An IVR call performance classification system using computational intelligent techniques
Date
2010-09-16
Authors
Patel, Pretesh Bhoola
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Speech recognition adoption rate within Interactive Voice Response (IVR) systems is on
the increase. If implemented correctly, businesses experience an increase of IVR
utilization by customers, thus benefiting from reduced operational costs. However, it is
essential for businesses to evaluate the productivity, quality and call resolution
performance of these self-service applications. This research is concerned with the
development of a business analytics for IVR application that could assist contact centers
in evaluating these self-service IVR applications. A call classification system for a pay
beneficiary IVR application has been developed. The system comprises of field and call
performance classification components. ‘Say account’, ‘Say amount’, ‘Select beneficiary’
and ‘Say confirmation’ field classifiers were developed using Multi-Layer Perceptron
(MLP) Artificial Neural Network (ANN), Radial Basis Function (RBF) ANN, Fuzzy
Inference System (FIS) as well as Support Vector Machine (SVM). Call performance
classifiers were also developed using these computational intelligent techniques. Binary
and real coded Genetic Algorithm (GA) solutions were used to determine optimal MLP
and RBF ANN classifiers. These GA solutions produced accurate MLP and RBF ANN
classifiers. In order to increase the accuracy of the call performance RBF ANN classifier,
the classification threshold has been optimized. This process increased the classifier
accuracy by approximately eight percent. However, the field and call performance MLP
ANN classifiers were the most accurate ANN solutions. Polynomial and RBF SVM
kernel functions were most suited for field classifications. However, the linear SVM
kernel function is most accurate for call performance classification. When compared to
the ANN and SVM field classifiers, the FIS field classifiers did not perform well. The
FIS call performance classifier did outperform the RBF ANN call performance network.
Ensembles of MLP ANN, RBF ANN and SVM field classifiers were developed.
Ensembles of FIS, MLP ANN and SVM call performance classifiers were also
implemented. All the computational intelligent methods considered were compared in
relation to accuracy, sensitivity and specificity performance metrics. MLP classifier
solution is most appropriate for ‘Say account’ field classification. Ensemble of field
classifiers and MLP classifier solutions performed the best in ‘Say amount’ field
classification. Ensemble of field classifiers and SVM classifier solutions are most suited
in ‘Select beneficiary’ and ‘Say confirmation’ field classifications. However, the
ensemble of call performance classifiers is the preferred classification solution for call
performance.