Mining Call Center Conversations exhibiting Similar Affective States

Automatic detection and identifying emotions in large call center calls are essential to spot conversations that require further action. Most often statistical models generated using annotated emotional speech are used to design an emotion detection system. But annotation requires substantial amount of human intervention and cost; and may not be available for call center calls because of the infrastructure issues. Therefore detection systems use models that are generated form the readily available annotated emotional (clean) speech datasets and produce erroneous output due to mismatch in training-testing datasets. Here we propose a framework to automatically identify the similar affective spoken utterances in large number of call center calls by using the emotion models that are trained with the freely available acted emotional speech. Further, to reliably detect the emotional content, we incorporate the available knowledge associated with the call (time lapse of the utterances in a call, the contextual information derived from the linguistic contents, and speaker information). For each audio utterance, the emotion recognition system generates similarity measures (likelihood scores) in arousal and valence dimension using pretrained emotional models, and further they are combined with the scores from the contextual knowledge-based systems, which are used to reliably detect the similar affective contents in large number of calls. Experiments demonstrate that there is a significant improvement in detection accuracy when the knowledge-based framework is used.