Conflict Cues in Call Center Interactions - Conflict and Multimodal Communication: Social Research and Machine Intelligence - page 421

Furthermore, this binary set of positive and negative values would facilitate the
inter-annotator agreement experiments, where different annotators are expected to
make faster and more reliable judgments when using binary labels. In addition,
this approach will facilitate the evaluators' task when the emotion is not distinctly
expressed (as in the case of phone calls, where noise is a constant factor).
The distinct 25 categorical labels initially selected were thus mapped to two
values, namely, positive and negative. Table 20.2 shows the diversity of emotion
types that positive and negative classes refer to.
The aforementioned process resulted in the annotation of 1,396 speech units.
Their distribution according to content categories, label type, gender, and speaker
role is depicted in Figs. 20.1 and 20.2 below. The call types where an emotional
behavior is detected more frequently are churns and customers. The negative
emotional label prevails over the positive one, suggesting that when speakers
exhibit an emotional behavior, it is usually targeted to expressions of complaints
and dissatisfaction. Furthermore, the majority of emotional units are uttered by
customers.

Table 20.2 Coarse categorical emotion labels used and mapping to detailed

values
Coarse Fine grained
Positive Pleasure, satisfaction, excitement, interest, politeness, certainty, relief,
trust, surprise, reassurement
Negative Anger, annoyance, irritation, disappointment, frustration, anxiety,
worry, helplessness, confusion, doubt, uncertainty, irony, indifference,
surprise, suspicion

Fig. 20.1 Annotated units per file type—ratio of the total duration of the annotated units over the
total duration of the audio files according to the call types ( left ) and the binary emotional labels
( right )