You should check out the VOLT paper, I think it would work well. It's a new technique for splitting up a vocabulary into subwords while minimizing entropy. These subwords could then be mixed and matched, maybe by a neural model, for better results.