An Open-domain Dialog Act Taxonomy

This document defines the taxonomy of dialog acts that are necessary to encode domain-independent dialog moves in the context of a task-oriented, open-domain dialog. Such taxonomy is formulated to satisfy two complementary requirements: on the one hand, domain independence, i.e. the power to cover all the range of possible interactions in any type of conversation (particularly conversation oriented to the performance of tasks). On the other hand, the ability to instantiate a concrete set of tasks as defined by a specific knowledge base (such as an ontology of domain concepts and actions) and within a particular language. For the modeling of dialog acts, inspiration is taken from several well-known dialog annotation schemes, such as DAMSL (Core & Allen, 1997), TRAINS (Traum, 1996) and VERBMOBIL (Alexandersson et al., 1997). INTRODUCTION In the context of the EU FP6 project ADAMACH (ADAptive meaning MAChines, EU contract no. 022593), we are designing an adaptive multi-modal spoken dialog system. One of the fundamental steps in such design is the formulation of a taxonomy of dialog acts which constitute the basic units of interaction of the dialog manager. In order to guide the creation of the ADAMACH dialog act taxonomy, the following requirements have been formulated: • The dialog acts must cover the range of basic dialog moves that can take place in opendomain, task-oriented dialog; • They must be sufficient to cover the range of conversational situations taking place in both human-human and human-computer dialog; • They must be domain independent, i.e. valid for any type of task-oriented dialog; • They must be used as the dialog moves encoded by the Dialog Move Engine and to annotate dialogs during training. This means they must have an appropriate level of specificity but must also guarantee that annotation and data-driven classification accuracy will not suffer; Two additional requirements for the future of the ADAMACH project are: • It must be possible to extend and complete the initial set of speech acts with domain-dependent speech acts; • There must be speech acts accounting for multi-modal dialog events.