Preserving Empirical Probabilities in BERT for Small-sample Clinical Entity Recognition