A Joint Optimization Method for Multi-Turn Dialogue Intent Prediction and Adaptive Human-Machine Transfer in Intelligent Customer Service Scenarios

Abstract

In multi-turn customer service dialogue, intent prediction and human handoff are often treated as separate tasks, leading to inconsistent decisions when model confidence is low. In this paper, we propose an uncertainty-aware joint modeling framework that connects these two processes. A shared dialogue representation is learned, followed by dual modules for intent prediction and transfer decision. We introduce an entropy-based uncertainty measure to capture prediction confidence and formulate handoff as a conditional decision dependent on both intent and uncertainty. A lightweight consistency constraint is further applied to align confidence with system behavior.The method we put forward offers a straightforward and flexible alternative to conventional rule‑based schemes.