AbstractThis paper proposes a new dialogue model that controls utterance initiatives dynamically according to the dialogue context. It classifies utterance intentions into two categories: active and passive. Based on this model, the authors have developed a telephone operator system that can understand utterance intentions by word‐spotting, respond suitably, induce utterances, and anticipate subsequent utterances. The system was developed through simulations using text data of real dialogues between a human telephone operator and callers. The system could understand the intentions of callers' utterances (97 percent) and generate the same reply as a human operator (96 percent