How to calculate the error at each timestep in a LSTM for binary classification?

Question

I\'m trying to understand how the error in a LSTM is backpropagated for binary classification. In other applications of LSTMs, there is usually a "correct" output