<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0614">
  <Title>Intentional Context in Situated Natural Language Learning</Title>
  <Section position="4" start_page="104" end_page="105" type="intro">
    <SectionTitle>
2 Intention Recognition
</SectionTitle>
    <Paragraph position="0"> The ability to infer the purpose of others' actions has been proposed in the psychological literature as essential for language learning in children (Tommasello, 2003, Regier, 2003). In order to understand how such intention recognition might be modeled in a computational framework, it is useful to specify the types of ambiguities that make intentional actions difficult to model. Using as an example the situation involving the cup described above, we propose that this interaction demonstrates two distinct types of ambiguity. The first type, which we refer to as a vertical ambiguity describes the ambiguity between the &amp;quot;move cup&amp;quot; vs. &amp;quot;offer drink&amp;quot; meanings of &amp;quot;blicket.&amp;quot; Here the ambiguity is based on the level of description that the speaker intended to convey. Thus, while both meanings are correct (i.e., both meanings accurately describe the action), only one corresponds to the word &amp;quot;blicket.&amp;quot; The second type of ambiguity, referred to as horizontal ambiguity describes the ambiguity between the &amp;quot;offer drink&amp;quot; vs. &amp;quot;ask for change&amp;quot; interpretations of &amp;quot;blicket.&amp;quot; Here there is an ambiguity based on what actually is the intention behind the physical action. Thus, it is the case that only one of these meaning corresponds to &amp;quot;blicket&amp;quot; and the other meaning is not an accurate description of the intended action.</Paragraph>
    <Paragraph position="1"> Figure 1 shows a graphical representation of these ambiguities. Here the leaf nodes represent a basic physical description of the action, while the root nodes represent the highest-level actions for which the leaf actions were performed  . Such a tree representation is useful in that it shows both the horizontal ambiguity that exists between the nodes labeled &amp;quot;ask for change&amp;quot; and &amp;quot;offer drink,&amp;quot; as well as the vertical ambiguity that exits between the nodes labeled &amp;quot;offer drink&amp;quot; and &amp;quot;move cup.&amp;quot;  horizontal ambiguities for actions.</Paragraph>
    <Paragraph position="2"> In order to exploit the intuitive value of such a tree representation, we model intention recognition using probabilistic context free grammars (PCFG)  . We develop a small set of production rules in which the left hand side represents a higher order intentional action (e.g., &amp;quot;offer drink&amp;quot;), and the right hand side represents a sequence of lower level actions that accomplish it (e.g. &amp;quot;grasp cup&amp;quot;, &amp;quot;move cup&amp;quot;, &amp;quot;release cup&amp;quot;). Each individual action (i.e. letter in the alphabet of the PCFG) is further modeled as a simple semantic frame that contains roles for an agent, an object, an action, and multiple optional modifier roles (see inset figure 1). While in this initial work productions are created by hand (a task made feasible by the  In other words, high-level actions (e.g. &amp;quot;be polite) are preformed by means of the performance of lower-level actions (e.g. &amp;quot;offer drink&amp;quot;).</Paragraph>
    <Paragraph position="3">  The idea of a &amp;quot;grammar of behavior&amp;quot; has a rich history in the cognitive sciences dating back at least to Miller et  constrained nature of situated domains) learning such rules automatically is discussed in section 4.2. Just as in the plan recognition work of Pynadath, (1999), we cast the problem of intention recognition as a probabilistic parsing problem in which sequences of physical actions are used to infer an abstract tree representation. Resolving horizontal ambiguities thus becomes equivalent to determining which parse tree is most likely given a sequence of events. Further, resolving vertical ambiguities corresponds to determining which level node in the inferred tree is the correct level of description that the speaker had in mind.</Paragraph>
  </Section>
class="xml-element"></Paper>