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

Information Technology Reference

In-Depth Information

in detecting anger in voice-controlled telephone services (Neiberg and Elenius 2008 ;

Burkhardt et al. 2009 ) (average recall of 83 % and 69 %, respectively). Duration

measures seem not to play an important role in emotion detection (Burkhardt

et al. 2009 ). Incorporating linguistics and training of “emotion salient” words

seems promising (Lee and Narayanan 2005 ), with fusion of acoustic and linguistic

cues slightly improving overall scores as in Polzehl et al. ( 2011 ). Approaches to

automatic conflict detection rely on extracting and analyzing nonverbal behavioral

informative cues (Kim et al. 2012 ) and exploit turn organization features (Pesarin

et al. 2012 ).

20.2

Data Collection

Our data collection consists of 135 audio files corresponding to call center human-

human dyadic conversations between an operator and a customer. The conversations

come from a customer support service of a Greek telecommunications company and

are classified according to their content in six major categories: (1) churns, outgoing

calls to contractual customers that have requested to cancel their contract with the

company (the operators ask the customers the reason for their choice and attempt

to change their mind); (2) customers, incoming calls about any issue related to

customer service (technical problems, bills, complaints, etc.); (3) telesales, outgoing

calls to customers aiming to sell regular phone contracts; (4) upgrade, outgoing calls

informing customers about new offers; (5) mobile, outgoing calls aiming to sell

mobile phone contracts; and (6) welcome, outgoing calls to welcome the customers

to the company's network. Each category consists of unequal number of files.

The overall duration of the corpus is approximately 9.5 h. Each conversation

corresponds to a unique customer, while the operator might be the same in more than

one conversation. The distribution of audio files per content, duration, and number

of speakers of the corpus is shown in Table 20.1 .

Table 20.1

Corpus details

# of speakers