<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2508">
  <Title>Experiments with Interactive Question Answering in Complex Scenarios</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 Domain-Dependent Complex Questions
</SectionTitle>
    <Paragraph position="0"> The decomposition of complex information-seeking scenarios represents a trio of pragmatic challenges for Q/A systems.</Paragraph>
    <Paragraph position="1"> First, effective question decomposition depends on the acknowledgment of the intentions that underlie a user's interaction with a Q/A system. Individuals participate in information-seeking dialogues (whether with other humans or with interactive Q/A systems) in order to learn new things - that is, to gather information that they do not currently possess. A user's behavior in a dialogue focuses on that set of speech acts which allow them to maximize the new information they obtain from the conversation while (at the same time) minimizing the amount of redundant or previously-established information that they encounter. We expect these same principles to govern the decomposition of complex scenarios as well: the decompositions generated by a user will focus on returning the domain-specific information that the user currently does not possess. Expert users (who are assumed to be familiar with a domain) will use interactive Q/A systems to (1) evaluate their existing knowledge with regards to changes in the context or (2) seek new information about known entities or events within the domain. In contrast, non-expert users (who remain unfamiliar with much of the ontological structure of a complex domain) will have very broad and potentially poorly-defined informational goals; in these cases, interactive Q/A systems will need to return information which will facilitate the novice users' exploration of the domain.</Paragraph>
    <Paragraph position="2"> Second, we suggest that question decomposition will depend on the development of semantic ontologies that are articulated enough to address the domain-specific questions characteristic of most complex information-seeking scenarios. Current Q/A technologies are unable to process (or decompose) complex questions without access to a large amount of domain-specific knowledge.</Paragraph>
    <Paragraph position="3"> Modeling domain-specific knowledge for complex domains, however, is an arduous task: complex domains necessarily consist of sets of structured concepts linked by classes of semantic relations. Although this kind of domain modeling is traditionally considered to be tangential to research in NLP, we believe that interactive Q/A systems must have access to not only the ontological structure of answers and complex semantic information, but also modes of probabilistic reasoning that can be used to induce categories of meanings between domain concepts.</Paragraph>
    <Paragraph position="4"> Finally, since complex questions represent such diverse informational goals, it should not be assumed that even the decompositions produced by expert users will be sufficiently simple enough to be processed by current Q/A systems. We propose that careful study needs to be conducted to identify the new types of context-dependent questions that are generated as part of interactive Q/A.</Paragraph>
    <Paragraph position="5"> The rest of this section is ordered as follows. Section 2.1 describes three new types of questions found in the sets of decompositions generated by human users. Section 2.2 details a simple solution that expands the coverage of interactive Q/A systems for specific topic domains. Finally, Section 2.3 distinguishes between two idealized types of users of interactive Q/A systems: experts and novices.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1 Scenarios and Questions
</SectionTitle>
      <Paragraph position="0"> Since complex questions represent such diverse informational goals, it should not be assumed that even the decompositions produced by expert users will be sufficiently simple enough to be processed by current Q/A systems.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>