Call Reason Prediction using Hierarchical Models
Vijay Varma Malladi, Suman Deb Roy · 2023
We propose a supervised method to predict a reason for making call to customer contact centers by exploiting domain knowledge wherein domain experts propose a few reason codes to label the calls. They also mark an utterance in the call that would indicate the reason for making the call, this is called a marked utterance. Thus we solve a two-level prediction problem for a call, first to classify whether an utterance is marked, and then predict a reason code around this marked utterance as a label for the call. We use hierarchical models (LSTM or BERT layered on top of a transformer) for modeling purpose. We carry out comprehensive experiments using our internal (labeled) call data set, marked by domain experts, and achieve some interesting results like the accuracy of prediction of reason codes leaps to 85% using hierarchical model.