A Compact Architecture for Dialogue Management Based on Scripts and Meta-Outputs

A Compact Architecture for Dialogue Management Based on Scripts ond Meta-Outputs MAnny Rayner, Beth Ann Hockey, FrAnkie J~mes Research Institute for Advanced Computer Science Mail Stop 19-39, NASA Ames Research Center Moffett Field, CA 94035-1000 { mauny, bahockey, l~ ames} ~riacs.edu Abstract We describe an architecture for spoken dialogue interfaces to semi-autonomous systems that trans- forms speech signals through successive representa- tions of linguistic, dialogue, and domain knowledge. Each step produces an output, and a meta-output describing the transformation, with an executable program in a simple scripting language as the fi- nal result. The output/meta-output distinction per- mits perspicuous treatment of diverse tasks such as resolving pronouns, correcting user misconceptions, and optimizing scripts. 1 Introduction The basic task we consider in this paper is that of using spoken language to give commands to a semi- autonomous robot or other similar system. As ev- idence of the importance of this task in the NLP community note that the early, influential system SHRDLU (Winograd, 1973) was intended to address just this type of problem. More recent work on spo- ken language interfaces to semi-antonomous robots include SRrs Flakey robot (Konolige et al., 1993) and NCARArs InterBOT project (Perzanowski et al., 1998; Perzanowski et al., 1999). A number of other systems have addressed part of the task. Com- mandTalk (Moore et al., 1997), Circuit Fix-It Shop (Smith, 1997) and TRAINS-96 (Traum and Allen, 1994; Tranm and Andersen, 1999) are spoken lan- guage systems but they interface to simulation or help facilities rather than semi-autonomous agents. Jack's MOOse Lodge (Badler et al., 1999) takes text rather than speech as natural language input and the avatars being controlled are not semi-autonomous. Other researchers have considered particular aspects of the problem such as accounting for various aspects of actions (Webber, 1995; Pyre et al., 1995). In most of this and other related work the treatment is some variant of the following. If there is a speech inter- face, the input speech signal is converted into text. Text either from the recognizer or directly input by the user is then converted into some kind of logi- cal formula, which abstractly represents the user's intended command; this formula is then fed into a command interpreter, which executes the command. We do not think the standard treatment outlined above is in essence incorrect, but we do believe that, as it stands, it is in need of some modification. This paper will in particular make three points. First, we suggest that the output representation should not be regarded as a logical expression, but rather as a pro- gram in some kind of scripting language. Second, we argue that it is not merely the case that the process of converting the input signal to the final represen- tation can sometimes go wrong; rather, this is the normal course of events, and the interpretation pro- cess should be organized with that assumption in mind. Third, we claim, perhaps surprisingly, that the first and second points are related. These claims are elaborated in …

Content truncated to 500 words