[kgwgk]

Implement the probabilistic model of token sequences and fit it to the training data?

You could say the same for a character recognition system.