<?xml version="1.0" standalone="yes"?>
<Paper uid="A92-1042">
  <Title>Acquiring and Exploiting the User's Knowledge in Guidance Interactions</Title>
  <Section position="3" start_page="0" end_page="249" type="intro">
    <SectionTitle>
3 Integrating Explicit and Im-
</SectionTitle>
    <Paragraph position="0"> plicit Methods The model suggested here integrates the explicit acquisition method with a variety of implicit methods in a context of a Guidance dialog. In contrast to advisory dialogs that usually contain an interview phase aiming to elicit from the user information required for the advice generation, in guidance interactions most of the dialog consists of instructions provided by the system, and user's responses to these instructions. Although questions to the guidance recipient - the guidee - do occur, they do not constitute a separate phase but are rather scattered in the instruction sequence. In the current study and  in other studies of human guidance dialogs it was found that in such dialogs any instruction made by the guide is followed by a guidee's response such as question, interruption, or confirmation-signal. In human interaction confirmation-signals are provided by either verbal (e.g., &amp;quot;ahm&amp;quot;) or non-verbal (e.g., head nodding) means, this can be simulated in human-computer interaction by a special &amp;quot;continue&amp;quot; button that is used to signal the system to move to the next instruction.</Paragraph>
    <Paragraph position="1"> We claim that any instruction-response pair in the guidance sequence can serve as a basis for inferences about the guidee. In general, each such pair type triggers different inferences and consumes a different amount of the user's resources (time, effort, patience). For example, inferences can be drawn from an instruction/confirmation-signal pair; the confirmation signal shows that the user understands the instruction and thus affirms the assumptions which led to its production. Such default inferences are less certain than those based on answers to explicit questions, but they do not consume the resources of the guidee: No special action from him is required; in fact the guidee may not even be aware to the fact that inferences about him are being made. A strategy of explicit questioning is justified only when the needed fact is informative, and there was no way in which the system could infer it from the previous dialog. In any other case a more implicit strategy should be preferred.</Paragraph>
    <Paragraph position="2"> This view extends the explicit-implicit distinction mentioned above since each utterance pair type facilitates a different acquisition method, and each such method has a different level of explicitness. Our model provides therefore, a new definition of the notion of explicitness which extends the previous notion in two ways. First, we distinguish between Acquisition Explicitness and Transmission Explicitness, this way the notion can be applied both for information flowing from guide to guidee (guidance information), and for information flowing from guidee to guide (user modeling information). Second, we allow various levels of explicitness, rather than the two extremes only. The system's explicitness level is defined in terms of the interactive effort that is invested in the dialog. This can either be the effort required to provide a fact to the guidee, or the effort required to acquire a fact about him.</Paragraph>
    <Paragraph position="3"> We define FIGS as a system that provides interactive, user-adapted guidance, and satisfies the following two conditions: (1) It is equipped with a set of instruction-acquisition strategies, each of which is characterized by a different level of acquisition explicitness and transmission explicitness; and (2) It has a mechanism to dynamically select a strategy.</Paragraph>
    <Paragraph position="4"> This mechanism attempts to reach optimMity with respect to the interactive effort required for providing the information needed by the guidee.</Paragraph>
    <Paragraph position="5"> The suggested model implements the foregoing considerations, it includes four instruction-acquisition strategies: (1) Question - the user provides a needed fact, (2) Explanation - no facts about the user are used, (3) Explicit Assumption - the instruction is based on an assumption about the user, and the assumption is mentioned, and (4) Implicit Assumption - an assumption is used without mentioning it.</Paragraph>
    <Paragraph position="6"> A major feature of our model is that UM acquisition considerations are integrated into the FIGS's utterance planning process. Hence, the user modeling and the instruction generation are done incrementally by inter-related processes. We use heuristic rules to select among the four strategies; these rules weigh four discourse parameters: information content, user knowledge (as described by the current UM), likelihood to acquire new facts about the user, and consumption of user's resources.</Paragraph>
    <Paragraph position="7"> The computational model we suggested was implemented by a computer program (called FIGS1) that offers directional instructions in a complex university building. FIGS1 uses heuristic rules to control both the strategy selection and the UM construction. A more detailed exposition of our model and its implementation can be found in (Shifroni &amp; Shanon, 1991), and (Shifroni &amp; Ornan, 1991).</Paragraph>
  </Section>
class="xml-element"></Paper>