You could probably use some heuristic on the tokens trained to try to weight customer service related data higher.