Instead of trying to get NELL to "un-learn" something, I would see if weighted metrics could be applied to multiple tags for an individual record. As the "learning" progressed the weights would favor a particular categorization, but others would still be on record.
So Klingons as an "Ethnic Group" would end up with a lower metric than Klingons as a "Fictional Alien Species" or "Humanoids" or "animal with bilateral symmetry", etc..
The weighting mechanism could be as simple as an integer hit count in NELL's matching logic. I couldn't get to the article to read the details so I would guess that the code operates as a neural network at some level, so allowing the creation of weighted links should be do-able.