<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1309">
  <Title>Tracing Actions Helps in Understanding Interactions</Title>
  <Section position="4" start_page="60" end_page="61" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> There are several main research directions on dialogue understanding. The one closest to our approach is activity-based dialogue analysis (Allwood, 1997; Allwood, 2000) contrasting BDIstyle approaches such as the one by (Cohen and Levesque, 1995). This research shows how speech acts are related to expectations expressed by means of language and inspired our approach.</Paragraph>
    <Paragraph position="1"> However, ALLWOOD does not work out in detail how the pragmatics of the application domain can be formalized in a tractable way. (Carletta, 1992)  shows in a corpus analysis that risk taking is a elementary behavior of dialogue participants. (Bos and Oka, 2002) uses first-order logic in a DRT environment to reason about the logical satisfiability of a new utterance given a previous discourse. For reasoning about action however, we think that a first-order theorem prover or model builder is not the ideal tool because it is too general. Additionally, in dialogues about acting in an environment, the primary interest of semantic evaluation is not whether a formula is true or false, but how a goal or task can be solved. Therefore, planning is more appropriate than proofing formulae. Work on planning as part of dialogue understanding is reported in (Zinn, 2004). This paper does not address selecting strategies for error recovery. Conflict resolution is addressed in (Chu-Carroll and Carberry, 1996). However, the presented discourse model is not computationally effective. (Huber and Ludwig, 2002; Ludwig, 2004) present an interactive system which uses planning, (Yates et al., 2003) and recently (Lieberman and Espinosa, 2006) reported on applying planning as a vehicle for natural language interfaces, but none of the papers discusses how a dialogue can be continued when a failure in the application occurs. In the WITAS system (see (Lemon et al., 2002)), activities are modelled by activity models, one for each type of activity the system can perform or analyse. A similiar recipe-based approach is implemented in COLLAGEN (Garland et al., 2003). As activities are hard-coded in the respective model, adaptation of the task and dialogue structure to the needs in a current situation arehardertoachievethaninourapproachinwhich only goals are specified and activities are selected by a planner depending on the current state. In addition, executing plans by verifying preconditions and effects of an activity that has been carried out recently lies the basis for a framework of understanding the pragmatics of a dialogue that is not implemented for a particular application, but tries to be as generic as possible.</Paragraph>
  </Section>
class="xml-element"></Paper>