Experimental conditions for the Ergodic HMM

To test these new techniques, experiments on stochastic network grammar acquisition were performed. For training data, the ATR Dialog Database was used. The learning conditions of these experiments are shown in Table 1 .

表 1: Learning conditions for Ergodic HMM

type of Hidden Markov model	Mealy type
number of training sentences	8,475
total number of words	57,354
vocabulary	6,418 words
end of HMM learning	40th iteration