<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-2016">
  <Title>Controlling Animated Agents in Natural Language</Title>
  <Section position="6" start_page="94" end_page="94" type="concl">
    <SectionTitle>
5 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> We have introduced our prototype systemK3, two distinctive features of which are described in this paper. Plan-based anaphora resolution enables K3 to interpret the user's intention more precisely than the previous, surface-cue-based resolution algorithms. The SPACE object is designed to bridge the gap between the symbolic system (language processing) and the continuous system (animation generation) . In what follows, we describe the research agenda of our project.</Paragraph>
    <Paragraph position="1"> One-to-many Conversation. Conversational agent systems should deal with one-to-many conversations as well as one-to-one conversations.</Paragraph>
    <Paragraph position="2"> In a one-to-many conversation, it is not easy to decide who is the intended listener. The situation gets worse when a speaker is concerned with only performing an action without caring who does it. In such cases, agents have to request clarifications or negotiate among themselves.</Paragraph>
    <Paragraph position="3"> Agent Coordination. In one-to-many conversations, agents must coordinate each other. Some sorts of coordination are explicitly requested by user, e.g., &amp;quot;Agent A and B tidy up the table, please.&amp;quot; But other kinds of coordination are implicitly requested, e.g., &amp;quot;Agent A hands agent B the box, please.&amp;quot; In this case, the speaker asks agent B nothing explicitly. However, agent B must react to the request for agent A and coordinate with agent A to receive a box.</Paragraph>
    <Paragraph position="4"> Parallel Actions. Most intelligent agent systems perform only one action at a time. Yet, if we want to make systems become more flexible, we must enable them to handle more than one action at a time. Hence, they must speak while walking, wave while nodding, and so on.</Paragraph>
    <Paragraph position="5"> Memory System. A history database is not enough to serve realistic dialogue in the domain of K3. In such a domain, people often mention a previous state, e.g., &amp;quot;Put the ball back to the place where it was.&amp;quot; To comply such a request, agents must have a human-like memory system.</Paragraph>
    <Paragraph position="6"> Interruption Handling. Agents sometimes misunderstand requests and perform not intended actions. In case of human conversations, a speaker usually interrupts hearer and try to repair misunderstanding. Conversational agents also should be able to accept such interruptions.</Paragraph>
    <Paragraph position="7"> Interruption handling is also essential to request a next action before agents finish actions.</Paragraph>
  </Section>
class="xml-element"></Paper>