Automatic Recognition of the General-Purpose Communicative Functions Defined by the ISO 24617-2 Standard for Dialog Act Annotation

From the perspective of a dialog system, it is important to identify intention behind segments in dialog, since provides an cue regarding information that present and how they should be interpreted. ISO 24617-2, standard for act annotation, defines hierarchically organized set general-purpose communicative functions which correspond different intentions are relevant context dialog. We explore automatic recognition these DialogBank, reference dialogs annotated according this standard. To do so, we propose adaptations existing approaches flat allow them deal with hierarchical classification problem. More specifically, use end-to-end network cascading outputs maximum posteriori path estimation predict function at each level hierarchy, preserve dependencies between path, decide stop. Furthermore, amount DialogBank small, rely on transfer learning processes reduce overfitting improve performance. The results our experiments show approach outperforms both one based multiple classifiers its components plays role towards functions.