Text this: Improving Text Recognition Accuracy for Serbian Legal Documents Using BERT