to this restriction? This is but one example of a myriad of similar questions that could be asked using such simulations. The answers to these questions will expedite the design of efficient human-machine collaborative systems.</Paragraph>
  </Section>
class="xml-element"></Paper>