Novel role-wise attention mechanism for predicting purchases made through chat-based online customer services

Abstract

Online chat has become a key channel for firms to provide pre-purchase customer services and generate purchases. Firms must deploy their service personnel based on the customers' purchase probability. Conversational contingency is indicative of purchase probability. However, the unstructured nature of textual data hinders the use of interaction information. We develop a novel role-wise attention mechanism to capture the conversational contingency among customers and service agents. Our results reveal that the role-wise attention mechanism is effective for measuring service interaction quality. When the mechanism is combined with a deep-learning model, it outperforms traditional self-attention mechanisms for predicting customer purchases. The area under the curve of the model is 0.85. Ablation studies indicate that external attention, a core element of the role-wise attention mechanism, substantially improves prediction performance by capturing the conversational contingency between communicators.