A multi-level annotation meta-scheme for dialogue acts

Abstract - This article describes a new principled framework for comparison, design and standardization of annotation schemes for dialogue acts. Previous attempts at comparing existing schemes in order to identify a common core of generally agreed-upon dialogue acts share the assumption that tags belonging to different schemes and describing the same general phenomena can always be related through hypo- or hyperonymy relationships. Consequently, general-purpose schemes have often been the result of a merger of different tag sets. In this article, we show the extent to which comparability of different annotation schemes is prevented by the very limited tag inter-translatability. We thus describe an alternative approach to the comparison of dialogue act taxonomies based on a compositional analysis of tags according to independent classificatory dimensions. The framework takes a recognition-based approach to dialogue tagging and defines four independent taxonomies of tags, one for each orthogonal dimension of linguistic and contextual analysis assumed to have a bearing on identification of dialogue acts. We also show how the same framework can be used to design a generalpurpose annotation scheme which combines the features of generality and expressivity by exploiting a modular structure. The advantages and limitations of this proposal over other previous attempts are discussed and concretely exemplified.