Real-life emotions detection on Human-Human spoken dialogs

L. Devillers, L. Vidrascu

In the paper we present emotional annotation for a corpus of naturalistic data recorded in a French Medical call center. When studying real-life data, there are few occurrences of full blown emotions but also there are many emotion mixtures. To represent emotion mixtures, an annotation scheme with the possibility to choose two verbal labels per segment was used by 2 expert annotators. A closer study of these mixtures has been carried out, revealing the presence of conflictual valence emotions. Results of the perceptive test show 85% of consensus between expert and naive labellers. When selecting the non-complex part of the annotated corpus, the performances obtained are around 60% of good detection between four emotions for respectively agents and callers.

PDF full paper