Annotations for Dynamic Diagnosis of the Dialog State

Abstract

This paper describes recent work aimed at relating multi-level dialog annotations with meta-data annotations for a corpus of real humanhuman dialogs. This work is carried out in the context of the AMITIES project in which spoken dialog systems for call center services are being developed. A corpus of 100 agent-client dialogs have been annotated with three types of annotations. The first are utterance-level DAMSL-style dialogic labels. The second set of annotations applies to exchanges and takes into account of the dynamic aspect of dialog progress. Finally, 5 emotions types are annotated at the utterance level. Some of these multi-style annotations were used in a multiple linear regression analysis to predict dialog quality. The predictive factors are able to explain about 80% of the dialog accidents.

Resources

Details