Spoken dialogue systems promise efficient and natural access to a largevariety of information sources and services from any phone. However, currentspoken dialogue systems are deficient in their strategies for preventing,identifying and repairing problems that arise in the conversation. This paperreports results on automatically training a Problematic Dialogue Predictor topredict problematic human-computer dialogues using a corpus of 4692 dialoguescollected with the 'How May I Help You' (SM) spoken dialogue system. TheProblematic Dialogue Predictor can be immediately applied to the system'sdecision of whether to transfer the call to a human customer care agent, or beused as a cue to the system's dialogue manager to modify its behavior to repairproblems, and even perhaps, to prevent them. We show that a ProblematicDialogue Predictor using automatically-obtainable features from the first twoexchanges in the dialogue can predict problematic dialogues 13.2% moreaccurately than the baseline.
展开▼
