I thought the idea of using unlabeled data to help increase the performance of the training was a good addition.

I did, however, have some reservations with the comparisons that the paper made with the other algorithms (REG, MAP, ADJ, etc.). The tests that the paper provided used both labeled and unlabeled data for the training. But (someone correct me if I'm wrong), it seems that the only algorithm that actually uses the unlabeled data is the ADA algorithm discussed in the paper. This does give the ADA algorithm a slight advantage because it is provided with more information than the others. Also, this is a comparison between algorithms with similarities but seem to fall in slightly different classes. Because of this, I think that we should be careful in interpreting the results of the tests.