PROBLEM TO BE SOLVED: To dispense with data formed by previously labeling satisfaction for every dialogue action; and to evaluate the satisfaction for every dialogue action in the dialogue.SOLUTION: A dialogue learning apparatus learns hidden Markov models M, ..., Mhaving such a state as outputting dialogue actions of a speaker for every evaluation phase. Alternatively, the dialogue learning apparatus may learn a hidden Markov model Mhaving such a state as outputting the dialogue actions of the speaker from all dialogues. All the states of the hidden Markov models (M,) M, ..., Mare connected together to create a hidden Markov model M. Likelihood may be improved by repeating the learning. A dialogue analyzer of the present invention obtains an evaluation value of each dialogue action by estimating from which evaluation value state of the hidden Markov model Mthe dialogue action is output using the hidden Markov model M, and estimates the evaluation value of the dialogue.
展开▼
