From a quick look of the link, to get the conditional softmax you create a softmax function for each discrete state of the discrete parent(s). And then I guess you can infer the parameter in the softmax function and the probability of the discrete state jointly.