The problem is that if there is some classifier bias in different domains, we would need to train the model on hand-labeled data for each domain. If the number of topics we are interested in is small, this could be solved with crowdsourced labeling, as you suggest. But how would we scale to hundreds of thousands of topics (i.e. all words and bigrams, or all entities in Freebase)? That’s an interesting problem that warrants some thought.