Is this correct? My sense from the article is that they did all the training on unsupervised, and then checked one of the recurrent lines for a correlation to sentiment, not adding a layer and doing more training.