<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1608">
  <Title>Incremental Generation of Multimodal Deixis Referring to Objects</Title>
  <Section position="4" start_page="0" end_page="0" type="relat">
    <SectionTitle>
3 Related Work
</SectionTitle>
    <Paragraph position="0"> While much work concerning the generation of referring expressions has been published over the last 15 years, work on the generation of multi-modal referring expressions is rare.</Paragraph>
    <Paragraph position="1"> Most of the approaches which can be found in this field use idealised pointing in addition or instead of referring expressions. [Claassen, 1992] and [Reithinger, 1992] highlight the referent in two-dimensional settings by an idealised pointing gesture represented by an arrow or a schematic hand. [Noma and Badler, 1997] and [Andr'e et al., 1999] introduce virtual agents in presentation tasks able to produce simple pointing gestures. [Lester et al., 1999] and [Rickel and Johnson, 1999] generate pointing gestures expressed by an agent which moves to the referent, and therefore, achieve unambiguous pointing. Only [Krahmer and van der Sluis, 2003] integrate pointing and definite descriptions in a more natural way and account for vague pointing. They distinguish three types of preciseness, i.e. precise, imprecise, or very imprecise pointing, and integrate pointing into the graph-based algorithm proposed by [Krahmer et al., 2003].</Paragraph>
    <Paragraph position="2"> Examining the generation of referring expressions realised as definite descriptions one has to mention, first of all, that the problem of selecting the minimal set (in the sense of Grice's quantity maxim) of object properties needed for an unambiguous description of the referent has exponential computational complexity [Reiter, 1990]. Each combination of properties has to be tested whether it is true only for the referent, and the shortest one of these combinations has to be chosen.</Paragraph>
    <Paragraph position="3"> Especially for real-time applications in domains with high object density and objects with a high number of properties this computation is intractable with brute-force methods. Several approaches have been proposed to deal with this problem, namely [Dale, 1992; Krahmer et al., 2003; Horacek, 1997; Gardent, 2002]. [Dale and Reiter, 1995] proposed an incremental algorithm which violates the quantity maxim in the strict sense, but achieves linear compute time and fits well with empirical findings.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 The Incremental Algorithm by Dale and Reiter
</SectionTitle>
      <Paragraph position="0"> To achieve linear compute time, [Dale and Reiter, 1995] propose a fixed sequence of property evaluation and avoid backtracking. This approach leads to over-specification, but they can show that the generation results fit well with empirical findings if the sequence of properties is chosen accurately w.r.t. the specific domain. Therefore, the content-selection algorithm (see Alg. 3.1) gets, in addition to the referent r and the context set C, also a sorted list of properties P as an input.</Paragraph>
      <Paragraph position="1"> The functionality of this algorithm can be described in short as follows. In the ordering of P each property Ai in P is evaluated concerning its discriminatory power, that means it is checked if there is at least one object in C which has another value for A than the referent r has. These objects are ruled out. If the contrast set C is empty the algorithm terminates and returns a list with the discriminating properties L.</Paragraph>
      <Paragraph position="2"> W.r.t. observations in their corpora Dale and Reiter add the property type everytime. The task of FINDBESTVALUE is the search for the most specific value of an attribute that (1), discriminates the referent r from more elements in D than the next general one does, and (2), is known by the addressee.</Paragraph>
      <Paragraph position="3"> We chose this algorithm as a starting point for our work and adapted it for multimodal expressions because of its appropriateness w.r.t. empirical data and its efficient compute  time .</Paragraph>
      <Paragraph position="4"> Algorithm 3.1: MAKEREFERRINGEXPRESSION(r,C,P)</Paragraph>
      <Paragraph position="6"> if C ={}then if (type,X)[?]L for some X</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>