My objection is more about the data used for training, assuming we are talking about unsupervised learning. Text alone just won't cut it.