A traditional “supervised” methodology requires manual labeling of data according to desired annotators or classes, and then training the supervised algorithm to “learn” the annotators or classes. This method has been deficient due, not only to the time consumption, but also to the fact that human supervision may bias the supervised algorithm according to the human supervisor(s). 

Additionally, in the scenario where the corpus of text is dynamic, it is not possible to extract new information without repeating the annotating and learning cycle.