<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-3002">
  <Title>WoZ Simulation of Interactive Question Answering</Title>
  <Section position="5" start_page="14" end_page="15" type="evalu">
    <SectionTitle>
5 Discussion
</SectionTitle>
    <Paragraph position="0"> The analysis of dialogue collected by the WoZ simulation showed very similar results to those obtained in a previous study, in which sequences of questions were written down by subjects without knowing the answers to questions. That is, as shown in Table 2, when users asked questions to get information for a report, the number of why-questions was relatively small. Moreover, there were fewer questions requesting an explanation or definition than expected, probably because definition questions such as &amp;quot;Who is Mr. Colin Powell?&amp;quot; were decomposed into relatively concrete questions such as those asking for his birthday and birthplace. The remainder (65%) could be answered in values and names. Table 3 indicates that 62% of the questions in our experiments were answered by values or names. If compound nouns describing events or situations, which are usually distinguished from names, are considered to be in the range of answers, the percentage of answerable questions reaches 68%. From these results, the setting of QACIAD looks realistic where users write reports interacting with a QA system handling factoid questions that have values and names as answers.</Paragraph>
    <Paragraph position="1"> A wide range of reference expressions is observed in information access dialogues for writing reports.</Paragraph>
    <Paragraph position="2"> Moreover, our study confirmed that those sequences of questions were sometimes very complicated and included subdialogues and focus shifts. It is expected that using an interactive QA system that can manage those pragmatic phenomena will enable fluent information access dialogue for writing reports.</Paragraph>
    <Paragraph position="3"> In this sense, the objective of QACIAD is appropriate. null It could be concluded from these results that the reality and appropriateness of QACIAD was reconfirmed in a more realistic situation. And yet suspicion remains that even in our WoZ simulation, the subjects were not motivated appropriately, as suggested by the lack of dynamic dialogue development in the example shown in Figure 2. Especially, the users often gave up too easily when they did not obtain answers to prepared questions.4 The truth, however, may be that in the environment of gathering information for writing reports, dynamic dialogue development is limited compared to the case when trained analysts use QA systems for problem solving. If so, research on this type of QA systems represents a proper milestone toward interactive QA systems in a broad sense.</Paragraph>
    <Paragraph position="4"> Another finding of our experiment is the importance of cooperative and helpful responses. Nearly half of WoZ utterances were not simple literal responses but included some cooperative and helpful behavior. This situation contrasts with a relatively small number of clarification dialogues. The importance of this behavior, which was emphasized in research on dialogues systems in the 80s and 90s, was reconfirmed in the latest research, although question-answering technologies were redefined in the late 90s. Some behavior such as providing alternative information could be viewed as a second-best 4It is understandable, however, that there were few rephrasing attempts since users were informed that paraphrasing such as &amp;quot;What is the population of the US?&amp;quot; to &amp;quot;How many people are living in the US?&amp;quot; are usually in vain.</Paragraph>
    <Paragraph position="5">  strategy of resource-bounded human WoZs. Even so, it is impossible to eliminate completely the need for such a strategy by improving core QA technologies. In addition, intrinsic cooperative and helpful behavior such as providing additional information was also often observed. These facts, accompanied by the fact that such dialogues are perceived as fluent and felicitous, suggest that the capability to behave cooperatively and helpfully is essential for interactive QA technologies.</Paragraph>
  </Section>
class="xml-element"></Paper>