Semiautomated Identification and Classification of Customer Complaints

This paper examines the feasibility of extracting useful information from customer comments using a Naïve Bayes classifier. This was done for a database, obtained from a large Korean mobile telephone service provider, of 533 customer calls to call centers in 2009. After eliminating calls not containing customer complaints or comments, the remaining 383 comments were classified by an expert panel into four domains and 27 complaint categories. The four domains were Transaction‐related (189 comments, 49%), Product‐related (120 comments, 31%), Customer Service or Support‐related (38 comments, 10%) and Customer Outreach and Marketing‐related (36 comments, 9%). The comments were then randomly assigned to either a training set (257 cases, 67%) or test set (126 cases, 33%). The training set was used to develop a Naïve Bayes classifier that correctly predicted the domain 75% of the time and the specific subcategory 51% of the time for the test set. Prediction accuracy was strongly related to prediction strength for both sets of predictions, suggesting that simple filtering strategies where difficult to understand comments are flagged for expert review and easy comments are automatically classified are both technically feasible and likely to be practically valuable. Several strong predictors were also identified that corresponded to categories more detailed than those originally assigned. © 2012 Wiley Periodicals, Inc.