question about the cross entropy

[phoenixzxj]

In the post "Recurrent Neural Networks in Tensorflow I"
it writes that
"If the network learns no dependencies, it will correctly assign a probability of 62.5% to 1, for an expected cross-entropy loss of about 0.66.If the network learns only the first dependency (3 steps back) but not the second dependency, it will correctly assign a probability of 87.5%, 50% of the time, and correctly assign a probability of 62.5% the other 50% of the time, for an expected cross entropy loss of about 0.52.If the network learns both dependencies, it will be 100% accurate 25% of the time, correctly assign a probability of 50%, 25% of the time, and correctly assign a probability of 75%, 50% of the time, for an expected cross extropy loss of about 0.45.".
How to get that probability “0.875， 0.625”? As the sample has been devided into x(t-3) is 1 and not 1, is it right to do so ?
The cross entropy should be the sum of real * log(expected). There is no model now, so the probability of correctly assigning 1gives to it?