Provided is a man-machine dialogue anomaly detection system and method. The method includes steps: acquiring and marking historical dialogue data, training an anomaly detection model by employing the marked data, conducting anomaly detection by employing the trained anomaly detection model when receiving real-time dialogue data, and obtaining a result. The system comprises an automatic speech recognition module (ASR module), a text-to-speech module (TTS module), a spoken language understanding module (SLU module), a dialogue state tracking module (DST module), a dialogue management module (DM module), a database query module (DATA module), a natural language generation module (NLG module), and an anomaly detection and processing module. According to the system and method, replies given by a machine can be reliable, and the system and method can be applied to any scene.