<?xml version="1.0" standalone="yes"?>
<Paper uid="W01-1629">
  <Title>Spoken Dialogue Control Based on a Turn-minimization Criterion Depending on the Speech Recognition Accuracy</Title>
  <Section position="3" start_page="0" end_page="0" type="intro">
    <SectionTitle>
2 Method
</SectionTitle>
    <Paragraph position="0"> Overview First, we describe about a system to which we assume this method will be applied. The system has belief state which is represented by the set of attributes, their values, and the certainty of the values. The certainty is in [0 .. 1], and the certainty for the determined value is 1. That is, if the user replies &amp;quot;Yes&amp;quot; to the confirmation, the systemchanges thecertaintyforthatvalue to 1. In practice, we can use the score from the recognition engine as this certainty. The system changes the recognition vocabulary according to the attributes to be confirmed at each confirmation. At any given time, the system either confirms or demands some attribute(s); it doesn't confirm and demand at the same time. Any values required in order to determine the user request are explicitlyconfirmed without exception. Words that are irrelevant to the present confirmation are excluded from the recognition vocabulary. The system knows the base recognition accuracy under acertainvocabularysize, whichisused to estimate the recognition accuracy.</Paragraph>
    <Paragraph position="1"> Our method can be divided roughly into five parts; the first three parts are used to obtaintheexpectednumberofturns,granting thatthe user request type are alreadyknown, the fourth part is used to approximate the probability distribution of the user request, and the last part is used to decide the next action to be taken by the system.</Paragraph>
    <Paragraph position="2"> The system needs to know only three sorts of information: 1) the vocabulary for each attribute; 2) the meaning constraints among words like &amp;quot;If the family name of the person is Yasuda, then his department must be accounting&amp;quot;; and 3) the required information for each type of user request like &amp;quot;To cancel an appointment; the day and the time are required&amp;quot;. No other domain-specific rules or training are necessary.</Paragraph>
    <Paragraph position="3"> Guessing the Recognition Accuracy  Hereweconsiderhowtoestimatetherecognition accuracy during confirmation from confirmation target. Once attributes for confirmation are decided, the recognition vocabulary will consist of the words accepted by the attributes and general words for moving the dialogue along that are at least necessary to progress the dialogue such as &amp;quot;Yes&amp;quot;, &amp;quot;No&amp;quot;, etc. We call the recognition accuracy at this time the &amp;quot;attribute recognition accuracy&amp;quot;. We adopt the rule ofthumb thatthe recognitionerrorrateisinproportiontothesquare null  rootofvocabularysize(Rosenfeld,1996;Nakagawa and Ida, 1998). Thus, the approximated attribute recognition accuracy can be derived from the number of words accepted by the attributes.</Paragraph>
    <Paragraph position="4"> Note that the attribute recognition accuracy can't be estimated beforehand, because  thecandidatesforsomeattributesaredynamically change, as a result of the meaning constraints among words; if the value of one attribute is fixed, then candidates for other attributes will be limited to values that satisfy the constraints. Besides, the degree of limitation varies with the values. The relation between the user's family name and department is such an example.</Paragraph>
    <Paragraph position="5"> Turn Estimation to Determine Some Attributes Next we consider how to estimate the expected number of turns for determining some attributes using the approximated attribute recognition accuracy.</Paragraph>
    <Paragraph position="6"> We assumethatthe user'sreplytothe confirmationmustcontaintheintentionthatcor- null responds to &amp;quot;Yes&amp;quot; or &amp;quot;No&amp;quot;, and the intention must be transmitted to the system without fail. Then, the expected number of turns to complete confirming for some attributes is equal to the expected number of turns in the case that the confirmation is incorrect (i.e.</Paragraph>
    <Paragraph position="7"> misrecognized). Therefore, we can derive the number of expected turns to complete con- null where r denotes the attribute recognition accuracyforattributesthataretobeconfirmed. null Turn Estimation to a Certain User Request Type Here weestimatetheexpected number of turns, granting that the type of user request is already known.</Paragraph>
    <Paragraph position="8"> If the user request type is fixed, the required attributes for that type are also fixed. By comparing the belief state with these attributes,wecanrepresenttherequiredactions null todeterminethe userrequest byasetofpairs made up of attributes and actions for the attribute (confirmation or demand). Once this setofpairsisgiven,wecanchoosetheoptimal plan, because we can estimate the expected turns of any permutations of any partitions of this set. The expected number of turns for this optiomal plan is used as the expected number ofturnsfora givenuserrequest type.</Paragraph>
    <Paragraph position="9"> Probability Distribution of User Request Types Here, we consider how to estimate the relevance between the belief state and each user request types.</Paragraph>
    <Paragraph position="10"> As it is hard to obtain the actual probability distribution, we define the degree of relevance between the belief state and each user request type as an approximation.</Paragraph>
    <Paragraph position="11"> Let a</Paragraph>
    <Paragraph position="13"/>
    <Paragraph position="15"> Choosing the Next Action Even if there is a highly possible user request type, choosingconfirmationplanforitisnotalwaysbest, null if the expected number of turns for that request is very large. In such case, confirming another type of request that is easily confirmed and medium possibility may better.</Paragraph>
    <Paragraph position="16"> Weassumethatwhen theuserrequesttype guessed by the system is not the real user request type, the number of turns required to know that the guess is incorrect is equal to the number of turnswhen the guess is correct and finish confirming the contents.</Paragraph>
    <Paragraph position="17"> Let p</Paragraph>
    <Paragraph position="19"> be the probability of user request</Paragraph>
    <Paragraph position="21"> be the expected number of turns to user request type R</Paragraph>
    <Paragraph position="23"> From permutations of request types, our method chooses the optimal order a(1),a(2),...,a(n) such that the expression</Paragraph>
    <Paragraph position="25"> ) is minimal. Then our method chooses the action that appears firstintheoptimalplanforrequesttype R a(1) as the next action.</Paragraph>
  </Section>
class="xml-element"></Paper>